From 4989f2865ea22976981db68518764512bbd39813 Mon Sep 17 00:00:00 2001 From: Jason Morse Date: Mon, 16 Nov 2026 11:35:17 -0800 Subject: [PATCH] docs(desktop): restructure package status docs into topic sections Replace the append-only dated logs in the current-status references of desktop-test, desktop-driver, desktop-driver-mcp and storybook-desktop with a status overview followed by stable topic sections, so changes to different topics merge cleanly. Each native endpoint keeps only its latest result with a link to the pull request or issue that holds the full evidence, and open work links to its tracking issue instead of restating it. - Add R-PKGDOC-010 describing the format. - Update README, reference-index and cross-reference wording and anchors. - Preserve the macOS activation source trace in the driver remaining-work plan. - Add a changeset. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 03bf8ad5-441c-41df-9738-791afdcbb6b5 --- .changeset/status-topic-sections.md | 8 + packages/agentic/desktop-driver-mcp/README.md | 6 +- .../desktop-driver-mcp/references/README.md | 10 +- .../references/current-status.md | 1122 ++--- .../desktop-driver-mcp/references/tools.md | 2 +- packages/agentic/desktop-driver/README.md | 6 +- .../desktop-driver/references/README.md | 28 +- .../references/current-status.md | 3641 ++------------- .../references/native-helpers.md | 2 +- .../references/remaining-work.md | 27 +- packages/agentic/desktop-test/README.md | 2 +- .../agentic/desktop-test/references/README.md | 22 +- .../desktop-test/references/current-status.md | 3993 ++--------------- .../references/native-qualification.md | 10 +- .../storybook-desktop/references/README.md | 2 +- .../references/current-status.md | 692 +-- .../agents/repository/package-guidance.md | 1 + 17 files changed, 1218 insertions(+), 8356 deletions(-) create mode 100644 .changeset/status-topic-sections.md diff --git a/.changeset/status-topic-sections.md b/.changeset/status-topic-sections.md new file mode 100644 index 0000000000..d0bff1c195 --- /dev/null +++ b/.changeset/status-topic-sections.md @@ -0,0 +1,8 @@ +--- +"@fluentui-react-native/desktop-driver": patch +"@fluentui-react-native/desktop-driver-mcp": patch +"@fluentui-react-native/desktop-test": patch +"@fluentui-react-native/storybook-desktop": patch +--- + +Restructure each package's current-status reference from a dated running log into a status overview with stable topic sections and the latest result per native endpoint. Per-run evidence now lives in pull request descriptions and issues, and remaining-work detail lives in the linked tracking issues. diff --git a/packages/agentic/desktop-driver-mcp/README.md b/packages/agentic/desktop-driver-mcp/README.md index 34c58b5f01..1e59a411ce 100644 --- a/packages/agentic/desktop-driver-mcp/README.md +++ b/packages/agentic/desktop-driver-mcp/README.md @@ -5,9 +5,9 @@ coding agents the Desktop Driver agent operations: bounded observation, element search, actions, checks with waits, captures, and Storybook story selection on a native macOS or Windows application. -**Work in progress:** [current status](references/current-status.md) records -implemented connection safeguards, orchestration gaps, native qualification -requirements, and measured MCP discovery costs. +**Work in progress:** [current status](references/current-status.md) summarizes +implemented connection safeguards, open orchestration work, per-endpoint native +qualification, and the latest measured MCP discovery costs. Agents working in this package start at [AGENTS.md](AGENTS.md). diff --git a/packages/agentic/desktop-driver-mcp/references/README.md b/packages/agentic/desktop-driver-mcp/references/README.md index b823f2ed68..c01b002fa1 100644 --- a/packages/agentic/desktop-driver-mcp/references/README.md +++ b/packages/agentic/desktop-driver-mcp/references/README.md @@ -7,8 +7,8 @@ owner_paths: # Desktop Driver MCP references -| ID | Document | Read when | -| ----------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------ | -| `desktop-driver-mcp.connection` | [Connection and setup](connection.md) | Publishing a descriptor, registering the server, or diagnosing connect | -| `desktop-driver-mcp.tools` | [Tools](tools.md) | Calling or changing tools, results, errors, resources, or progress | -| `desktop-driver-mcp.current-status` | [Current status](current-status.md) | Checking implemented lifetime safeguards, native qualification gaps, and measured discovery cost | +| ID | Document | Read when | +| ----------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------- | +| `desktop-driver-mcp.connection` | [Connection and setup](connection.md) | Publishing a descriptor, registering the server, or diagnosing connect | +| `desktop-driver-mcp.tools` | [Tools](tools.md) | Calling or changing tools, results, errors, resources, or progress | +| `desktop-driver-mcp.current-status` | [Current status](current-status.md) | Checking implemented lifetime safeguards, per-endpoint qualification, and measured discovery cost | diff --git a/packages/agentic/desktop-driver-mcp/references/current-status.md b/packages/agentic/desktop-driver-mcp/references/current-status.md index f32709034b..5c5400d45d 100644 --- a/packages/agentic/desktop-driver-mcp/references/current-status.md +++ b/packages/agentic/desktop-driver-mcp/references/current-status.md @@ -7,868 +7,262 @@ owner_paths: # Current adapter and measurement status -**Work in progress, updated 2026-11-12.** The adapter is a thin client, not a -supervisor. Keep this package name until actual harness-owned management tools -land; a later `desktop-mcp` facade must retain driver-only use and plan published -compatibility. Done below describes implementation, not full native qualification. -Follow-up work is tracked in -[desktop driver improvements (#4435)](https://github.com/microsoft/fluentui-react-native/issues/4435); -connection isolation is [#4440](https://github.com/microsoft/fluentui-react-native/issues/4440) -and the generalized management facet is -[#4450](https://github.com/microsoft/fluentui-react-native/issues/4450). - -## Capability matrix - -| Capability | State | Boundary/remaining work | -| --------------------------------------------------- | ------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Six tools over shared driver schemas | Done | Session, observe, act, check, capture, story. No separate execution path. | -| Private descriptor and native identity verification | Done | Verifies before sending a grant. Uses the driver's `/native`, `/authority`, and `/client` subpaths, not its WebdriverIO-heavy root; boundary tests guard this dependency. No model URL/token/config/executable input. | -| Single-flight connection/session lifetime | Done | Serial acquisition, close, operations and resource reads; late acquisition after dispose is cleaned up. | -| Same-port protected restart/revocation | Done in deterministic contracts | Invalid authority is forgotten; next call re-verifies discovery. No mutation replay. Live endpoint reruns remain required. | -| Same-authority transport retry | Done, bounded semantics | Retain known session cleanup responsibility and rebind only after verification of the same incarnation/principal. This is not cross-MCP-connection resume. | -| Lost create response | Fenced; child drain implemented | `session-uncertain`; no repeated acquisition. Connection disposal revokes and drains the child, including unknown session ids. Live descriptor-path reruns not run for this change. | -| Failed delete/disconnect | Done with limits | Keep known-session responsibility; bounded dispose wait leaves late cleanup attached. Unreachable-service cleanup cannot be guaranteed; service drain/expiry remains necessary. | -| Launch-mode discovery | Done | Status lists effective launch/attach modes; no predictable attach-first failure against launch-only targets. | -| Progress, cancellation, paged image resources | Done | Progress does not renew idle sessions; artifacts are session-bound and hash-checked. | -| Storybook control | Experimental | Driver selection/marker path exists; channel ingress/reconnect and full exclusive context authority remain incomplete. | -| Per-connection attenuated grants | Done in deterministic contracts | Each controller delegates a distinct driver-only child after verification; target/scope subsets and terminal-only cleanup retry. Native descriptor-path qualification not run for this change. | -| Explicit controller relinquish | Done in deterministic contracts | Immediate admission fence and cancellation; return only after child/session drain. Failures retain retryable cleanup responsibility. Fresh open expires old references, cursors, proofs and artifact links. Generic attach-only runner integration is in desktop-test; Storybook and native endpoint qualification remain separate. | -| Cross-connection session resume | Intentionally unsupported | No lease/grace or ownership-transfer protocol. Same-controller rediscovery retains its existing child; no automatic replay or cross-connection resume. | -| Workload/test orchestration tools | Work in progress | Belong to a future harness control/client surface; adapter never becomes the resource owner. | -| General MCP-client qualification | Work in progress | SDK stdio evidence exists; client-specific resource/image/cancellation behavior needs independent runs. | - +This document covers the state of the MCP adapter and its measured discovery and interaction costs. +The adapter is a thin client, not a supervisor. +Done means the stated implementation exists and is covered by deterministic tests; it does not mean every native endpoint is qualified. +Parent issue: [desktop driver improvements (#4435)](https://github.com/microsoft/fluentui-react-native/issues/4435). See [connection](connection.md) and [tools](tools.md) for exact behavior. -Native and Storybook limitations are tracked in -`@fluentui-react-native/desktop-driver/references/current-status.md`; -supervision gaps in -`@fluentui-react-native/desktop-test/references/current-status.md`. - -## Measured discovery and interaction costs - -Historical measurements migrated from the root capability report, **not rerun** -by this update. `o200k_base` was used over all MCP text blocks; tokenizer package -version/raw portable artifact locator were not retained. - -| Date / revision | Measurement | Result | Boundary | -| ------------------------ | -------------------------------------- | ------------------------------- | ----------------------------------------------------- | -| 2026-11-04 / `41fc7473f` | Six-tool `tools/list` | 4,688 text tokens; 20,269 bytes | Discovery once per connection; not per operation | -| Same | Server instructions | 120 text tokens | Additional discovery context | -| Same | Named Toggle + check + closing observe | 386 text tokens; 48 ms | One warm macOS Fabric run, not a latency distribution | -| Same | Physical click + checked check | 181 ms | One run | -| Same | Observe default | 29-31 ms median, max 35 ms | 20 warm calls per button/checkbox/menu story | -| Same | Delta | 32-35 ms median, max 37 ms | Same 20-call-per-story method | - -The app was macOS Fabric Storybook at 2720 x 1988 pixels on a 2x display, -served through `storybook driver --macos --agent`, driven by an MCP SDK stdio -client. Operation text/scoping comparisons and their completeness are retained -in the driver's status reference rather than duplicated here. - -The schema test limits serialized JavaScript string length to 22,000 -**characters**. It is not a token or UTF-8 byte limit. The earlier proposed -8 KiB discovery budget was not achieved; measured discovery is about 20.3 KB. -Changed descriptions/schema text invalidate an exact present-day 4,688-token -claim: retain that number only with its historical revision. - -For each new run append revision, SDK/client version, platform/renderer, -tokenizer/version, serialized UTF-8 bytes and characters, text tokens, -instructions overhead, projection/budgets, sample count, cold/warm timing, -success/retry/cancellation/cleanup counts, and approved artifact locator. -Do not claim token reductions from reducing information scope or omitting -discovery. New macOS Paper samples are recorded in the dated F15 -permission-enabled follow-up below; they do not fill earlier provenance gaps. - -### Windows Fabric and Win32 Paper - 2026-11-08 UTC - -Revision `9d3e11516` plus the uncommitted Windows ownership and foreground -helper changes, which do not alter tool text. MCP SDK 1.32.2 stdio client -through the default descriptor; `gpt-tokenizer` 4.1.0 with `o200k_base` over -every text block. Windows 11 build 26300, AMD64; Storybook served by -`yarn desktop-test --target windows|win32 --reporter json serve --agent`. The -console session was disconnected, so `click` was not measured; the toggle -row uses `activate`. Each endpoint made 139 calls with no errors or retries, -then closed its session. Runs `caba9e12-d9da-42a0-8b06-7300940e8318` -(Windows) and `5f75aac2-d74c-4969-aba3-bd8c925d5725` (Win32), each with its -`run.json` under `apps/storybook/artifacts/desktop//`. - -| Measurement | Windows Fabric | Win32 Paper | -| ------------------------------------------------------- | ----------------------------------- | ----------------------------------- | -| Six-tool `tools/list` | 4,691 tokens; 20,277 bytes (chars) | Same | -| Server instructions | 120 tokens; 591 bytes | Same | -| `open` (attach) | 45 ms | 31-40 ms | -| Checkbox `activate` + `checked` check + default observe | 1,923 tokens; 5,346 bytes; 2,288 ms | 1,645 tokens; 5,187 bytes; 2,302 ms | -| Of which `activate` | 2,190 ms | 2,196 ms | -| Observe default (100 of about 113 nodes, 20 calls) | 66.7-73.7 ms median, max 93.1 ms | 81.1-85.8 ms median, max 97.3 ms | -| Observe text per call | 1,801-1,843 tokens | 1,594-1,633 tokens | -| Delta (20 calls, idle) | 72.7-77.5 ms median, max 98.4 ms | 83.3-88.3 ms median, max 106.4 ms | -| Delta text per call | 368 tokens | 220 tokens | - -Ranges span the Button, Checkbox, and Menu stories (`components-button--appearance`, -`components-checkbox--constrained-text`, `components-menu--default`); the -toggle row used `components-checkbox--default`. Observe used the default -budget, which emitted 100 nodes. On Windows an idle delta is never empty: -the React Native Test App's React, Debug, and System menu-bar items get new -references on every read (three removed, three added). The helper reuses a -reference only when the UI Automation runtime identity matches; whether these -items lack stable runtime identities is unconfirmed. Without those changes -the envelope is about 151 tokens. The Win32 -idle delta was not decomposed. The first Win32 toggle attempt chose the -wrapping-label checkbox in `constrained-text` and returned a `stopped` -sequence; it is retained here, not diagnosed, and the default story was used -instead. - -Discovery grew by 3 tokens and 8 bytes since `41fc7473f` from tool-text -changes. The Windows toggle rows are not comparable with the 386-token macOS -row: they target a different story and their closing observe is an unscoped -default observe, which accounts for most of their text. - -## Before full submission - -Re-run same-port replacement, revocation, dispose during open, failed delete, -ambiguous create, and resource invalidation through the real descriptor path on -macOS and Windows. Qualify Windows Fabric and Win32 Paper independently, with -physical Ctrl+C and interactive input/capture cases where supported. Current -fake-host tests validate portable controller semantics, not native ownership, -signing, TCC, Windows ACLs, DPI, or process containment. This requalification is -tracked in [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). - -## Provisional F15 native endpoint records — 2026-11-06 - -**Exact source revision:** `388a60cc803ab78202a6f64f92e3f53b779c5c39`. -**App:** Agentic Components Storybook / ReactTestApp (configured bundle ID -`com.microsoft.fluentui.agenticstorybook`). -**OS:** macOS 27.2, arm64. -The reusable procedure is -`packages/agentic/desktop-test/references/native-qualification.md`. - -These rows document the native smoke environment only. The smoke used the -non-agent WDIO path; no MCP descriptor was consumed, so these rows are not -descriptor-path passes. - -| Renderer | Command and selection | Passed / failed / skipped / unsupported | Cleanup | Artifact | -| ------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------- | ------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | -| macOS Fabric | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook smoke --macos --mode stories-and-tests`; 211 selected/executed | 211 / 0 / 0 / 0 | No owned processes or listeners on 21338/30029/45382; `interactive.lock` absent | `apps/storybook/artifacts/desktop/0f99740f-631b-444a-a705-76a9b4a761b7/run.json` | -| macOS Paper | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook smoke --macos --paper --mode stories-and-tests`; 211 selected/executed | 211 / 0 / 0 / 0 | No owned processes or listeners on 21338/30029/45382; `interactive.lock` absent | `apps/storybook/artifacts/desktop/e82a83c2-a01d-41c0-b308-2551630e8b1e/run.json` | -| Windows Fabric descriptor path | Not run — Windows is unavailable on this macOS-only machine | Not run; 0 selected | Not applicable | None | -| Win32 Paper descriptor path | Not run — Win32 is unavailable on this macOS-only machine | Not run; 0 selected | Not applicable | None | - -| macOS descriptor-path renderer | Command / selection | Passed / failed / skipped / unsupported | Cleanup | Artifact | -| ------------------------------ | ---------------------------------------------- | --------------------------------------- | -------------------------------------------- | -------- | -| Fabric | Not run; no descriptor consumed; selection N/A | 0 / 0 / 0 / 0; Not run | N/A for descriptor path; smoke cleanup above | None | -| Paper | Not run; no descriptor consumed; selection N/A | 0 / 0 / 0 / 0; Not run | N/A for descriptor path; smoke cleanup above | None | - -Real-descriptor lifecycle, grant revocation, ambiguous create, resource -invalidation, protected args, quarantine/recovery, F2/F3/F4/F5 cases, and MCP -discovery/text/latency measurements remain **Not run** on this revision. Do not -attribute the WDIO callback counts to MCP. The initial Fabric smoke attempt -failed before selection because the generated CocoaPods sandbox was out of -sync; supported `prep` resolved it before the passing run. The story -environment selector did not constrain callback selection; the artifacts show -211 selected. - -## 2026-11-06 partial descriptor-path evidence at `3c780cfd1bccaf510ece080e4be2b516c645989a` - -App: Agentic Components Storybook / ReactTestApp, -`com.microsoft.fluentui.agenticstorybook`; macOS 27.2 arm64. The MCP SDK stdio -client used the default `agent-connection.json` descriptor against the real -native helper. The exact inline client launch command and connection/session -identifiers were not retained. No descriptor grants or story-args values are -recorded here. Manual MCP operations are not suite selections; the companion -driver `run.json` files each report 0 selected/0 executed tests. - -| Renderer | Tool/operation evidence | Result | Artifact / limitation | -| -------- | ---------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Fabric | `desktop_session.status/open/close`; `desktop_story.story` select/reset; `desktop_observe` find/observe; `desktop_capture.capture` | Discovery showed attach-only. Select and reset were applied with marker verification (regions 1 then 2). A separate preview-child reference was rejected with `stale element reference`, reason `ref-region-changed`. Protected args returned `unsupported operation` / `story-args-unsupported`; an intermediate unchanged-story assertion was false, the final assertion was true, and no probe value appeared in error or journal. | Private `fabric-launch-mode-discovery.json`, `fabric-mcp-markers-resources.json`, `fabric-mcp-observe-resource-invalidation.json`, `fabric-protected-args-retry.json`, `fabric-protected-args-final.json`. | -| Fabric | Anonymous/foreign-authority HTTP select, prepare, args; foreign WebSocket; idle channel | HTTP returned 403 for all three mutation classes and native preview stayed unchanged. The first two foreign-WS assertions failed; the final attempt passed with no observer-message delta. The channel remained open after 21,003 ms idle and marker verification succeeded. | Private `fabric-f5-ingress-bypass.json`, `fabric-f5-ingress-bypass-retry.json`, `fabric-f5-ingress-bypass-final.json`, `fabric-f5-channel-idle.json`, `fabric-mcp-live-selection.json`. | -| Paper | `desktop_session.status/open/close`; `desktop_story.story` select/reset; protected args; bypass and idle-channel checks | Attach-only discovery, marker-settled selection/reset, protected args refusal, revocation/drain, bypass checks, and 21-second channel lifetime were reported. A protected preview reference was reported rejected after reset, but handle kind/generation and exact operation-level fields were not retained; classify that invalidation as Partial. | Private Paper driver/MCP logs in `f15-native-qualification-3c780cfd-20261106T0200/`; Paper run metadata: `apps/storybook/artifacts/desktop/a58d2fdd-5584-47f4-8747-5307607b4c0d/run.json`. | - -The initial Fabric observation that a pre-reset reference was accepted after a -later observation remains **Inconclusive**: its handle kind (preview element, -MCP PNG resource, or chrome/window reference), stable identity relationship, -and explicit session continuity were not retained. Do not count an immutable -PNG resource or chrome/window reference surviving reset as a failure. The -separate retained preview-child stale-element rejection is expected behavior, -but does not prove the F5 same-native-ID-after-fresh-find seam. No additional -probe of the inconclusive sequence was run. The later F5 fix at -`2d8e8db001b64ebea85f0d6dce141f040fb5182d` is not part of this candidate; the -original observation is not attributed to that defect. - -No aggregate MCP call count was retained. Dispose-during-open, failed delete, -ambiguous create, story quarantine/recovery, failed-cleanup claim retention, -F4 cancellation/readiness/render/lost-context/secondary-window cases, F3 -failure-survivor injection, F5 legacy separation, and exact manifest-smoke -single-session/no-HTTP-fallback evidence remain Not run. Windows Fabric and -Win32 Paper are Not run on this macOS-only host. At handoff, the default -descriptor and workload lock were absent and the Storybook, Metro, and driver -ports had no listeners. The Fabric cleanup record reports owner exit, -descriptor removal, grant revocation, and attached-app preservation. Paper -cleanup was reported, but its process IDs were not retained for independent -PID-level confirmation. - -## 2026-11-06 combined-candidate descriptor evidence at `a19fd3e1e7b3f1f443a241b1571cb029162f4ef2` - -App: Agentic Components Storybook / ReactTestApp, -`com.microsoft.fluentui.agenticstorybook`; macOS 27.2 arm64. The helper was -rebuilt for this exact source and the fresh nonprompting permission preflight -reported Accessibility, PostEvent, and ScreenCapture granted. The MCP SDK -stdio client consumed the default private `agent-connection.json`; no -descriptor grants, private descriptor data, or story-args values are retained. -The descriptor-path commands were `yarn storybook driver --macos --agent` -and `yarn storybook driver --macos --paper --agent`, paired with -`yarn storybook run --macos` and `yarn storybook run --macos --paper`. -Manual MCP operation sequences are not suite selections; companion driver -run.json counts were 0 selected/0 executed. - -| Renderer | Sanitized operation sequence and application outcomes | Result and locator | -| ---------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Fabric | `desktop_session.status` → `desktop_session.open(launch:false)` → `desktop_story.story(select)` → `desktop_observe.find` → authorized `desktop_story.story(select)` → `desktop_observe.observe(within)` → authorized `desktop_story.story(reset)` → old `observe(within)` → fresh `find` → fresh `observe(within)` → old `observe(within)` → `desktop_session.close`. | Passed. Same driver session/incarnation 1 throughout; preview-element aliases `element-ref-1` and `element-ref-2` were distinct. Select marker settled at region 2 and reset at region 3. Old ref returned `stale element reference` / `ref-region-changed` immediately after select and reset and after fresh find; fresh ref worked. Private sanitized record: `fabric-reference-sequence-retry.json`. | -| Paper | Same operation sequence as Fabric. | Passed with the same identity, marker, region, stale-old-ref, and fresh-ref outcomes. Private sanitized record: `paper-reference-sequence.json`. | -| Fabric idle channel | `desktop_session.status` → `open` → idle 21,002 ms → `status` → `close`. | Passed; status and close succeeded after idle. Private record: `fabric-idle-channel-21s.json`. No equivalent Paper interval on this SHA. | -| Fabric same-port replacement | Old client: `desktop_session.status` → `open` → `desktop_observe.find` (0 matches) → service stopped/replaced on the same ports → old client `desktop_session.status` → `desktop_session.close`. | Partial (recorded Failed at qualification; reclassified on contract review): the first post-replacement status returned `mcpIsError:true`, `error:"adapter"`, `reason:"service-unavailable"`, the designed fail-closed rejection that the same-port contract asserts before the next call re-verifies discovery; close then succeeded. Close is a no-op once invalid authority forgets the session, so native rediscovery was not demonstrated. No old-reference operation was possible because find returned 0 matches, and no mutation was sent. Sanitized record: `fabric-same-port-replacement-final.json`. | -| Fabric physical input | `desktop_act` physical click was sent once after one corrected request shape. | Partial/inconclusive: MCP returned success and sequence completed, but the client parser did not retain `result.steps[0].act` mechanism/disposition. Not a physical-input pass and not repeated. Private record: `fabric-physical-input-retry-corrected.json`. | - -Separate non-agent manifest-backed smoke passed 211/211 callbacks in both -renderers, with 0 failed, skipped, unsupported, incomplete, retries, or flaky -results. These runs are the legacy lane, not protected-ingress evidence, and -the recorded artifacts do not establish the requested one-session/no-HTTP- -fallback invariant. Fabric artifact: -`apps/storybook/artifacts/desktop/a2475e8d-7ea9-465f-8461-b8f115cfcee5/run.json`; -Paper artifact: -`apps/storybook/artifacts/desktop/87f0452f-5fd7-48eb-8494-a63202842521/run.json`. - -The initial Fabric service replacement attempt encountered a retained -`interactive.lock`; it was not removed until its exact recorded owner exited -and the lock showed no resources or unresolved effects. A subsequent launch -attempt was correctly refused while that claim remained; after verified -cleanup, the replacement service started normally. Final cleanup verified -owned app and service PIDs exited, ports 21338/30029/45382 had no listeners, -the descriptor and lock were absent, and no external app was stopped. -Anonymous/foreign HTTP and WS bypass, protected args refusal, grant revocation, -dispose-during-open, failed delete, ambiguous create, resource invalidation -beyond preview elements, quarantine/recovery, F3/F4 injections, and Paper -same-port/idle/physical-input cases are Not run on this exact SHA. - -## 2026-11-06 Windows Fabric and Win32 Paper descriptor evidence - -Source: the PR branch at `1f5333e25` plus `8458bde48` and `a6b89e924`; host -Windows 11 Enterprise build 26300, AMD64. Commands: -`yarn storybook driver --windows|--win32 --agent`, then -`yarn storybook run --windows|--win32`; the MCP SDK stdio client consumed the -default private `agent-connection.json`. No grants, descriptor data, or story -args values are retained. Manual MCP calls are not suite selections. The full -matrix (Windows Fabric 15 Pass, 1 Partial, 1 Failed, 7 Not run; Win32 Paper -15 Pass, 2 Partial, 7 Not run), run records, and cleanup are in the "Windows -Fabric and Win32 Paper qualification — 2026-11-06" section of -`packages/agentic/desktop-test/references/current-status.md`. -These rows supersede the macOS-host "Not run" Windows rows above for the -Windows host only. - -| Area | Windows Fabric and Win32 Paper outcome | -| ----------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Same-port replacement | Pass on both. The old client's status returned `service-unavailable` and its refs returned `no-session`; a later call rediscovered the new service, and close and a new open succeeded. The first select after replacement returned `relaunch-required` (rejected, effect none). No old mutation replayed. | -| Authority faults | Pass on both for grant revocation, dispose during open, failed delete, and ambiguous create (`session-uncertain` until the grant was revoked). Windows grant revocation and failed delete each passed on a rerun; the first attempts are retained in the desktop-test ledger. | -| Resources and refs | Pass on both. Preview element refs became stale after reset and after the next select; chrome/window refs kept their identity; the immutable MCP PNG resource stayed readable after reset with a matching digest. One session and incarnation 1 throughout each sequence. | -| Quarantine and recovery | Pass on both. An interrupted select returned `story-not-ready` (indeterminate, quarantined); later mutations in the same session returned `target-quarantined` while reads worked. A new session's exit probe lifted the quarantine, and select and reset applied with marker verification. | -| Protected args | Pass on both: `story-args-unsupported`, with no value in the response. | -| F4 cancellation | Pass on both. After a client abort at 150 or 40 ms, replaying the request ID returned the recorded applied result, and the next mutation applied in the same session. | -| F5 ingress and lifetime | Pass on both. Unauthenticated HTTP returned 403, an anonymous WebSocket received only the hello request, and the session stayed usable after about 21 s idle. | -| Not run | Failed-cleanup claim retention, live cleanup and restart, F4 readiness/render faults, F4 secondary windows, F3 survivors, F5 manifest smoke, and F5 legacy separation. No MCP discovery-size or latency measurements were recorded on Windows. | - -## 2026-11-07 UTC Windows follow-up - -Source `386f0585edae2450ab22e6d359e2fe4db3399555`, plus validation, -atomic-report repairs, and guidance. The seven Not run rows above were exercised; the -additive desktop-test ledger now records Windows 21 Pass, 2 Partial, 1 Fail -and Win32 21 Pass, 3 Partial. Full run IDs, ref kinds/identities/generations, -failure history, and operator cleanup proof are in -`packages/agentic/desktop-test/references/current-status.md`. - -The real SDK stdio client continued to use only the default private -descriptor and native-verified service. Delayed readiness rejected a stale -correlation; correlated render faults quarantined the same session while -chrome reads remained usable. Old preview refs expired. No fresh preview -root was exposed after those errors, so this does not fill Win32's -fresh-preview fence gap. Secondary Callout probes were Partial: no separately -registered secondary target was selected. - -Exact failed-cleanup and startup-survivor claims blocked replacement until -the original stop callbacks settled. After exact operator reconciliation, new -descriptor-backed services selected/reset with marker verification in -Windows session `fbbdadc4-aa0e-45a9-9862-124052bc5e4e` and Win32 session -`0245de0c-c004-4c76-865d-7458b62f999f`. Normal shutdown removed discovery, -revoked authority, freed ports/resources, and preserved attached apps. - -Manifest smoke/legacy traces were separate **unprotected** workloads. They -used no descriptor or grant and provide no additional protected-ingress -evidence. The initial local probe's cached `bridge-unavailable` loop was a -request-ID reuse error: the corrected client issued a new ID only after -rejected/effect-none proof, never after an indeterminate mutation. Earlier -failed attempts and the local control-file crash remain recorded. - -The final separate non-agent smoke runs passed: Windows -`1825d9aa-2c25-4a2c-8524-ceadcf52ad68` (234 passed, 78 skipped) and -Win32 `8c6f58e1-3591-4f1b-b008-c52a718d3ca2` (214 passed, 52 skipped). -Windows recovered one atomic-report sharing conflict with a journal warning; -Win32 needed none. All sessions closed and cleanup removed owned processes, -listeners, and the workload claim; no default descriptor remained. These -runs issued no descriptor grant and do not change the descriptor matrix's -Partial or Fail outcomes. - -## 2026-11-08 UTC connected descriptor-path checks - -Source `fec94cec4995c7372b798877c5412ddcb156523b`, Windows 11 Enterprise -build 26300, AMD64. An MCP SDK stdio client used the default private -descriptor and attached to the real, instance-matched Storybook app under an -exclusive claim. These are manual operations, not selected suite tests. - -| Endpoint | Driver run / MCP session | Physical checkbox | Target PNG | -| -------------- | ------------------------------------------------------------------------------- | ----------------------------------------------------------------------------- | --------------------------------------------- | -| Windows Fabric | `af97fdb3-68de-4ce9-836e-935331528c8e` / `6ab236be-dc25-4b53-97f9-4e6ce8875ab2` | Pass: 2 applied `physical-pointer` actions, unchecked -> checked -> unchecked | Pass: 42,436 bytes, resource SHA-256 verified | -| Win32 Paper | `3294d66d-9332-4639-95fe-8d3047f47e46` / `8d75052d-ddae-48f9-a1c1-2c8abcf8a019` | Pass: 2 applied `physical-pointer` actions, unchecked -> checked -> unchecked | Pass: 31,639 bytes, resource SHA-256 verified | - -Each session selected `components-checkbox--default`, checked every state, -read the PNG through its MCP resource, verified its signature and digest, -and closed. The Windows case supersedes the earlier failed physical-click -attempt for this bounded descriptor-path operation. The first Windows -readiness attempt ended after 30 rejected, effect-none retries while Metro -was still building; a later attempt passed. - -On Win32, a session-local injection closed the authenticated bridge after -selection. A newly observed preview checkbox was unavailable (`find: none`) -while the chrome root remained present; fresh-preview mutation fencing and -chrome mutation remain Partial, not passes. An initial 10-second wait timed -out before the observed bridge retirement; the subsequent attempt waited for -retirement. No raw image or grant is retained in this reference. Driver -shutdown, app cleanup, and the failed Win32 app-run reports are -recorded in `packages/agentic/desktop-test/references/current-status.md`. - -## Windows Fabric and Win32 Paper — #4441 physical MCP click measurement — 2026-11-12 UTC - -**Observed:** revision `505a2187b0a2451aed7081c9260b343d55742f93`, -Windows 11 Enterprise build 26300 x64, Node 26.13.0. Packaged Windows -Fabric reused the native build at `bdeb0b291` after confirming no native -implementation differences; Win32 Paper used REX 0.81.1. Metro/catalogs/ -bundles came from the tested checkout. Each run held the W3 campaign and -machine lock with `-RequirePhysical`. Before/after gates were eligible, -unlocked **remote RDP**, one 1728 x 1084 monitor at origin zero, 96 DPI / -100%; this was not console-session qualification. - -### Method and missing click cost - -Real MCP SDK **1.32.2** stdio connection through the private descriptor and -native-verified service; tokenizer **gpt-tokenizer 4.1.0 / o200k_base**. -After selecting `components-checkbox--default`, each endpoint made six -actual `click` actions, each followed by a checked-state assertion and an -**unscoped default observe**. Every receipt was applied/physical-pointer; -no semantic fallback counted. All six state assertions passed and the -checkbox ended unchecked. Each connection made 23 tool calls, zero error -responses/readiness rejections, and closed its session. These runs did not -exercise retries or cancellation. - -Discovery serialized the six-tool array using the November 8 baseline method: -**4,691 tokens / 20,277 UTF-8 bytes (characters)**, once per connection. -The full `tools/list` result envelope was separately 4,693 tokens / 20,287 -bytes; server instructions added 120 tokens / 591 bytes. Discovery has not -been omitted from the accounting. - -| Measurement, six samples per endpoint | Windows Fabric | Win32 Paper | -| ------------------------------------------------------ | -------------------------------------- | -------------------------------------- | -| Click response text, all blocks | **99–106 tokens / 313–314 bytes** | **96–103 tokens / 311–312 bytes** | -| Click request JSON, separately | 56–63 tokens / 186 bytes | 57–64 tokens / 186 bytes | -| Request + response-text pair | 155–169 tokens / 499–500 bytes | 153–167 tokens / 497–498 bytes | -| Click RPC p50 | **299.107 ms** | **308.003 ms** | -| Click RPC p90 / max | 369.976 ms | 318.208 ms | -| Click + checked check + default observe, response text | 1,999–2,005 tokens / 5,537–5,546 bytes | 1,585–1,592 tokens / 5,302–5,309 bytes | -| Same three RPCs, summed elapsed time | 411.849–550.601 ms | 431.774–473.477 ms | -| Default observe completeness | Incomplete: maxNodes, 13 omitted | Complete, no truncation | -| Default observe emitted / matched / visited | 100 / 113 / 115 | 99 / 99 / 124 | - -Response text blocks were independently encoded; request counts encode -user-facing `JSON.stringify({name, arguments})`, **not** JSON-RPC wire or -model-billing totals. Opaque identifiers cause normal token variation. -Percentiles use nearest rank; with six samples p90 is the maximum. -`rpcElapsedMs` stops immediately after SDK `callTool` resolves, before -tokenization, excluding service/app startup and discovery. Raw artifacts -also retain encoding-inclusive elapsed time: p50/max 299.192/370.036 ms -(Fabric), 308.128/318.281 ms (Paper). Earlier encoding-inclusive samples -were not silently substituted for this RPC distribution. - -Default budgets on both were deadline 3000 ms, maxBytes 12288, maxDepth 24, -maxNodes 100, maxStringLength 120, maxVisited 1500. No reduced scope or -budget was used to claim a saving. These six local samples complete the -previously missing physical-click row; they are not a controlled speedup -comparison with November 8's `activate` tuple or a platform ranking. - -### Commands, evidence and qualification limits - -Exact supervisor invocations from the qualification root, inside -`with-native-lock.ps1 -Lane W3 -RequirePhysical`: - -```powershell -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --mcp-only --with-secondary -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs win32 --mcp-only -``` - -The retained supervisor owns declared Storybook driver/bundle/run commands, -exact app identity, MCP child connection and cleanup. It is local -qualification tooling, not a new published CLI surface. Artifacts are -relative to that checkout: -`artifacts\desktop\w3-qualification\runs\\mcp-click-.json`. - -| Endpoint | Qualification ID | Driver run / MCP session | Physical session gates, UTC | -| -------- | -------------------------------------- | ------------------------------------------------------------------------------- | --------------------------- | -| Fabric | `f188fae5-67c9-47e4-9800-ea584185fed9` | `57d4f6cb-2dc3-4b00-bb95-4331efd8d34b` / `b739a296-4647-4c34-bb80-a8f898fddab3` | 15:15:13.606–15:15:47.712 | -| Paper | `6d474fd4-61cb-41f6-824e-7a80df68db45` | `918930ba-e498-4e20-ad98-52e4187b5618` / `6df40ce1-32db-44d1-a9ba-64cc2dc6ec90` | 15:14:30.979–15:14:57.843 | - -Both MCP cases passed independently. Fabric's enclosing run also retained -a **failed** secondary semantic-effect assertion; its overall exit 1 is -not relabeled passed. A later exact-secondary physical click passed, but -semantic activation reproduced applied/accessibility with actual status -`Idle`. Win32's visible untitled tool-window popup was not separately -registered. These are **Observed** open gaps being fixed on this branch; -accessibility-pattern dispatch and the untitled-toolwindow registration -filter are **Inferred** causes, not proven fixes. - -200% was **not run** (100% scale; changes unauthorized), mixed DPI was -**not run** (one monitor), negative-origin layout was **not run** (only an -origin-zero monitor), and physical console Ctrl+C was **not run** (separate -implementation owner). Full per-case native results are in -`@fluentui-react-native/desktop-test/references/current-status.md`. -Sessions, applications and owned services drained; the campaign ended and -pre/post-end audits passed: -`FURN-AUDIT: no FURN desktop processes found`. - -## F15 macOS Paper measurement prerequisite — 2026-11-12 - -Source `505a2187b0a2451aed7081c9260b343d55742f93`; macOS 27.2.1 -(26B131), arm64. The stable-signed helper's noninteractive doctor reported -Accessibility=false and PostEvent=false, with ScreenCapture=true. The -driver ledger at `packages/agentic/desktop-driver/references/current-status.md`, -section "F15 non-AX macOS supporting evidence — 2026-11-12", -retains the exact diagnostic command and signing identity; its local artifact -is `artifacts/desktop/f15-20261112/doctor.json`. - -The initial pass was **blocked**: macOS Paper MCP interaction/text/latency -measurements needed Accessibility/PostEvent. That prerequisite was resolved -later the same day; it is no longer an owner-action block. -In the initial pass no Paper Storybook app or MCP session was launched, no measurement command -was run, and 0 tests were selected/executed (no new pass, failure, or skip -counts). Cleanup is N/A for MCP; the diagnostic helper exited. The non-AX -helper self-test does not qualify the adapter or change historical measurement -claims. - -### F15 permission-enabled MCP follow-up — 2026-11-12 - -**Observed, partial; not macOS requalification.** Source: -`141b88b4b81a1bd9a0472de533681f9274d8cb00`; macOS 27.2.1 (26B131), -arm64. App: Agentic Components Storybook / ReactTestApp, bundle -`com.microsoft.fluentui.agenticstorybook.id915e78595`. -Fabric target: `macos-a6dfeaccb13e643a`; Paper target: -`macos-de43c71d45415315`. Nonprompting doctor reported -Accessibility/PostEvent/ScreenCapture true; no privacy changes were made. -All native/GUI commands used the machine-wide native lock and the stable -`FURN Desktop Driver Development` identity/cache described in the driver -ledger's F15 follow-up. - -The local evidence root is `artifacts/desktop/f15-20261112T115411Z/`. -The session-local probe used the public MCP SDK 1.32.2 and shipped -desktop-driver-mcp 0.1.3 stdio CLI with the real private -`agent-connection.json` descriptor. No grant or descriptor contents were -printed or copied to evidence. Tokenizer: `gpt-tokenizer` 4.1.0, -`o200k_base`; compact JSON for discovery, newline-joined MCP text blocks for -operation text. UTF-8 bytes and JavaScript characters were recorded -independently. These bounded samples are not a controlled cold-start -benchmark, reliability window or cross-client qualification. - -Fabric command, from the repository root: -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs fabric lifecycle`. -It starts `yarn storybook driver --macos --agent --reporter json`, then -`yarn storybook run --macos --reporter json` in `apps/storybook`. -Results: **10 passed, 1 failed, 0 skipped manual checks** in -`fabric-lifecycle-qualification.json`. Revocation, dispose-during-open, -ambiguous CREATE and failed DELETE before/after dispatch used the SDK -in-memory transport with the real native service and verified descriptor: -response loss was injected at the public typed-client method boundary, -not with a TCP proxy or by forcing native cleanup failure. Each connection -drained and the target reservation could be reacquired. - -The stdio pass observed same-port service replacement, authoritative -invalidation of the old session, fresh verified discovery, `no-session` -refusal of the old mutation, target capture/resource reading, session-close -resource expiry and read cancellation. The first group-SIGTERM harness -attempt retained a descriptor and failed restart; it remains recorded in -`fabric-qualification.json`, with exact-record reconciliation in -`descriptor-reconciliation.json`. It is not a successful ordinary shutdown. - -Fabric's physical checkbox click **failed**, not passed: the sequence stopped -before dispatch (`element click intercepted`, `obscured`, effect `none`). -Its 107-token / 438-byte / 1,187.5-ms response measures a refusal. -No successful physical click/check/observe interaction cost is claimed. -The driver's F15 open finding records the reproduction and routing. - -| Fabric measurement | Observed result | -| -------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | -| Six-tool discovery | 4,693 tokens; 20,287 UTF-8 bytes / characters | -| Server instructions, additional to discovery | 120 tokens; 591 UTF-8 bytes / characters | -| First attach after discovery | 846.4 ms; one sample, excludes service/app startup | -| Default observe, 20 calls per story | Median 21.9–27.8 ms across stories; p90 1,127.7–1,669.0 ms; maximum 1,680.2 ms; 905–948 tokens, 2,778–3,061 bytes per call | -| Idle delta, 20 calls per story | Median 21.5–26.3 ms; p90 1,122.0–1,671.9 ms; maximum 1,688.1 ms; 151–152 tokens, 464–469 bytes per call | -| Read cancellation | Caller aborted after 301.5 ms; subsequent observe passed; no mutation cancellation claim | - -The three stories were `components-button--appearance`, -`components-checkbox--constrained-text` and `components-menu--default`. -The default 100-node projection completed in all 60 observations and all -60 deltas; no retries. The slow maxima are retained rather than hidden by -the medians. These samples do not establish equal information content to -historical Windows or other story projections. - -#### Paper descriptor, interactions and completeness - -From the repository root, -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper` -used `yarn storybook driver --macos --paper --agent --reporter json` -and `yarn storybook run --macos --paper --reporter json`. -It recorded **14 passed, 4 failed, 0 skipped manual assertions** -(`paper-qualification.json`), not selected suite tests. The four failures -were keyboard clear, its dependent semantic-clear precondition, refused -physical Checkbox click and unsupported semantic Checkbox activation. -The latter two sequences were `stopped` despite non-error MCP envelopes; -they were not accepted as interactions. - -The same real-descriptor lifetime/fault boundaries described for Fabric -passed on Paper, including same-port invalidation, fresh discovery, -`no-session` old-mutation refusal, read cancellation followed by a usable -session, target PNG resource reading/hash verification and expiry on close. -The exact app PID/start-time/executable survived session and service close. -Paper's channel survived one 22-second idle interval. The fault probes -wrapped public typed-client methods, not native cleanup or TCP failures. - -Discovery independently matched Fabric: six tools, **4,693 tokens / -20,287 UTF-8 bytes / characters**, plus 120 tokens / 591 bytes of server -instructions. First attach was 831.6 ms, excluding service/app startup. -All latency values below are client wall time; p90 uses nearest rank. -Each row has 20 calls, zero retries and zero call failures. Defaults used -the 100-node projection, not a full-tree request. - -| Paper story / operation | p50 / p90 / max, ms | Text tokens | UTF-8 bytes | Complete projections | -| ------------------------------------------------ | ------------------------ | ----------- | ----------- | -------------------- | -| `components-button--appearance`, observe | 31.8 / 1,142.7 / 1,689.3 | 1,188–1,189 | 3,533–3,537 | 20 / 20 | -| Same story, delta | 31.5 / 1,152.0 / 1,671.9 | 151–152 | 465–467 | 20 / 20 | -| `components-checkbox--constrained-text`, observe | 34.1 / 1,688.0 / 1,698.8 | 1,201–1,202 | 3,750–3,752 | 20 / 20 | -| Same story, delta | 33.9 / 1,696.4 / 1,701.6 | 151–152 | 465–467 | 20 / 20 | -| `components-menu--default`, observe | 34.1 / 1,158.7 / 1,706.5 | 1,313–1,322 | 3,862–3,866 | **0 / 20** | -| Same story, delta | 36.4 / 1,159.7 / 1,710.1 | 151–152 | 468–470 | 20 / 20 | - -**Observed:** each Paper Menu observation emitted 100 of 103 matching -nodes, `maxNodes`, three omitted; these are incomplete observations, not -full-information comparison samples. Complete deltas do not repair that -truncation. Fabric's complete projections and these Paper projections are -not equivalent scopes. Window dimensions and display scale were not -retained for this run; the cached stable helper and running app make these -warm operation samples, not controlled cold-start measurements. -Paper read cancellation took 302.2 ms; post-dispatch mutation cancellation -was **Not run**. - -The Paper physical-click refusal cost 1,172.4 ms / 106 tokens / 438 bytes; -the provider's **Unsupported** semantic activation cost 743.9 ms / -79 tokens / 329 bytes. Neither is a successful click/check/observe cost. -The independent -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper input` -probe recorded **4 passed, 1 failed, 0 skipped manual checks**: -semantic clear, physical-input delivery, protected args and the following -real stdio MCP interaction passed; Paper keyboard clear still failed after -bounded polling. Exact harnesses and defect reproduction are retained in the -driver and consumer ledgers. - -| Successful Paper Input interaction, one sample | ms | Tokens | UTF-8 bytes | -| --------------------------------------------------------------- | ------- | ------ | ----------- | -| `desktop_act` type `F15`, keyboard mechanism, applied/completed | 837.5 | 74 | 261 | -| `desktop_check` value equals `F15`, passed on first read | 30.1 | 96 | 322 | -| Closing default `desktop_observe`, complete, 90 emitted nodes | 1,148.2 | 1,201 | 3,611 | -| Sum of those three calls | 2,015.7 | 1,371 | 4,194 | - -This sum excludes discovery, attach, story reset and close. No mutation was -retried; each measured call ran once. It is a successful text-entry/value -interaction, **not physical Checkbox activation** and not directly comparable -to historical toggle costs or projections. Artifact: -`paper-input-qualification.json`. - -**Cleanup passed:** `cleanup-audit.json`, 2026-11-12T12:20:26Z, records no -worktree processes, no listeners on 39860/23935/28029, no descriptor and -no per-user workload claim; the native-lock wrapper exited and the lock -was unheld afterward. Reliability, durable TCC continuity, other MCP clients, -native cleanup/survivor faults and the outstanding F3/F4/F5 acceptance -cases listed in the consumer ledger remain **Not run** on this source. -Both renderer smokes failed; the permission prerequisite is resolved but -#4441's macOS acceptance is not met. - -## F11 explicit relinquish — 2026-11-12 - -Source `5c9d0191e` includes explicit `desktop_session` relinquish with an -immediate admission fence, cancellation, and awaited connection-subtree -drain. A failed drain retains retryable cleanup responsibility. A fresh -session after relinquish rejects old references, cursors, proofs and -artifact links; no cross-connection resume or replay is added. -Authoritative rejection discards the cached child binding. After operator -preemption and runner drain, the same MCP connection delegates a fresh child; -old references, cursors, proofs and artifact links remain invalid. -Package format, lint, build, and all 32 tests passed, including -deterministic resource invalidation, late/lost acquisition and failed-drain -cases. The default-`TMPDIR` focused server selection passed 31 cases, -including the preempt → runner drain → fresh MCP session regression. -These are portable semantics, not native qualification. - -The real private-descriptor AppKit consumer at `87bc8b710` reached verified -discovery but failed native session creation on macOS 27.2.1 arm64: -`The target does not support any click input mode.` The stable-signed -helper doctor reported Accessibility/PostEvent `false`; no privacy settings -were changed. Native relinquish/handoff and worker-crash drain did not run. -Both descriptors and the workload claim were removed; the externally owned -fixture survived service shutdown and only its spawning test stopped it. -The Accessibility-blocked native contract was not repeated for review fixes. - -Exact commands, revisions, one corrected-fixture retry, cleanup evidence -and local artifact locators are in -`@fluentui-react-native/desktop-test/references/current-status.md`. -Windows/Win32 native handoff is not covered by the existing PR smoke/E2E -selection and remains not run. Storybook integration and retained-launch -ownership remain separate follow-up phases. - -## Windows Fabric and Win32 Paper — F11 handoff qualification — 2026-11-12 UTC - -**Observed:** source `c9753208d0f5602d28510d84ededf3f2fc3c3855`; -Windows x64, active/unlocked RDP (`sessionKind: remote`). Live runs were -inside W1's campaign and the machine-wide `with-native-lock.ps1 -Lane W1 --RequirePhysical` wrapper; before/after session gates stayed eligible. -No physical-input delivery is claimed. - -The real MCP SDK/server used `connectWithDescriptor`, native Windows listener -verification and a separate attenuated controller child. The generic -`DesktopSupervisor` published agent/runner descriptors in the validated -default private store. It served externally owned apps, not the unsupported -Storybook integration handoff. From the worktree root, the retained -reproduction commands are: - -```powershell -node artifacts\desktop\f11\w1\qualify.cjs windows -node artifacts\desktop\f11\w1\qualify.cjs win32 -``` - -Windows attached to the exact registered packaged Storybook AUMID (Fabric), -with Metro from this worktree using the app's declared -`yarn start --no-interactive --port 8081`. The registered build at -`bdeb0b291` was reused because the native-source comparison found only -native-core changelog/version metadata changes. Win32 attached to the exact -resolved prebuilt REX `ReactTest.exe` (Paper), after the app's declared -`yarn storybook bundle --win32` passed. The public native launcher maintained -application ownership separately from the agent service. - -| Native MCP/runner case | Windows Fabric | Win32 Paper | -| --------------------------------------------------- | --------------------------------------------------------------- | --------------------------------------------------------------- | -| Runner defaults to busy while MCP owns a session | Passed (`target-busy`) | Passed (`target-busy`) | -| `desktop_session` explicit relinquish | Passed; returned confirmed relinquish | Passed; returned confirmed relinquish | -| Operator-authorized preemption | Passed; old adapter binding refused, then runner-exclusive busy | Passed; old adapter binding refused, then runner-exclusive busy | -| Normal runner drain and fresh MCP reopen | Passed | Passed | -| Identity-checked, literal-PID worker kill and drain | Passed recovery; intentionally failed WDIO run | Passed recovery; intentionally failed WDIO run | -| Lost real HTTP 200 runner-grant response | Passed; parent drain restored admission | Passed; parent drain restored admission | -| Old reference/cursor/capture-link rejection | Passed | Passed | - -During preemption the old adapter operation reported `service-unavailable`; -an attempted open then returned `target-busy` during runner ownership. -After runner cleanup, the same MCP connection opened a new child/session. -No old session or resources were resumed or replayed. Lost-response injection -discarded the real native service's acquisition-response body before the typed -client received it; it did not substitute a fake listener. - -| Endpoint | Local evidence root | Normal / preempt / crash run IDs | -| -------------- | ----------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ | -| Windows Fabric | `artifacts\desktop\f11\w1\windows-3tYoml` | `a3bf02e2-df22-4a63-92f8-686057ef2a80` / `24d38403-b4e5-49d6-9e0f-d1d123e4fb17` / `8f702133-e18a-41f0-b05d-ce5fcb5cff1a` | -| Win32 Paper | `artifacts\desktop\f11\w1\win32-BefbqE` | `d79fd672-df2d-4943-8b40-c017a8978380` / `f1f17781-08aa-4cac-a6d8-34cb13bd749b` / `a09a8c20-1380-4933-a7a3-9c2d786352c3` | - -Each root retains `evidence.json`, `run.json`, command logs and crash-process -identity. Normal/preempt runs each have one selected/executed/passed test, -zero skips and `terminal: true`. Crash runs are terminal **failed**, with one -selected, zero completed and one incomplete test: killing workers `34752` -(Windows) and `29856` (Win32) intentionally prevented the final reporting -handshake. Parent cleanup completed and fresh MCP attach proved reuse. -Both service lifecycle reports are terminal passed; both service workload -claims and descriptors were removed. The same external app process -incarnation survived service shutdown, then its recorded launcher closed it. -No processes survived; audit: `FURN-AUDIT: no FURN desktop processes found`. -The campaign is ended. - -Native worker environment checks found no broker/worker credential variables; -model output and a scan of 45 retained text/config/report files contained no -bearer credential. This is observed confinement, not isolation from hostile -same-account code or Node preloads. Helper artifact: -`d8ee05e4796fb3aa22f7b5dbf481bd8cdfb0c20b6d5d24676d689dc7900d5b3c`. - -**Observed, supporting checks:** `yarn build` passed. With native opt-ins unset, -uncached `yarn workspace @fluentui-react-native/desktop-driver-mcp test ---runInBand` passed all 32; driver passed 1,211/skipped 88; harness passed -376/skipped eight. Logs are under `artifacts\desktop\f11\w1`. The additional -Windows native identity/backend selection passed 23 and skipped 19 macOS -cases, including all 15 Windows cases skipped by the earlier macOS selection. -The protected-service selection passed three; its recovering-quarantine case -was **failed/environment-blocked** by project scratch ancestry (`ancestor`). -Default protected storage opened and supported both live runs; no ACL change -was made. Windows permission diagnostics are **unsupported**, not passed. - -**Remaining CI gap:** desktop-test's native MCP/WDIO consumer is darwin-gated, -with a macOS-only fixture. On Windows, `FURN_NATIVE_DRIVER_TEST=1 yarn -workspace @fluentui-react-native/desktop-test test nativeRunner.contract ---runInBand` skips its sole case, including a locked repeat with the fixture -path opt-in naming the installed REX executable. Both plain endpoint -`desktop-driver doctor` commands passed (`ready: true`); their `--permissions` -variants are unsupported. Driver native contracts/Storybook smoke -selection do not cover this consumer handoff. A Windows fixture and explicit -consumer selection are handed back for implementation; no source or CI fix -was made here. The existing three-package `scoped-runner-handoff` patch -changeset remains. Storybook handoff, retained launches and cross-connection -resume/replay remain separate phases. - -**Evidence validation:** `yarn markdown-link-check -c -.github\markdown-link-check-config.json -packages\agentic\desktop-driver-mcp\references\current-status.md` passed -(six links). Root `yarn oxfmt --check`, `yarn change:check` and -`git diff --check` passed. No clean-cache `buildci` rerun was required for -these evidence-only markdown edits. - -## Windows native consumer and physical serve cleanup — 2026-11-12 UTC - -**Observed, tested source `b1bbb26df880252a184ea5a3d2e1848bd811e83f`:** -the Windows generic private-descriptor MCP/WDIO consumer now executes against -an explicitly provisioned plain Win32 fixture. In the granted W1 campaign, -with an active unlocked RDP session (`sessionKind: remote`), native opt-in and -the exact fixture path, the locked command -`yarn workspace @fluentui-react-native/desktop-test test -nativeRunner.contract --runInBand --json --outputFile -\artifacts\desktop\f11\w1-native\native-runner-jest.json` **Passed** all -five outer cases, with zero skips. - -| MCP/runner mode | Result | Native WDIO run | -| ---------------------------------------------------- | --------------------------------------------- | -------------------------------------- | -| Explicit handoff and fresh MCP reopen | Passed, 2/2 inner specs | `a9849169-7f2b-46e6-87de-3e0b8b8e7a3f` | -| Operator preemption and exclusive runner reservation | Passed, 2/2 | `2a1f63bd-3ac0-4951-beb5-860a2a92b69c` | -| Worker crash and confirmed parent drain/reuse | Passed recovery; expected terminal failed run | `9c1a8447-2a6c-4ca6-b027-8bc302df3fc7` | -| Lost real HTTP-200 runner-grant response | Passed, 2/2 | `4e7d50fa-b055-407c-bfe2-443fe2bc504f` | -| Target capture and stale-resource invalidation | Passed, 2/2 and real capture | `047e8166-c3d2-4fb9-b785-288519b3342d` | - -Every non-crash mode executed the accessibility-input spec, rather than -silently skipping it. Each mode preserved the exact attached fixture -incarnation through service cleanup, removed both private descriptors and -the claim, then ended only its own fixture. Credential confinement and -fresh-session/stale-resource checks passed. This plain fixture qualifies -generic native ownership, not a React Native renderer or Storybook runner -handoff. The separate Windows driver identity/backend selection passed -24 and skipped 19 macOS cases. - -The previous hosted consumer in workflow `43284867711` ran on both Windows -job `131415841258` and Win32 accessible job `131415842551`: handoff, -preemption and crash recovery passed, but lost-response failed before WDIO -because it compared the injected error's message with the correctly wrapped -transport error. Capture was policy-skipped. The current harness correction -checks the typed failure's original cause and shares the actual successful -response-loss injection with the portable HTTP runner. Its native -lost-response pass above confirms that reservation/drain/reopen path; no -adapter runtime or error-redaction change was needed. - -**Observed, real Storybook consumers:** physical Ctrl+C also **Passed** -after authenticated `components-button--appearance` native-marker selection -and MCP observation on Windows Fabric and Win32 Paper. The MCP session and -adapter were closed before interruption. An isolated, visible, focused -console received four `SendInput` key events and recorded console `SIGINT`; -no `GenerateConsoleCtrlEvent` or synthetic Node signal was used. The exact -CLI, Yarn and observer each exited 0; both lifecycle reports are terminal -passed, descriptors were withdrawn, claims released and apps preserved until -their external launcher closed them. - -Fabric lifecycle: `75c40800-571c-4d8f-9fa1-debf876f0664`. -Paper lifecycle: `9a5c618f-d54d-4599-b36f-e3fd5b5436b9`. -Reports are under `apps\storybook\artifacts\desktop\`. -The unchanged registered Fabric native build was reused after identity/source -comparison. Paper used a serving-instance/driver-manifest-matched bundle -and the prebuilt REX host; its native build remains **Unsupported**. - -Every native/input command used -`with-native-lock.ps1 -Lane W1 -RequirePhysical`. Evidence and reproduction -are retained under `artifacts\desktop\f11\w1-native`, including -`contracts.json`, `physical-windows-UmfSj6`, `physical-win32-sssEyg` and -the session/lock/audit receipts. No private descriptor was copied into those -artifacts. The campaign is ended with zero W1 native/GUI survivors. The broad -audit's four remaining candidates are foreign portable W5 Jest/Yarn processes, -with no listeners; none was terminated. Corrected-head hosted confirmation is -pending publication at this receipt, not inferred from earlier green checks. - -**Portable closeout:** locked default-environment -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` **Failed** -three fresh attempts on the same Storybook integration readiness test, -`hosts the Storybook channel and Stage 1 smoke-test target in one process` -(`A stopping command cannot publish readiness.`). The test and lifecycle -source are unchanged from main/review base; the full owning package's -default-concurrency `yarn workspace -@fluentui-react-native/storybook-desktop test` **Passed** 340 with three -skips. No unrelated source/deadline or environment workaround was added, -and no passing aggregate is claimed for this head. Exact failures and -source comparison are retained under the same evidence root and in the -harness ledger. The last aggregate began with a clean audit, so the failure -is not attributed to the foreign portable chain. Owned portable workers -exited naturally; that foreign owner removed its abandoned processes. -Final audit **Passed**: `FURN-AUDIT: no FURN desktop processes found`. -Changed-ledger format, instruction and link checks also passed. - -## Windows default-aggregate follow-up — 2026-11-12 UTC - -**Observed at `2bf750f9e`:** locked default -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` **Passed** -178 tasks, zero failed, at 20:52:21–20:55:29 UTC. No source, test deadline -or environment/concurrency workaround changed. Detached main first failed -a different driver cancellation/reuse assertion, then its unchanged default -repeat passed 178; the earlier Storybook readiness error's exact cause is -not proven pre-existing or attributed to campaign state. - -No agent/runner descriptor, active claim or generated lease survived. -Only two exact exited campaign application epochs were compare-and-removed -through verified native APIs; unrelated recovery records were retained. -The scratch worktree was removed, no app/OS setting changed, and final -process audit **Passed**, no FURN processes found. The owning harness -receipt in `packages/agentic/desktop-test/references/current-status.md` -records locations, timestamps and the causal-evidence limits. - -Hosted workflow `43297198219` finished **successfully**, zero failed jobs: -Windows `131458546719` and Win32 accessible `131458546805` each executed -four passing ownership modes with one capture-policy skip. Handoff, -preemption and lost-response each passed both inner specs, zero skips, -with `accessibilityClick: passed`. This confirms hosted generic native -consumer coverage; it does not broaden the earlier physical Storybook -or macOS evidence. - -## F11 rebased macOS positive qualification — 2026-11-12 UTC - -**Observed at `5dbadb53b10cd81ccf1e15739e213f0046d93323`:** real native -descriptor verification, MCP SDK transport, and WDIO/Jasmine passed all five -generic AppKit modes: relinquish/handoff, operator preemption, worker crash, -lost successful runner-grant response, and target capture. Every non-crash -mode passed both inner ownership/accessibility-input specs with zero skips; -the intentional crash's terminal failed inner run was drain-confirmed before -a fresh MCP session. Stale resources and credential confinement passed. -All fixtures were preserved through service shutdown, descriptors removed, -claims released, and exact spawned fixtures subsequently ended. - -The stable-signed helper's noninteractive doctor reported Accessibility, -PostEvent and ScreenCapture **true**, with no prompt. This supersedes the -earlier macOS permission-blocked attempt, not the preserved Windows receipts. -The rebase retained all Windows additions from -`11820d6330f4ea591bed400193b13e793970b5bd` onto -`041ac8b9894aeaa7e1aa273baa2ece4e46298b94`; no runtime fix was needed. -MCP format, lint, build and **32 portable tests passed**; focused server tests -passed **31**. See `packages/agentic/desktop-test/references/current-status.md`, -under "F11 rebased macOS positive qualification — 2026-11-12 UTC", for exact -commands, native identities, run IDs and cleanup. -Storybook/F9, retained launches/F10, physical input, general MCP-client -qualification, cross-connection resume and replay remain separate. -Final portable aggregate acceptance remains **blocked** by the unchanged -driver bounded-command `SIGKILL`/`SIGTERM` assertion under plain default -parallel execution, despite passing owning/MCP suites and native qualification. -The owning harness receipt retains both failures, the additional first-run -private-storage timeout, and the passing focused diagnostic; no unrelated -test or environment workaround was applied. +Native and Storybook limitations are tracked in `@fluentui-react-native/desktop-driver/references/current-status.md`; supervision gaps in `@fluentui-react-native/desktop-test/references/current-status.md`. + +## Status overview + +The adapter's tools, connection lifetime, grants and relinquish are implemented and covered by deterministic tests. +Native qualification through the real descriptor path is partial and per endpoint. +Management and workload tools are not part of this adapter. + +- **Tools and shared schemas:** Done. Six tools over shared driver schemas. See [Tools and discovery](#tools-and-discovery). + +- **Descriptor, identity and grants:** Done. Native identity is verified before a grant is sent, and each connection gets an attenuated child. See [Descriptor, identity and grants](#descriptor-identity-and-grants). Tracked by [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Connection and session lifetime:** Qualified (generic AppKit consumer on macOS, generic native consumer on Windows and Win32). Storybook lifecycle faults are not qualified. See [Connection and session lifetime](#connection-and-session-lifetime). Tracked by [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441) and [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451). + +- **Relinquish and handoff:** Done. Qualified for the generic consumer on macOS Fabric, Windows Fabric and Win32 Paper. See [Relinquish and handoff](#relinquish-and-handoff). Storybook handoff is tracked by [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Storybook control:** In progress. The experimental selection and marker path exists. See [Storybook control](#storybook-control). Tracked by [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Cross-connection resume:** Unsupported. See [Cross-connection resume](#cross-connection-resume). + +- **Workload and test orchestration tools:** Deferred. See [Workload and test orchestration tools](#workload-and-test-orchestration-tools). Tracked by [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450). + +- **General MCP-client qualification:** In progress. SDK stdio evidence only. See [MCP client qualification](#mcp-client-qualification). Tracked by [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Native qualification:** In progress. See [Native qualification](#native-qualification). Tracked by [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Measurements:** In progress. Latest values are recorded for all four endpoints except macOS Paper click and cancellation cost. See [Measurements](#measurements). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +## Tools and discovery + +### Behavior and limits + +- The adapter exposes six tools over shared driver schemas: session, observe, act, check, capture, story. +- There is no separate execution path from the driver. +- Launch-mode discovery lists the effective launch and attach modes, so a launch-only target never fails a predictable attach-first call. +- Progress and cancellation are supported; progress does not renew idle sessions. +- Image artifacts are paged resources that are bound to the session and hash-checked. +- The schema test limits serialized JavaScript string length to 22,000 characters; this is neither a token nor a UTF-8 byte limit. +- Discovery is about 20.3 KB; the earlier 8 KiB discovery budget is not met. +- Changed descriptions or schema text invalidate an exact token claim; keep token counts tied to their revision. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): remaining MCP cost measurements (macOS Paper click and cancellation cost). + +## Descriptor, identity and grants + +### Behavior and limits + +- Preview element references go stale after reset or the next select; chrome and window references and immutable PNG resources keep their identity. +- Manifest smoke and legacy traces are unprotected workloads that use no descriptor or grant, so they are not protected-ingress evidence. +- The adapter verifies the private descriptor and native identity before sending a grant. +- It imports the driver's `/native`, `/authority` and `/client` subpaths, not the WebdriverIO-heavy root; boundary tests guard this. +- The model supplies no URL, token, configuration or executable input. +- Each controller delegates a distinct driver-only child after verification, with target and scope subsets. +- Terminal-only cleanup is retried. +- Invalid authority is forgotten and the next call re-verifies discovery; mutations are never replayed. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): live descriptor-path requalification of per-connection grants on every endpoint. + +## Connection and session lifetime + +### Behavior and limits + +- Acquisition, close, operations and resource reads are serialized single-flight; a late acquisition after dispose is cleaned up. +- After a same-port protected restart or revocation the next call re-verifies discovery. +- Same-authority transport retry keeps the known session cleanup responsibility and rebinds only after verifying the same incarnation and principal. +- Same-authority retry is not cross-connection resume. +- A lost create response is fenced as `session-uncertain` with no repeated acquisition. +- Connection disposal revokes and drains the child, including unknown session ids. +- A failed delete or disconnect keeps known-session responsibility; the bounded dispose wait leaves late cleanup attached. +- Cleanup against an unreachable service cannot be guaranteed, so service drain and expiry remain necessary. +- Fake-host tests validate portable controller semantics, not native ownership, signing, TCC, Windows ACLs, DPI or process containment. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): descriptor-path reruns of same-port replacement, revocation, dispose during open, failed delete, ambiguous create and resource invalidation on the endpoints not yet qualified. +- [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451): repeated-run reliability and failure-path support matrix. + +## Relinquish and handoff + +### Behavior and limits + +- `desktop_session` relinquish applies an immediate admission fence and cancellation, and returns only after the connection subtree drains. +- A failed drain keeps retryable cleanup responsibility. +- A fresh open expires old references, cursors, proofs and artifact links. +- Authoritative rejection discards the cached child binding; after operator preemption and runner drain the same connection delegates a fresh child. +- The generic attach-only runner integration is owned by `@fluentui-react-native/desktop-test/references/current-status.md`. +- Handoff serves externally owned apps; Storybook integration handoff is not part of it. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): Storybook handoff and preview lifecycle. + +## Storybook control + +### Behavior and limits + +- The driver selection and marker path exists and is experimental. +- Channel ingress, reconnect and full exclusive context authority are incomplete. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): channel ingress, reconnect and exclusive context authority. + +## Cross-connection resume + +### Behavior and limits + +- There is no lease, grace or ownership-transfer protocol, by design. +- Same-controller rediscovery retains its existing child. +- There is no automatic replay or cross-connection resume. + +## Workload and test orchestration tools + +### Behavior and limits + +- These tools belong to a future harness control or client surface; the adapter never becomes the resource owner. +- Keep this package name until harness-owned management tools land. +- A later `desktop-mcp` facade must retain driver-only use and plan published compatibility. + +### Open work + +- [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450): generalized management facet and transport. + +## MCP client qualification + +### Behavior and limits + +- Evidence comes from the MCP SDK 1.32.2 stdio client; it is not universal client qualification. +- Client-specific resource, image and cancellation behavior needs independent runs. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): independent runs for supported clients. + +## Native qualification + +Results are per endpoint and tied to the revision that produced them. +Skipped or not-run checks are not passes. +Full per-run evidence (commands, counts, run ids, artifact locators, cleanup outcome) lives in the linked pull request or issue. + +### macOS Fabric + +- **Generic AppKit consumer (relinquish/handoff, operator preemption, worker crash, lost runner-grant response, target capture):** all five modes passed through the real descriptor, MCP SDK transport and WDIO/Jasmine at `5dbadb53b`, with stale-resource and credential-confinement checks passing. Evidence: [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Storybook descriptor lifecycle and faults:** 10 passed, 1 failed, 0 skipped at `141b88b4b`; the failure is the physical Checkbox click (`element click intercepted`, effect none). Fault injection used public typed-client methods, not native cleanup or TCP failures. Evidence: [#4493](https://github.com/microsoft/fluentui-react-native/pull/4493). [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500) fixed driver physical-input postconditions but records no MCP-path rerun. + +### macOS Paper + +- **Storybook descriptor lifecycle and interaction:** 14 passed, 4 failed, 0 skipped at `141b88b4b` (keyboard clear and its dependent precondition, physical Checkbox click, unsupported semantic Checkbox activation); a separate input probe passed 4 and failed 1 (keyboard clear). Cleanup audit passed. Evidence: [#4493](https://github.com/microsoft/fluentui-react-native/pull/4493). No MCP-path rerun followed [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500), so macOS acceptance for [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441) is not met. + +### Windows Fabric + +- **Physical MCP click:** six physical-pointer checkbox clicks, each state-checked, with 23 calls and zero errors at `505a2187b`; remote RDP session, 100% scale, one origin-zero monitor. Evidence: [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Generic consumer handoff, preemption, worker kill and lost runner-grant response:** the plain Win32 fixture passed 5/5 with zero skips at `b1bbb26df`; this qualifies generic native ownership, not a React Native renderer or Storybook handoff. Evidence: [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Connected descriptor-path checks:** physical checkbox and target PNG passed at `fec94cec4`. Evidence: [#4469](https://github.com/microsoft/fluentui-react-native/pull/4469). + +- **Descriptor lifecycle:** same-port replacement, revocation, dispose during open, failed delete, ambiguous create, quarantine and recovery, protected args, cancellation replay and idle-channel lifetime passed with the real descriptor at `1f5333e25` plus the follow-up fixes `8458bde48` and `a6b89e924`. Evidence: [#4463](https://github.com/microsoft/fluentui-react-native/pull/4463). + +### Win32 Paper + +- **Physical MCP click:** six physical-pointer checkbox clicks, each state-checked, with 23 calls and zero errors at `505a2187b`. Evidence: [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Generic consumer handoff, preemption, worker kill and lost runner-grant response:** same result as Windows Fabric at `b1bbb26df`. + +- **Connected descriptor-path checks:** physical checkbox and target PNG passed at `fec94cec4`. Evidence: [#4469](https://github.com/microsoft/fluentui-react-native/pull/4469). Injected fresh-preview fencing and the secondary tool popup were later recorded in [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Descriptor lifecycle:** the same real-descriptor lifecycle set as Windows Fabric passed at `1f5333e25` plus the follow-up fixes `8458bde48` and `a6b89e924`. Evidence: [#4463](https://github.com/microsoft/fluentui-react-native/pull/4463). + +- **Process lifetime:** after MCP session close the exact REX process can outlive its closed window; it is stopped only after identity verification. + +### Hosted CI and portable evidence + +- **Portable:** the package passes 32 tests (31 in the focused server selection) at `5dbadb53b`; these do not qualify native behavior. Evidence: [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Hosted generic native consumer:** the Windows and Win32 accessible jobs each passed four ownership modes with one capture-policy skip, which is not a pass. Evidence: [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- Not run: 200% scale, mixed DPI and negative-origin layout. Windows and Win32 doctor `--permissions` is unsupported. + +## Measurements + +Measurements keep the latest values per endpoint. +Tokens use `o200k_base` over all MCP text blocks; Windows and later macOS runs used `gpt-tokenizer` 4.1.0 with MCP SDK 1.32.2 stdio through the private descriptor. +The macOS Fabric baseline at `41fc7473f` is retained where it is the latest value; its tokenizer package version was not retained. +Discovery is paid once per connection, and each value stays tied to its revision. + +### Discovery + +| Item | Value | Revision / source | +| --------------------------------- | --------------------------------------------- | ------------------------------------- | +| Six-tool `tools/list` array | 4,691 tokens; 20,277 UTF-8 bytes (characters) | `9d3e11516` and `505a2187b` (Windows) | +| Full `tools/list` result envelope | 4,693 tokens; 20,287 bytes | `505a2187b`; macOS `141b88b4b` | +| Server instructions | 120 tokens; 591 bytes | Same | + +### macOS Fabric + +| Measurement | Result | Caveat | +| ---------------------------------------------------- | ------------------------------------------------------------ | ------------------------------------------------- | +| Named Toggle + check + closing observe (`41fc7473f`) | 386 tokens; 48 ms | One warm run at 2720 x 1988 on a 2x display | +| Physical click + checked check (`41fc7473f`) | 181 ms | One run | +| Default observe (`141b88b4b`) | median 21.9-27.8 ms, max 1,680.2 ms; 905-948 tokens per call | 20 calls per story; 100-node projection, complete | +| Idle delta (`141b88b4b`) | median 21.5-26.3 ms, max 1,688.1 ms; 151-152 tokens per call | Same method | +| First attach after discovery (`141b88b4b`) | 846.4 ms | One sample; excludes service and app startup | +| Read cancellation (`141b88b4b`) | 301.5 ms, then a usable observe | No mutation cancellation claim | + +No successful physical click cost is recorded for the latest macOS revision, because the click was refused. + +### macOS Paper + +| Measurement (`141b88b4b`) | Result | Caveat | +| ------------------------------------------ | ------------------------------------- | -------------------------------------------------------------- | +| Observe, button / checkbox / menu stories | p50 31.8-34.1 ms; 1,188-1,322 tokens | 20 calls each; menu emitted 100 of 103 nodes, 0 of 20 complete | +| Idle delta | p50 31.5-36.4 ms; 151-152 tokens | 20 calls each, all complete | +| First attach after discovery | 831.6 ms | One sample | +| Text entry + value check + closing observe | 2,015.7 ms; 1,371 tokens; 4,194 bytes | One sample; not a physical Checkbox activation | +| Read cancellation | 302.2 ms | Mutation cancellation not run | + +The p90 values (1,142-1,710 ms) are slow outliers; samples are warm and window size and scale were not retained. + +### Windows Fabric and Win32 Paper + +Revision `505a2187b` (click) and `9d3e11516` (other rows); Windows 11 build 26300, AMD64; MCP SDK 1.32.2; remote RDP for the click run. + +| Measurement | Windows Fabric | Win32 Paper | +| --------------------------------------------------- | --------------------------------------- | --------------------------------------- | +| `open` (attach) | 45 ms | 31-40 ms | +| Physical click response (6 samples) | 99-106 tokens; p50 299.1 ms, max 370.0 | 96-103 tokens; p50 308.0 ms, max 318.2 | +| Click + checked check + default observe (6 samples) | 1,999-2,005 tokens; 411.8-550.6 ms | 1,585-1,592 tokens; 431.8-473.5 ms | +| Checkbox `activate` + check + default observe | 1,923 tokens; 2,288 ms | 1,645 tokens; 2,302 ms | +| Observe default (20 calls, 100 nodes) | median 66.7-73.7 ms; 1,801-1,843 tokens | median 81.1-85.8 ms; 1,594-1,633 tokens | +| Idle delta (20 calls) | median 72.7-77.5 ms; 368 tokens | median 83.3-88.3 ms; 220 tokens | +| Default observe completeness in click run | Incomplete: `maxNodes`, 13 omitted | Complete | + +- Click request JSON adds 56-64 tokens; click RPC time excludes startup and discovery; with six samples p90 is the maximum. +- On Windows an idle delta is never empty: three menu-bar items (React, Debug, System) get new references on every read, and whether they lack stable runtime identities is unconfirmed; without them it is about 151 tokens. +- The Win32 idle delta was not decomposed. +- The `activate` rows come from a disconnected console session and are not comparable with the click rows or the macOS row. +- These are local samples, not a latency distribution or platform ranking; do not claim token reductions from reducing scope or omitting discovery. + +## Maintaining this document + +- Edit only the topic your change affects, and keep that topic's line in the status overview in step with it. +- Describe what is true now. Do not add dated entries, run logs, or "updated" stamps. +- Record per-run evidence (commands, counts, run ids, artifact locations, cleanup outcome) in the pull request description or the tracking issue, and keep only the latest result for each endpoint here with a link to that record. +- Put remaining-work detail in the tracking issue and link the issue from the topic's open-work list; remove the link when the issue closes. +- Write one sentence or fact per line, and keep a blank line between adjacent list entries that different changes are likely to edit, so concurrent changes merge independently. +- Historical results stay tied to the revision that produced them and are never reported as qualifying a later revision. +- Skipped checks are not passes; report each endpoint (macOS Fabric, macOS Paper, Windows Fabric, Win32 Paper) separately. +- The detailed per-run ledger that preceded this format is preserved in git history: `git show a7af2b4b2:packages/agentic/desktop-driver-mcp/references/current-status.md`. diff --git a/packages/agentic/desktop-driver-mcp/references/tools.md b/packages/agentic/desktop-driver-mcp/references/tools.md index b4e1a1a9f7..9e16dcefdd 100644 --- a/packages/agentic/desktop-driver-mcp/references/tools.md +++ b/packages/agentic/desktop-driver-mcp/references/tools.md @@ -13,7 +13,7 @@ result, and failure reasons are defined once in `@fluentui-react-native/desktop- service rejects is rejected here first, before anything reaches the service. Unknown fields are refused. The historical tool list measured about 20.3 KB; the test bounds serialized JavaScript **characters** at 22,000, not tokens or -UTF-8 bytes. See [measurement history](current-status.md). +UTF-8 bytes. See [measurements](current-status.md#measurements). | Tool | Calls | Hints | | ----------------- | ------------------------------------------------------------------- | --------------------------- | diff --git a/packages/agentic/desktop-driver/README.md b/packages/agentic/desktop-driver/README.md index d8c8d74895..b0f7b0a36a 100644 --- a/packages/agentic/desktop-driver/README.md +++ b/packages/agentic/desktop-driver/README.md @@ -13,8 +13,8 @@ fake hosts for contract tests. It does not depend on Storybook, React, or React Native. **Work in progress:** see [current status](references/current-status.md) for -implemented capabilities, incomplete lifecycle guarantees, endpoint qualification, -and measured token/latency history. +implemented capabilities, incomplete lifecycle guarantees, the latest +per-endpoint qualification results, and measured token/latency results. Story providers use the existing `/storybook` subpath for `StoryProviderContext`, `StoryBridgeUnavailableError`, and the versioned marker verifier. Built-in routes @@ -35,7 +35,7 @@ a new verified selection; finding fresh elements alone cannot restore readiness. V1 permits one active WebDriver session per registered target in one service. Native helpers also exclude controlling the same exact application incarnation -across services; endpoint qualification and recovery limits are recorded in +across services; endpoint qualification and recovery limits are summarized in [current status](references/current-status.md). Native builds run only on the target operating system. Installation never compiles or downloads a helper. diff --git a/packages/agentic/desktop-driver/references/README.md b/packages/agentic/desktop-driver/references/README.md index 78dedfe89b..3e9980bc88 100644 --- a/packages/agentic/desktop-driver/references/README.md +++ b/packages/agentic/desktop-driver/references/README.md @@ -9,17 +9,17 @@ owner_paths: Start with [current status](current-status.md) when assessing readiness. -| ID | Document | Read when | -| --------------------------------- | ------------------------------------------ | ------------------------------------------------------------------------------ | -| `desktop-driver.getting-started` | [Getting started](getting-started.md) | Installing, building the package, or first setup | -| `desktop-driver.architecture` | [Architecture](architecture.md) | Choosing a module boundary or tracing ownership | -| `desktop-driver.service` | [Service integration](service.md) | Embedding the WebDriver service or registering targets | -| `desktop-driver.native-helpers` | [Native helpers](native-helpers.md) | Building or resolving native helpers | -| `desktop-driver.test-integration` | [Test integration](test-integration.md) | Connecting WDIO or using driver testing APIs | -| `desktop-driver.ci-integration` | [CI integration](ci-integration.md) | Provisioning helpers and validating CI lanes | -| `desktop-driver.security` | [Security model](security.md) | Reviewing target trust, ownership, or evidence boundaries | -| `desktop-driver.protocol` | [WebDriver contract](protocol.md) | Changing or consuming WebDriver behavior | -| `desktop-driver.operations` | [Agent operations](operations.md) | Calling or changing observe, find, journal, refs, or budgets | -| `desktop-driver.contributing` | [Contributor reference](contributing.md) | Changing package source, tests, or build tooling | -| `desktop-driver.remaining-work` | [Agent efficiency plan](remaining-work.md) | Implementing MCP-first agent workflows and macOS/Windows parity | -| `desktop-driver.current-status` | [Current status](current-status.md) | Checking implemented capabilities, qualification gaps, and measurement history | +| ID | Document | Read when | +| --------------------------------- | ------------------------------------------ | ------------------------------------------------------------------------------------------ | +| `desktop-driver.getting-started` | [Getting started](getting-started.md) | Installing, building the package, or first setup | +| `desktop-driver.architecture` | [Architecture](architecture.md) | Choosing a module boundary or tracing ownership | +| `desktop-driver.service` | [Service integration](service.md) | Embedding the WebDriver service or registering targets | +| `desktop-driver.native-helpers` | [Native helpers](native-helpers.md) | Building or resolving native helpers | +| `desktop-driver.test-integration` | [Test integration](test-integration.md) | Connecting WDIO or using driver testing APIs | +| `desktop-driver.ci-integration` | [CI integration](ci-integration.md) | Provisioning helpers and validating CI lanes | +| `desktop-driver.security` | [Security model](security.md) | Reviewing target trust, ownership, or evidence boundaries | +| `desktop-driver.protocol` | [WebDriver contract](protocol.md) | Changing or consuming WebDriver behavior | +| `desktop-driver.operations` | [Agent operations](operations.md) | Calling or changing observe, find, journal, refs, or budgets | +| `desktop-driver.contributing` | [Contributor reference](contributing.md) | Changing package source, tests, or build tooling | +| `desktop-driver.remaining-work` | [Agent efficiency plan](remaining-work.md) | Implementing MCP-first agent workflows and macOS/Windows parity | +| `desktop-driver.current-status` | [Current status](current-status.md) | Checking implemented capabilities, per-endpoint qualification, open work, and measurements | diff --git a/packages/agentic/desktop-driver/references/current-status.md b/packages/agentic/desktop-driver/references/current-status.md index baf7645fac..8ba3ba7526 100644 --- a/packages/agentic/desktop-driver/references/current-status.md +++ b/packages/agentic/desktop-driver/references/current-status.md @@ -5,3309 +5,338 @@ owner_paths: - packages/agentic/desktop-driver --- -# Current capability and qualification status - -**Work in progress, updated 2026-11-12.** "Done" below means the stated -implementation exists, not that every native endpoint is qualified or that the -complete orchestration product is ready for submission. Historical measurements -remain tied to their original revision; they were not rerun by this update. -See [architecture](architecture.md) for boundaries and -[remaining work](remaining-work.md) for open delivery gates. Follow-up -lifecycle, orchestration, and qualification work is tracked in -[desktop driver improvements (#4435)](https://github.com/microsoft/fluentui-react-native/issues/4435) -and its sub-issues, which record dependencies and parallel waves. - -## Capabilities - -| Capability | State | Boundary or remaining work | -| ---------------------------------------------------------- | -------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| W3C service and typed/WDIO clients | Done | Registered targets only; native browser-inapplicable commands explicitly unsupported. | -| Application-neutral driver | Done | No Storybook, React, or React Native dependency; package-boundary tests reject forbidden imports/dependencies. Optional story provider is injected. | -| Native sessions and command/input queues | Done | One reservation per registered target in one service. Exact macOS application exclusion and the physical-input mutex passed signed AppKit contracts on 2026-11-12. Renderer and failure-path qualification remains separately scoped below. Windows behavior is unchanged. | -| Verified private discovery and grants | Done | Native identity verified before credential transmission; all protected driver routes authenticate. Storybook agent manifests additionally disable alternate channel mutations in the integration adapter. | -| Attenuated connection principals and permission scopes | Done in deterministic contracts | Driver/workload/recovery scopes are distinct; fixed driver-only HTTP delegation, target/scope subsets, synchronous subtree fencing and awaitable/retryable drain. Native descriptor-path requalification not run for this change. | -| Mutation effects, settlement, quarantine | Done | Native mutations settle before advancing. Uncertain story changes now quarantine instead of assuming a later render is safe. | -| Durable settlement storage and trusted recovery API | Done | Generic/Storybook native paths share the harness factory's private storage and exact-incarnation probes. Direct library embedders must pass `recovery`; restart is not proof. | -| Bounded observation, references, paging, deltas | Done | Provider capability/budget checks, explicit completeness and truncation. Reference generation is session-local; reopen after application replacement. | -| Find, action, check, sequence | Done | Up to eight sequence steps; request-id outcomes prevent replay within their retained session history. No replay across service replacement. | -| Capture and artifact pages | Done | Session-bound bounded artifacts and hashes; pixels are opt-in. Screenshots are evidence, not visual comparison. | -| Named accessibility actions | Done | Agent operations and WDIO (`desktopPerformAccessibilityAction`) invoke one uniquely named action; the W3C route is fenced against stale Storybook preview context like click, clear and value. | -| Story catalog/select/reset | Done, experimental integration | Local lifetime/signal context, authenticated epochs persisted per native-verified app incarnation, versioned marker conformance, prompt disconnect failure, and stale preview rejection. Bounded macOS Paper/Fabric marker and stale-reference checks passed at `6fca14788`; native cancellation/fault injection remain unqualified. Unknown navigation effects still quarantine. | -| Story args on protected driver | Intentionally unsupported | `story-args-unsupported`, including the session-owning grant and controlled orchestrator. HTTP and foreign WebSocket args are refused/not relayed without logging values; legacy unprotected args remain separate. | -| Exclusive Storybook control | Work in progress: native qualification | Deterministic endpoint evidence: controlled HTTP mutations return 403; foreign WS selection/prepare/args never reach the runtime; authenticated driver prepare/select uses F4 settlement and fencing. Non-agent channel remains legacy unprotected. Native endpoint evidence not yet run: macOS Paper/Fabric, Windows Fabric/Win32 Paper. Preview context checks preserve chrome/secondary references. | -| Supervisor crash containment and automatic reclaim | Work in progress outside driver | Exact ownership ledger, surviving-resource reconciliation, and lifelines belong to desktop-test. The Windows helper's Job Object ends an unpackaged application it launched when the helper dies; packaged launches are `uncontained`. | -| Application-exit/reconnect continuity | Work in progress | Bounded macOS Paper/Fabric full-service restart checks passed at `6fca14788`: private authenticated-epoch record plus exact live-incarnation proof yields renderer-specific relaunch-required, effect none. Initial absence and exited/replaced apps stay bridge-unavailable; failed/unknown probes are contract-tested only. No transparent restoration or quarantine recovery; inspected RN Storybook 10.7.1 / Storybook 10.6.5 transport does not reconnect. | -| Recording/replay export, event-tail API, visual comparison | Work in progress/deferred | No agent-session-to-WDIO export; window-events still returns its full bounded ring. | -| Intel macOS and distributed/remote control | Unsupported | Current macOS helper is arm64; service is loopback-only. | - -The approved direction is a stronger `desktop-test` workload host, with native -authority here and MCP as a client. No coordinator package or daemon is being -added. Management/run schemas stay in the harness, not this package. - -## Native evidence and full-submission gates - -The macOS permission gate was **resolved, observed 2026-11-12**: the -stable-signed helper reports Accessibility/PostEvent true without prompting. -This alone is not requalification; see the -[2026-11-12 supporting evidence](#f15-non-ax-macos-supporting-evidence--2026-11-12). - -| Lane | Recorded evidence through 2026-11-12 | Still required before full submission | -| ------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| macOS Fabric | M3 at `41fc7473f`; bounded F4/restart at `6fca14788`; signed neutral ownership/inspection/exclusion and 224 accessible smoke checks on 2026-11-12 | Native cancellation/fault injection, fresh preview after lost context, secondary windows, physical pointer/normal-keyboard/capture paths, supervisor crash, descendants, interrupted publication and reliability window. | -| macOS Paper | Hosted build/smoke and bounded F4/restart at `6fca14788`; separately signed neutral ownership/inspection/exclusion and 224 accessible smoke checks on 2026-11-12 | Native cancellation/fault injection, fresh preview after lost context, secondary windows, physical pointer/normal-keyboard/capture paths and complete failure/reliability qualification. | -| Windows Fabric | M3 real-app observation and story MCP path recorded at `600235c35` | Same-branch lifetime/restart and failure gates; 200%/mixed DPI, minimized/negative-origin windows, live lock/RDP/elevation, popups, capture cancellation. | -| Win32 Paper | M3/agent path recorded; native fixture physical input/capture at 100% and 150% | Interactive rerun of current native contracts; physical Ctrl+C, mixed DPI and failure paths. Hosted input/capture skips are not passes. | -| Generic native applications | Historical macOS Calculator/plain Win32 evidence; signed AppKit ownership and physical-input mutex passed on 2026-11-12 | Integrated survivor/interrupted-publication/recovery parity; a probe alone is not real-app lifecycle qualification. | -| This update's portable changes | Deterministic fake-host MCP, authority, and story regression coverage; the Windows oversized-tree fallback passed hosted CI | CI and matching-OS live evidence are distinct. The oversized-tree fallback is compiled and contract-tested only on hosted CI; no live Windows qualification. | - -Hosted Windows policy disables physical input and can reject capture with -`0x80070057`. Capability-gated skips must stay explicit. `activate` succeeding -does not qualify a refused physical click. Each new lifecycle/native change -needs its own endpoint evidence; a final aggregate matrix is not a substitute. -Live requalification of these lanes is tracked in -[#4441](https://github.com/microsoft/fluentui-react-native/issues/4441), and -the repeated-run support matrix in -[#4451](https://github.com/microsoft/fluentui-react-native/issues/4451). -The parameterized procedure is -`packages/agentic/desktop-test/references/native-qualification.md`. - -### Provisional F15 macOS evidence — 2026-11-06 - -**Provisional; exact source revision `388a60cc803ab78202a6f64f92e3f53b779c5c39`.** -App: Agentic Components Storybook / ReactTestApp (configured bundle ID -`com.microsoft.fluentui.agenticstorybook`). OS: macOS 27.2, arm64. -The helper was explicitly built and the noninteractive doctor reported -Accessibility, PostEvent, and ScreenCapture preflight `true`, console unlocked, -ad-hoc signing, and designated requirement -`cdhash H"0d2e7a81dfd4695b7fb84aa1b3242f5e1773e8b4"`. - -| Renderer | Command and actual selection | Result | Cleanup | Artifact | -| -------------- | --------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | -| macOS Fabric | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook smoke --macos --mode stories-and-tests`; 211 selected/executed | Passed 211, failed 0, skipped 0, unsupported 0; retries 0 | No owned app/driver/channel/Metro process; ports 21338/30029/45382 had no listeners; `interactive.lock` absent | `apps/storybook/artifacts/desktop/0f99740f-631b-444a-a705-76a9b4a761b7/run.json` | -| macOS Paper | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook smoke --macos --paper --mode stories-and-tests`; 211 selected/executed | Passed 211, failed 0, skipped 0, unsupported 0; retries 0 | No owned app/driver/channel/Metro process; ports 21338/30029/45382 had no listeners; `interactive.lock` absent | `apps/storybook/artifacts/desktop/e82a83c2-a01d-41c0-b308-2551630e8b1e/run.json` | -| Windows Fabric | Not run — this machine is macOS-only | Not run; 0 selected | Not applicable | None | -| Win32 Paper | Not run — this machine is macOS-only | Not run; 0 selected | Not applicable | None | - -| Descriptor-path renderer | Command / selection | Passed / failed / skipped / unsupported | Cleanup | Artifact | -| ------------------------ | --------------------------------------------------- | --------------------------------------- | -------------------------------------------- | -------- | -| macOS Fabric | Not run; no real descriptor consumed; selection N/A | 0 / 0 / 0 / 0; Not run | N/A for descriptor path; smoke cleanup above | None | -| macOS Paper | Not run; no real descriptor consumed; selection N/A | 0 / 0 / 0 / 0; Not run | N/A for descriptor path; smoke cleanup above | None | - -The `STORYBOOK_SMOKE_STORY` value did not constrain the callbacks in these -runs; `run.json` recorded 211 selected. An initial Fabric attempt failed before -test selection because CocoaPods was out of sync. `yarn storybook prep --macos` -resolved that environment prerequisite; the failed attempt is retained at -`apps/storybook/artifacts/desktop/2e42a0aa-506d-4edb-a3ca-32e5b7c26b04/run.json`. -These smoke runs used the non-agent WDIO path. They do **not** qualify the MCP -descriptor lifecycle, native story quarantine/recovery, F2/F3/F4/F5 integration, -or physical-input/capture requirements. The descriptor rows are not run on -this baseline revision; update them on the combined integration SHA after the -real descriptor-path pass. - -### 2026-11-06: bounded F4 restart qualification - -Source revision `6fca1478871f41ec884fb4bd2cff66cff8197833` includes mainline -`cd0ede21489c9ab8cb76fea07c4dec2e9699389f` through merge -`35057966aeb20b70756af4e83c58c3242675aa72`. macOS arm64 Paper and Fabric -were run separately. These are bounded acceptance checks, not full smoke or -reliability passes. - -**Passed:** the nonprompting preflight -`FURN_DESKTOP_DRIVER_BUILD_POLICY=never yarn desktop-driver doctor --platform macos --permissions` -reported AX, PostEvent and screenCapture all `true`. The cached helper retained -cdhash `0d2e7a81dfd4695b7fb84aa1b3242f5e1773e8b4`; no helper rebuild, permission -prompt, TCC change or unrestricted capture was used. - -**Passed separately for Paper and Fabric:** authenticated button navigation, -actual native marker conformance, rejection of an old preview reference after -another selection, and continued usability of an existing chrome reference. -The bounded authored command was -`FURN_DESKTOP_DRIVER_BUILD_POLICY=never yarn storybook test --macos --test '*exposes enabled button semantics*'` -(append `--paper` for Paper). Passing authored run IDs, beneath -`apps/storybook/artifacts/desktop/`, were -`c462e891-a755-4a27-9b63-9eb07f64934b` (Paper) and -`cd1ef4ac-8f38-4d3e-ac06-0d957ade0207` (Fabric). - -**Passed separately for Paper and Fabric:** stop the entire driver supervisor -while preserving the authenticated app; start a fresh driver; verify the -unchanged authenticated epoch and exact live PID/start time. The story route -returned `relaunch-required`, the matching renderer, and mutation effect -`none`. The same bounded authored command intentionally failed navigation -with that typed error; it is not an authored-test pass. Independent native -route assertions passed. The private endpoint/target storage namespace and -epoch survived the service restart without relying on the rotated manifest. - -**Passed separately for Paper and Fabric:** quit the recorded app and driver, -launch a replacement app before its channel exists, then restart the driver. -The route returned `bridge-unavailable`, effect `none`, and removed the exited -incarnation's record. A subsequent no-record check remained unavailable. -Initial absence before authentication also returned unavailable, effect `none`. -Early Paper startup attempts failed navigation while the runtime was not ready; -they remain failures, not passes hidden by the later settled run. The earlier -Fabric full-catalog traversal with app-stop AppleEvent `-1712` remains a failed -smoke command; bounded reruns used verified-process cleanup instead. - -**Passed:** focused and owning-package checks, clean-cache `yarn lage buildci` -(178 tasks), and `yarn lage test-links --verbose --grouped` (90 tasks). -Private session evidence is retained under `files/f4-validation/`: -`restart-{paper,fabric}-{markers,record,assert,replaced-assert}.log`, -`restart-permissions.log`, `restart-buildci.log` and `restart-test-links.log`. -After both lanes, owned apps, supervisors, Metro, channels and helpers were -stopped; ports 23489, 39627 and 30542 and the interactive lock were verified -clear, and the exclusive validation slot was released. - -**Not run natively:** post-dispatch cancellation, delayed ready/error/render -fault injection, fresh-preview actions after lost context, secondary windows -and physical input. **Not run:** Windows Fabric and Win32 Paper qualification. -Portable contract coverage is not native evidence for these gaps. - -## Measurement ledger - -Preserved from the former root capability report, not newly measured here. -Use **Measured** for observed product/recorded-run data, **Estimate** for -synthetic/token-formula calculations, and **Not run** for missing evidence. -Do not silently replace historical rows with current claims. - -### 2026-11-04: delivered operation text costs at `41fc7473f` - -**Measured.** macOS Fabric Storybook, 2720 x 1988-pixel window (1360 x 994 -points, 2x display), served by `yarn storybook driver --macos --agent`. -An MCP SDK stdio client rendered `components-button--default`, -`components-checkbox--default`, and `components-menu--default`, each -marker-confirmed. Baseline routes and new operations sampled the same app -state. `o200k_base` token counts include every returned text block and the -result envelope; all observations were complete, without truncation. - -| Result | Measured text tokens | Scope/notes | -| ------------------------------ | -------------------- | ------------------------------------------------------ | -| `furn/tree`, pretty JSON | 12,972-13,643 | 63-67 KB; 142-150 nodes | -| Same tree, minified | 7,102-7,469 | Same nodes | -| Page source XML | 6,244-6,562 | Same app state | -| `furn/window-events` | 5,331-5,352 | Entire bounded ring, not a recent tail | -| `observe`, `all` outline | 1,602-1,859 | Same 142-150 nodes; 86-88% below pretty JSON | -| `observe`, `all` JSON nodes | 3,290-3,375 | Explicit JSON projection | -| `observe`, default interactive | 1,122-1,304 | 74-76 elements; not the same scope as all nodes | -| `observe`, content | 1,199-1,373 | 79-80 elements | -| `observe`, story-root scope | 252-311 | 2-6 elements | -| Unchanged delta | 149-150 | Retained-sample comparison | -| One-element delta | 192 | Checkbox toggled | -| Find story root | About 200 | Unique result | -| Find button role | About 490 | Multiple matches | -| Check | About 81 | Small predicate result | -| Capture without pixels | About 140 | Metadata, hash, and resource link | -| Journal, default 50-entry page | About 3,570 | Bounded page | -| Stories, first 25 | 637 | Full 209-story catalog in three 100-entry pages: 5,429 | - -Envelope overhead was about 150 tokens for observation/delta payloads, not a -universal floor for every operation. Include MCP discovery separately: -`@fluentui-react-native/desktop-driver-mcp/references/current-status.md`. - -### 2026-11-02 baseline: recorded evidence at `c84acb741` - -**Measured from 61 recorded macOS failure-evidence sets**, not a new live -session. Same tokenizer and window dimensions; different app states from the -November 4 comparison. Median full tree: 157 nodes, 69 interactive roles, -12 testIDs, maximum depth three. - -| Representation | Median tokens | Range or limitation | -| ---------------------------- | ------------- | ---------------------------------------- | -| Pretty tree evidence | 14.2k | 11.6k-18.4k; about 70 KB | -| Minified tree | 7.8k | Same retained evidence | -| Minified tree without rects | 4.8k | Less information, not equivalent content | -| Legacy application describe | 9.4k | Depth 3 / 100-node bound | -| Legacy story-scoped describe | 223 | Maximum 2.1k | -| XML page source | 6.8k | Recorded source | -| Window events | 5.0k | Full 512 + 128-entry ring | - -Earlier synthetic compact-outline estimates (824 tokens all named nodes, -683 interactive, 39 story-scoped) excluded delivered references/action metadata -and envelopes. They are **not measurements** and must not be used as product -targets already achieved. The screenshot estimate of about 1,530 Claude or -1,105 OpenAI high-detail image tokens applied then-published resize formulas to -the 2720 x 1988 PNG; it is not measured billing or a current model guarantee. -Third-party agent-device comparisons were not same-app measurements and are not -retained as product evidence. - -### Historical latency, not a current performance guarantee - -| Date/revision and measurement boundary | Operation | Median | Tail/qualification | -| ------------------------------------------------- | ------------------------------ | -------- | ---------------------------------- | -| 2026-11-04 / `41fc7473f`, MCP client end-to-end | Default observe | 29-31 ms | 20 warm calls per story; max 35 ms | -| Same | Delta | 32-35 ms | 20 warm calls per story; max 37 ms | -| Same | Physical click + checked check | 181 ms | One run, not a distribution | -| Same | Named Toggle + check + observe | 48 ms | One run; 386 text tokens | -| 2026-11-02 / `c84acb741`, recorded helper journal | Snapshot/windows | 0-2 ms | Different app states | -| Same | Attach | 2 ms | Max 983 ms | -| Same | Probe | 5 ms | Helper-side only | -| Same | Perform actions | 19 ms | Helper-side only | -| Same | Tree | 58 ms | Full tree, not bounded observe | -| Same | Find | 65 ms | p90 498 ms; max 693 ms | -| Same | Source | 79 ms | Max 1.16 s | -| Same | Physical click | 121 ms | Helper-side only | -| Same | Capture window | 152 ms | Helper-side only | - -Do not subtract helper timings from end-to-end timings or compare them as a -controlled speedup. Five reads estimated as one interactive observation and -four unchanged deltas cost about 1.7k-2.1k text tokens **plus discovery**; that -is a calculated scenario, not an observed five-turn agent session. - -### Recording the next measurement - -Append a dated run with commit, OS/architecture, renderer, application/story, -window dimensions/scale, helper build/signing mode, cold/warm state, sample -count, tokenizer/version, exact projection/budgets, completeness, text blocks -included, discovery overhead, and client-versus-helper timing boundary. -Retain an approved artifact/report locator without credentials, private paths, -application text, or raw screenshots. Record missing artifacts as unavailable. -The historical report did not retain a portable raw-artifact locator or tokenizer -package version; those provenance gaps remain, rather than inventing them. - -Report p50/p90/max with sample count, retries, failures, skipped capabilities, -and cleanup status. Compare equivalent assertions and complete scopes. -Windows and Paper measurements need their own rows; do not extrapolate macOS -Fabric numbers. - -### 2026-11-06: partial native evidence at `3c780cfd1bccaf510ece080e4be2b516c645989a` - -App: Agentic Components Storybook / ReactTestApp -(`com.microsoft.fluentui.agenticstorybook`), macOS 27.2 arm64. The exact -candidate helper was built with `yarn storybook build-driver --macos`; fresh -nonprompting preflight was -`yarn desktop-driver doctor --platform macos --permissions`. Accessibility, -PostEvent, and ScreenCapture were all `true`; no prompt or permission change -was made. The console was on and unlocked. The helper was arm64/release, -ad-hoc signed, with designated requirement -`cdhash H"0d2e7a81dfd4695b7fb84aa1b3242f5e1773e8b4"`. Host load was 3.25 / -4.41 / 3.81 at 1/5/15 minutes. Detailed preflight output is kept privately as -`f15-native-qualification-3c780cfd-20261106T0200/doctor.json`. - -| Renderer | Command and selection | Outcome | Cleanup and artifact | -| ------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| macOS Paper | `yarn storybook driver --macos --paper --agent`; app via `yarn storybook run --macos --paper`; MCP calls used the default descriptor. Driver run selected/executed 0 tests. | Attach-only discovery, capture, normal selection/reset, protected-args refusal, revocation, and 21-second channel check were reported. Same-port restart/rediscovery was partial because explicit old-session close was not retained. | Cleanup was reported; ports 21338/30029/45382 had no listeners and the lock/descriptor were absent at handoff. Paper process IDs were not retained for PID-level confirmation. Driver run: `apps/storybook/artifacts/desktop/a58d2fdd-5584-47f4-8747-5307607b4c0d/run.json`; private `paper-native-capture.png`. | -| macOS Fabric | `yarn storybook driver --macos --agent`; app via `yarn storybook run --macos`; MCP calls used the default descriptor. Driver run selected/executed 0 tests. | Attach-only discovery and capture passed. Selection/reset markers settled; F5 HTTP/WS bypass checks and 21,003 ms idle-channel check passed. The old-reference acceptance remains inconclusive; the physical-input attempt returned `physicalInputApplied:false`. | Fabric cleanup record reports the service owner exited, descriptor removed/grant revoked, and attached app preserved through service shutdown; ports had no listeners and the lock/descriptor were absent at handoff. Driver run: `apps/storybook/artifacts/desktop/c15a35b9-7b8d-4edb-adba-84e3cd5a3023/run.json`; app launch: `apps/storybook/artifacts/desktop/8fb9a576-cde8-470c-a8bb-a76b6d8fabc0/run.json`; private `fabric-native-capture.png`. | - -The capture operation succeeded for both renderers. Fabric's first capture -caller attempt used the wrong argument shape and failed; the corrected -documented `image` field succeeded. This caller error is retained and is not -counted as a product failure. The separate Paper non-agent manifest smoke -selected 319, executed/passed 308, failed 0, skipped 11, unsupported 0, -incomplete 0, retries 0, flaky 0; it is not descriptor-path evidence. The -three 319-count run artifacts do not all preserve renderer metadata, so only -the Paper-attributed artifact is credited to Paper. - -Fabric's first two foreign-WebSocket bypass assertions failed, followed by a -final attempt that passed with no observer-message delta; the cause is -unknown. An intermediate protected-args record had a false unchanged-story -assertion, while the final record was unchanged and value-free. The overall -bypass/args results are therefore recorded as Partial in the harness matrix. -The physical-input attempt returned `physicalInputApplied:false`; it was not -retried. - -This is not full native qualification. Dispose-during-open, failed delete, -ambiguous create, quarantine/recovery, failed-cleanup claim retention, faulted -F3 survivor reconciliation, F4 cancellation/readiness/render/lost-context/ -secondary-window cases, and legacy-lane separation were not run. Physical -input did not pass. Windows Fabric and Win32 Paper are Not run on this host. -The F5 reference-generation fix at -`2d8e8db001b64ebea85f0d6dce141f040fb5182d` is outside the tested SHA and needs -a new native run; do not attribute the original inconclusive probe to it. - -## 2026-11-06 combined-candidate macOS evidence at `a19fd3e1e7b3f1f443a241b1571cb029162f4ef2` - -App: Agentic Components Storybook / ReactTestApp -(`com.microsoft.fluentui.agenticstorybook`), macOS 27.2 arm64. The exact -arm64/release helper was rebuilt with -`yarn storybook build-driver --macos`; fresh nonprompting -`yarn desktop-driver doctor --platform macos --permissions` reported -Accessibility, PostEvent, and ScreenCapture granted. No permission changes -were made. Helper fingerprint: -`b58e4d8692702fd1dfa16b73ac987d1738e6e17d16757cbc3c6007722b8e3869`; -ad-hoc signing designated requirement: -`cdhash H"0d2e7a81dfd4695b7fb84aa1b3242f5e1773e8b4"`. -Observed quiet host load was 2.49 / 5.27 / 5.47 (1/5/15-minute averages). - -| Renderer | Command and selection | Result | Cleanup and artifact | -| ------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| macOS Fabric | `yarn storybook driver --macos --agent`; app `yarn storybook run --macos`; manual MCP descriptor session, 0 automated tests selected/executed. Separate `yarn storybook smoke --macos --mode stories-and-tests`: 211 selected/executed. | Retained preview-element sequence passed: same session/incarnation 1; select settled at region 2, reset at region 3; old element alias rejected after both transitions; fresh find produced a distinct alias that worked; old alias remained rejected. Fabric idle session passed at 21,002 ms. Same-port replacement partial (recorded Failed, reclassified on contract review): old client's first post-replacement `desktop_session.status` returned the designed fail-closed `adapter` / `service-unavailable`; close succeeded, but rediscovery on a later call was not demonstrated. One physical click was issued but its nested mechanism/disposition was not retained, so partial only. | App PID 26792 survived service shutdown/restart and was stopped afterward. Owned PIDs gone; ports 21338/30029/45382 free; descriptor and lock absent. Smoke artifact `apps/storybook/artifacts/desktop/a2475e8d-7ea9-465f-8461-b8f115cfcee5/run.json`; service runs `apps/storybook/artifacts/desktop/414cf449-bfd4-4b9f-b630-63e86ebdd5c4/run.json`, `apps/storybook/artifacts/desktop/d732ae38-df56-4b81-a56c-2791bdbac5e8/run.json`; blocked-claim attempt `apps/storybook/artifacts/desktop/73563809-2d77-49cb-a297-c74fd57ecd36/run.json`. Sanitized operation records are in the private `f15-native-qualification-a19fd3e1-20261106T0406/` log directory. | -| macOS Paper | `yarn storybook driver --macos --paper --agent`; app `yarn storybook run --macos --paper`; manual MCP descriptor session, 0 automated tests selected/executed. Separate `yarn storybook smoke --macos --paper --mode stories-and-tests`: 211 selected/executed. | Retained preview-element sequence passed with the same-session region transition and stale/fresh alias assertions. No same-port, idle-channel, or physical-input case was run for Paper on this SHA. | App, service, and Metro processes exited; ports free, descriptor absent, stale lock removed only after verifying its recorded owner had exited with empty resources/unresolved effects. Smoke artifact `apps/storybook/artifacts/desktop/87f0452f-5fd7-48eb-8494-a63202842521/run.json`. Sanitized record is in the private log directory above. | - -The Fabric replacement case was not retried. It was recorded as Failed and -reclassified as Partial on contract review: the replacement service started on -the same ports, and the old MCP client's first status call was rejected with -`service-unavailable`, which is the designed fail-closed response before the -next call re-verifies discovery. The following close cannot prove rediscovery, -so a native status call after the rejection remains required. No mutation was -issued. Manual MCP calls are -not counted as selected tests. Manifest smoke is the separate non-agent legacy -lane and does not prove protected-ingress behavior or the no-HTTP-fallback -instrumentation requirement. Windows Fabric and Win32 Paper measurements are -Not run on this host. Fault injection, args refusal, ingress bypass, and F3/F4 -survivor/fault cases were not run. - -## 2026-11-06 Windows Fabric and Win32 Paper descriptor evidence - -Source: the PR branch at `1f5333e25` plus `8458bde48` and `a6b89e924`; host -Windows 11 Enterprise build 26300, AMD64. The full 24-case matrix, counts, -retry history, run records, and cleanup are in the "Windows Fabric and Win32 -Paper qualification — 2026-11-06" section of -`packages/agentic/desktop-test/references/current-status.md`. -`yarn storybook build-driver --windows|--win32` succeeded, and -`yarn desktop-driver doctor --platform windows|win32 --permissions` reports -`--permissions` as **Unsupported** on Windows. No text-size or operation -latency measurements were recorded on Windows, so none are added here. - -Reference kinds, each from one driver session at service incarnation 1: - -| Endpoint | Target | Preview element refs | Chrome/window ref | Fresh ref after reset | -| -------------- | -------------------------- | -------------------------------------------------------------------------- | -------------------------------------------------- | ---------------------------------------- | -| Windows Fabric | `windows-f33cce69c1a7d0ae` | `9z4z23` (story root) and `9z4z66` became `ref-region-changed` after reset | `9z4z1` survived on `window-df4a4ab9-b2` (allowed) | `9z4z102` became stale after next select | -| Win32 Paper | `win32-5e01f2ae23282f34` | `okdw21` (story root) and `okdw38` became stale after reset | `okdw1` survived on `window-9951db02-b3` (allowed) | `okdw54` became stale after next select | - -After a Fabric Metro reload, the old story root `7fqk23` returned `ref-stale` -and a new story root `7fqk101` was observable in the same session. After a Win32 app exit, observing the old preview returned -`window-closed`. Physical input: the Windows helper rejected a physical click -because the target window did not become the foreground window (effect none); -the Win32 helper applied one with mechanism `physical-pointer`. - -## 2026-11-07 UTC Windows follow-up - -Source `386f0585edae2450ab22e6d359e2fe4db3399555`; validation tests, -tooling, atomic reporting, and guidance changed during this follow-up. The additive native -matrix in `packages/agentic/desktop-test/references/current-status.md` -records 21 Pass, 2 Partial, 1 Fail for Windows and 21 Pass, 3 Partial for -Win32, with no remaining Not run rows. This does not resolve Windows -physical input or the partial launch/secondary-window coverage. - -Real descriptor/MCP probes held correlated readiness for 1,500 ms after -native render and delivered a stale correlation first. Windows settled after -1,637 ms and Win32 after 1,618 ms. Within one uninterrupted session at -incarnation 1, regions changed 1 -> 2 -> 3: - -| Endpoint | Session | Preview element / chrome | Exact window | Runtime delayed/error generation | -| -------- | -------------------------------------- | ------------------------ | --------------------- | -------------------------------- | -| Windows | `46a48bed-5f9c-40c6-aac7-c2929461d222` | `0yoo23` / `0yoo1` | `window-4124f4f1-1be` | 2 / 3 | -| Win32 | `a29dc7af-c96f-4b0a-a98f-2eae04d7f8c1` | `oyno25` / `oyno1` | `window-5d019501-11b` | 5 / 6 | - -The preview refs returned `ref-region-changed`; chrome checks passed. -Correlated render errors produced unknown-effect quarantine and rejected -later story mutations. A fresh preview root was not exposed after that error, -so that subcase is Partial, not a bypass/fencing proof. Win32's earlier -fresh-preview fencing gap remains Partial. Native Callout probes confirmed a -shown native window and cleanup by subsequent selection, but addressed only -the registered primary identity: separately registered secondary-window -selection remains Partial on both endpoints. - -Startup and shutdown failure probes retained exact owned-resource claims and -refused replacement. Original stop callbacks reconciled the recorded -generation-2 channel/helper; the retained owner and claim were removed only -after resource cleanup proof. Fresh service incarnations then passed -descriptor discovery, marker-verified selection/reset, normal shutdown, -authority revocation, and attached-app preservation. The desktop-test ledger -retains the injected failures, one local control-reader crash, operator -reconciliation details, every supervisor run ID, and the separate unprotected -smoke history. No new physical-input or capture capability is inferred. - -Final non-agent smoke with bounded atomic-report recovery passed on both -endpoints: Windows run `1825d9aa-2c25-4a2c-8524-ceadcf52ad68` -(234 passed, 78 skipped) and Win32 run -`8c6f58e1-3591-4f1b-b008-c52a718d3ca2` (214 passed, 52 skipped). -Every session had a successful delete; readiness traversal ended before -authored tests. Windows journaled one recovered sharing conflict, Win32 none. -Processes, listeners, per-user claim, and default descriptor were absent -after cleanup. Artifacts remain under -`apps/storybook/artifacts/desktop//run.json`; earlier reporting and -physical-input failures are not erased. - -## 2026-11-08 UTC Windows application ownership and foreground diagnosis - -Source `9d3e11516` plus the Windows helper changes for -[#4441](https://github.com/microsoft/fluentui-react-native/issues/4441) and -F10-W ([#4447](https://github.com/microsoft/fluentui-react-native/issues/4447)). -Same host, with the console session disconnected throughout. - -- **Foreground diagnosis (#4441).** Activation now re-attaches to the - foreground holder when it changes during an attempt, and - `window-activation-failed` names the holder by opaque handle and process id - with its cause. The helper self-test covers each cause; a live physical - click on Windows Fabric was not rerun, so that row stays Fail. -- **Application ownership (F10-W).** Leases report `containment`. Unpackaged - launches run in a helper-owned Job Object and end with the helper; - packaged launches are `uncontained` and survive it in a fenced claim; - attach is `external` and fails closed. The fixture contracts and live REX - and packaged-Storybook launch probes passed, which moves launch-capable - discovery to Pass on both endpoints. Runs and survivor handling are in - `packages/agentic/desktop-test/references/current-status.md`. -- **Measurements.** Windows Fabric and Win32 Paper discovery, observe, delta, - and toggle text and latency are now recorded in - `packages/agentic/desktop-driver-mcp/references/current-status.md`. On - Windows Fabric three React Native Test App menu-bar items get new - references on every read, so an idle `delta` is never empty. - -## 2026-11-08 UTC connected Windows qualification - -Source `fec94cec4995c7372b798877c5412ddcb156523b`, Windows 11 Enterprise -build 26300, AMD64, with the console connected. The opt-in `nativeDriver` -contract passed 6 tests with 2 skipped. Under an exclusive workload claim, -`FURN_NATIVE_DRIVER_TEST=1 yarn test win32Fixture --runInBand` passed all -11 tests; retained elements were 20 at cycle 100 and 23 at cycle 500. -The full `FURN_NATIVE_DRIVER_TEST=1 yarn test --runInBand` suite then passed -909 tests with 31 skipped across 65 passed and 2 skipped suites. Its separate -fixture run retained 20 elements at both cycles 100 and 500; the exclusive -claim was released. - -A separate generic Win32 fixture probe (`dfb100b1-94b2-4007-946e-9172c1f97d54`) -switched between the exact main and secondary dialog handles, read controls -in each, and verified the dialog closed. Capture of the minimized main -window returned `unsupported operation` without restoring it; an explicitly -restored window produced a valid PNG. This is not the Storybook F4 -separately registered secondary-window case, which remains Partial. - -Connected physical-input Storybook smokes passed on both endpoints, including -the checkbox and popup cases. The separate real MCP descriptor path also -passed two physical checkbox clicks and checked-state assertions on each -endpoint, and returned hash-verified target PNGs. This replaces the earlier -Windows physical-click failure for that bounded case, not for the full -native matrix. Run IDs, selection counts, receipts, and cleanup are in -`packages/agentic/desktop-test/references/current-status.md`; MCP session -evidence is in `packages/agentic/desktop-driver-mcp/references/current-status.md`. - -The Win32 authenticated bridge was disconnected after a settled story. -Chrome remained observable, but a fresh preview reference was not exposed -after the disconnect, so fresh-preview fencing remains Partial. The first -probe timed out after 10 seconds; bridge retirement occurred about 19 seconds -after injection. A later attempt waited for retirement and retained the -same Partial result. Physical Ctrl+C and DPI changes were not exercised. - -## Windows Fabric and Win32 Paper — #4441 connected-session qualification — 2026-11-12 UTC - -**Observed baseline:** `505a2187b0a2451aed7081c9260b343d55742f93`, -Windows 11 Enterprise build 26300, x64, Node 26.13.0. All native runs held -the machine-wide campaign/lock; physical gates before and after each run -reported an active, unlocked **remote RDP** session, not a console session. -The single attached monitor was 1728 x 1084, origin `(0,0)`, 96 DPI / 100%. -No display settings changed. The packaged Fabric app's registered native -build was `bdeb0b291`; comparison of the Windows app project and native-core -sources to the tested SHA showed only native-core -version/changelog changes. **Inferred:** native-build reuse was valid. -Metro, generated catalogs and protected bundles came from the qualification -checkout. Paper used the prebuilt REX 0.81.1 host. - -### Matching-OS F16 clear contracts - -F16 is issue #4467, implemented by merged PR #4468, -`ac20660ff039f4e0dc914706ff24f548610f30aa`. From the qualification checkout, -with `$CampaignTools` identifying the private campaign helper directory: - -```powershell -$QualificationRoot = (Get-Location).Path -$env:TEMP = Join-Path $QualificationRoot 'artifacts\native-working' -$env:TMP = $env:TEMP -$env:TMPDIR = $env:TEMP -$env:FURN_NATIVE_DRIVER_TEST = '1' -pwsh -NoProfile -File "$CampaignTools\with-native-lock.ps1" ` - -Lane W3 -Label 'W3 qualification native strict-clear and fixture contracts' ` - -RequirePhysical ` - -WorkingDirectory (Join-Path $QualificationRoot 'packages\agentic\desktop-driver') -- ` - yarn test --runInBand nativeDriver win32Fixture operations.contract -``` - -**Passed:** 3 suites, 27 tests; **skipped:** 2 macOS-only signed/unsigned -build cases. This is one shared Windows-provider/Win32-fixture run, not 27 -tests per renderer. The real helper self-test negotiated `strictClear` and -covered semantic/keyboard mechanism selection and no-dispatch refusal; -the Windows native provider serves both endpoints. Helper artifact SHA-256: -`d8ee05e4796fb3aa22f7b5dbf481bd8cdfb0c20b6d5d24676d689dc7900d5b3c`. -The self-test's keyboard-denial instrumentation is not itself successful -React Native keyboard deletion. - -Separate live endpoint probes **passed** actual nonempty RN keyboard clear -(keyboard receipts 1 -> 2), accessible clear of an **initially empty** -TextInput, and readonly accessible clear refusal with effect `none` and -unchanged receipts. Nonempty semantic removal is native self-test evidence, -not a claim about the initially-empty RN probe. - -### Secondary targeting, capture and remaining gaps - -All rows below are **Observed** at the baseline SHA: - -| Case | Windows Fabric packaged | Win32 Paper REX | -| ---------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------- | -| Separately registered secondary | Passed: exact new opaque handle, explicit switch, complete 11-node native observation, unique `modern-callout-focus`, physical-pointer click, `focus: confirmed`, then handle removal | Not run / partial: real Callout shown, but only main registered; no secondary targeting claim | -| Additional secondary semantic activation | Failed: applied/accessibility on the exact unique button, but six checks / 2090 ms and final native observation stayed `Idle` | Not run: no registered secondary | -| Minimized capture | Passed: `unsupported operation`, still minimized, no implicit restore; explicit restore then valid 586 x 793 capture | Passed: same refusal/restore contract; restored capture 1010 x 761 | -| 200% scale | Not run: 100% RDP scale; changing it was not authorized | Not run: same | -| Mixed DPI | Not run: only one attached monitor | Not run: same | -| Negative-origin layout | Not run: only an origin-zero monitor; no approved eligible layout | Not run: same | -| Physical console Ctrl+C | Not run: owned by the separate implementation lane | Not run: same | - -The semantic failure was reproduced **without** switching foreground back -to main: main status was read through its explicitly cited native root. -The same selector and status assertion passed with physical click in one -attempt / 79 ms. **Inferred:** the renderer/accessibility-pattern dispatch -path may acknowledge activation without reaching Pressable `onPress`. -`native/windows/src/automation.cpp` tries several patterns and accepts a -successful HRESULT; the precise selected pattern was not instrumented. -An applied receipt alone is not application-level effect proof. - -On Win32, exact-process native inventory **observed** a visible, untitled, -root `Net UI Tool Window` with ToolWindow true. **Inferred:** the deliberate -untitled `WS_EX_TOOLWINDOW` exclusion in `CollectWindow` explains the missing -registration. Modern Callout was absent from the actual Win32 catalog. -No native title/style manipulation was attempted. Both open findings are -being fixed on this connected-session branch; this entry records the -unfixed baseline, not successful fix qualification. - -Exact endpoint supervisor invocations, each inside the same W3 -`with-native-lock.ps1 -RequirePhysical` wrapper: - -```powershell -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs win32 -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only --physical-secondary -``` - -Retained artifacts are relative to that qualification checkout, not shipped -package files: `artifacts\desktop\w3-qualification\runs\\`. - -| Evidence | Qualification ID | Report in that directory | -| --------------------------------------------------------------------------------- | -------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- | -| Fabric clear / minimized capture | `5c97fc26-71ef-483c-aa2a-b454d6ff4458` | `descriptor-cases-windows-clear.json`, `descriptor-cases-windows-minimized-capture.json` | -| Paper clear / minimized capture / missing secondary | `22332065-6fa3-4061-a2b6-93f83295ca2e` | `descriptor-cases-win32-clear.json`, `descriptor-cases-win32-minimized-capture.json`, `descriptor-cases-win32-secondary.json` | -| Fabric semantic failure, driver run `a9f2e71e-71ab-40f0-a375-fd6d76a0b971` | `675160f7-e069-413c-a6c8-486de0c37b69` | `descriptor-cases-windows-secondary---modern.json` | -| Fabric physical secondary pass, driver run `16aac92b-bb36-4742-bd29-01df381d9594` | `7fbe3b0b-df33-476b-80e3-95b97bf19f02` | `descriptor-cases-windows-secondary---modern---physical.json` | - -The physical secondary pass does not erase the semantic failure or earlier -receipt-only/investigation failures. The full matrix and injected Win32 -fresh-preview fence are recorded in -`@fluentui-react-native/desktop-test/references/current-status.md`; real MCP -click costs are in -`@fluentui-react-native/desktop-driver-mcp/references/current-status.md`. -Exact owned applications, helpers, Metro and service processes drained. -The campaign ended; pre/post-end audits both exited 0: -`FURN-AUDIT: no FURN desktop processes found`. No physical Ctrl+C claim is -made from the supervisor's cooperative programmatic shutdown. - -## Windows Fabric and Win32 Paper — connected-session fixes — 2026-11-12 UTC - -**Observed test base:** `a1cee02547b7aae4c72134902fe43ed007215069` plus the -then-uncommitted fixes described here. This section supplements, rather than -rewrites, the unfixed qualification above. No macOS qualification is added. -The native runs used the serialized W3 campaign/lock. Eligible runs reported -an active, unlocked **remote RDP** session on `Default` before and after; -they were not console-session tests. The packaged Fabric build at -`bdeb0b291` was reused: neither the Windows app native project nor native-core -Windows native sources differed. Metro and bundles came from the fix -checkout. Paper used the prebuilt REX 0.81.1 host; its native build remains -unsupported. - -### Observed causes and corrections - -Instrumented Fabric and Paper runs -`6488435c-6222-4c35-a670-d50655753da4` and -`ecd1cdff-6b77-4c42-ad57-89733dc154d0` found no Invoke, Toggle, -SelectionItem, or ExpandCollapse pattern. LegacyIAccessible was selected: -its default-action read returned `S_OK` and a **nonempty** action, and -`DoDefaultAction` returned `S_OK`, but the application did not change. -This occurred on main-window Pressables and agentic Buttons on both -endpoints, and on the Fabric secondary button. The saved fixed-format traces -are `native-trace-21424.jsonl` and `native-trace-38384.jsonl` under -`artifacts\desktop\w3-qualification`. - -**Inferred from that dispatch trace and the independent effect checks:** a -nonempty legacy default action is not sufficient activation evidence. -An empty-action-only guard would not repair the observed defect. The helper -therefore refuses legacy-only activation before mutation with -`unsupported-operation`, reason `activation-unsupported`, and effect `none`. -The four explicit actionable pattern paths remain unchanged. The public -error preserves its class, native code, and reason; it records no delivery, -does not quarantine, and never implicitly retries physically. Temporary -native tracing was removed. - -The Paper inventory also proved an exact-process, visible, untitled, -280 x 144 `Net UI Tool Window` popup owned/root-owned by the main window. -The strict untitled-tool exclusion originated in provider commit -`5b3f74c7b`; the available history supplied no more specific rationale. -The correction leaves global discovery conservative and qualifies only -lease-local popup roots using verified primary identity, exact-process -ownership, positive visible bounds, helper-class exclusions, and a bounded -UIA exact-root/content check, followed by native identity rechecks. - -The final helper registered that real Paper secondary. Its unique Callout -content observation was complete at eight nodes; its public `find` proved -the actual focus **button** unique, cited activation refused honestly, and -story cleanup removed the secondary handle. The host's whole-window UIA tree -also exposes the owner subtree, so this is scoped popup-content evidence, -not a claim that the default whole-window observation was complete. -Stable focus/blur test IDs were added to the legacy Callout story rather -than weakening the public uniqueness-proof requirement. - -### Red/green checks and helper identity - -All commands below ran from `packages\agentic\desktop-driver` unless stated -otherwise. Native commands used the W3 lock/campaign; interactive runs also -used `-RequirePhysical`. Test working directories were confined beneath the -ignored checkout `artifacts\desktop\w3-fix\native-working`. - -```powershell -$env:FURN_NATIVE_DRIVER_TEST = '1' -yarn test --runInBand operations.contract win32Fixture.contract --testNamePattern 'nonactionable legacy|legacy-only|untitled tool roots' -yarn test --runInBand operations.contract win32Fixture.contract nativeDriver -yarn desktop-driver build-driver --platform windows -``` - -- **Expected red:** five behavioral failures, 18 skipped, in - `native-passive-red.log`: MSAA no-action and ScrollIntoView-only UIA - activation both falsely succeeded through raw/public routes, and the - owned content tool root was missing. -- **Passed on the initial fixed helper:** five new cases, 18 filtered cases - skipped (`native-fix-focused-green.log`); then three suites, 32 tests - passed and two macOS build/signing cases skipped - (`native-fix-regressions.log`). This includes existing operation, fixture, - cancellation, retention, capture, and helper build/handshake checks. -- That initial helper was artifact - `29b67c2c775f51cd360016c6abd09eb2ca377b7868b0cd7c6feb7432799e871e`, - source digest `e99e3246963a0fa4848a31c5753a026200492d9ad92c036a132bba16f9d1b469`. -- **Passed final helper build:** after adding an exact UIA native-root - handle check, artifact - `b47212dcd9e63aba388907b930da19d9a6d602f7d3de32e364f1584890f4d6dc`, - source digest `16ecdda1e41112cf2dafaf57eaad843054fbeeb2e9add0518f9c41886e8c867c`, - build fingerprint `3628b82e4280727dce2b265db184baa52285cc9f88f571621182d3893b051397`. - The build log is `native-fix-final-build.log`. The real Paper pass below - used this final helper and story blob - `f96de651a8433fc446f9c02b7b5ee3f5075c945c`. -- The **final-helper native-contract rerun is Not run**, not a replay of - the initial helper's 32 passes: the desktop was paused after disconnect. - Logs in this list are relative to `artifacts\desktop\w3-fix`. - -### Endpoint results and session boundary - -The copied qualification supervisor ran these exact commands from the -checkout root, inside the W3 `-RequirePhysical` wrapper: - -```powershell -node artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only --activation-diagnostics -node artifacts\desktop\w3-qualification\run-endpoint.mjs win32 --secondary-only --activation-diagnostics -``` - -| Case | Outcome and evidence | -| --------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Fabric main Pressable and agentic Button | **Passed** honest refusal on the initial fixed helper; original statuses remained `Closed` and `0`, with unchanged receipts. | -| Fabric modern secondary semantic action | **Passed** honest refusal on the initial fixed helper: exact secondary handle, complete 11-node observation, unique button, no delivery/quarantine, main status stayed `Idle`, handle retired. | -| Paper main Pressable and agentic Button | **Passed** honest refusal on the final helper; original statuses remained `Native window: Shown` and `0`, with unchanged receipts. | -| Paper secondary registration, observation, semantic action, retirement | **Passed** on the final helper: main plus one secondary, complete eight-node content sample, unique focus button, `activation-unsupported`/effect `none`, no receipt/quarantine/fallback, then main alone. | -| Final-helper Fabric semantic rerun | **Not run**: first startup failed before readiness with `spawn rnx-cli ENOENT`; dependency resolution confirmed the CLI was installed. Its one retry was session-gated before the supervisor started. | -| Final-helper Fabric and Paper secondary physical clicks | **Not run (session disconnected)**. No physical input was attempted after the gate. The earlier baseline Fabric physical pass remains valid; it is not a final-helper rerun. | -| DPI, mixed-monitor, negative-origin monitor layout, physical console Ctrl+C | **Not run** by this fix campaign; no new claim beyond the baseline section. | - -Artifacts are beneath -`artifacts\desktop\w3-qualification\runs\\`: - -| Evidence | Qualification ID | Driver run ID | -| -------------------------------------------------------- | -------------------------------------- | -------------------------------------- | -| Initial fixed Fabric main/secondary semantic pass | `a583840b-7ad5-4691-9e60-f4f83d49fd48` | `0644505e-3642-4fed-b6aa-157275fb519f` | -| Final fixed Paper main/secondary semantic pass | `7edaa34a-fefd-4a8b-846b-dc8aa8c3880e` | `2d342e8c-dd72-4553-bc90-af070716add2` | -| Final Fabric startup failure, no app-level case executed | `c807faf4-31e8-4aa1-a3de-5b78e720e20f` | `07fe422f-7c94-4bbc-b6e8-379b9ee989c7` | - -Each successful directory retains `qualification.json` and -`descriptor-cases--main-activation.json` / -`descriptor-cases--secondary[---modern].json`, with cleanup -acknowledgments. Earlier failed Paper probes were qualification-harness -failures, not passes: the owner subtree exhausted the observation byte -budget, searching _inside_ the unnamed button did not prove that button, -and an observation sample was correctly refused as `sample-not-unique`. -The successful run uses a real unique `find` with the stable story test ID. -The optional managed UIA probe now serializes nonfinite geometry as null; -its bounded 32-node content result is explicitly incomplete, not a complete -window-tree claim. - -At **2026-11-12T17:03:45Z** the wrapper observed session 1 as remote, -Disconnected, locked, LogonUI running, with no input desktop. It returned -75 / `NOT RUN` before starting the retry. All native/GUI/physical work was -then paused; no reconnect or session move was attempted. The campaign ended -and both audits passed: -`FURN-AUDIT: no FURN desktop processes found`. - -### Portable checks and remaining qualification - -**Passed:** desktop-driver `yarn test --runInBand` with native opt-in unset: -73 suites / 1,201 tests, with eight native suites / 93 native tests -**skipped**, not passed. Native-core `yarn test` passed its build, story -typecheck, 18 Node checks, 56 Jest checks, and one snapshot; native-core -format/lint passed. These do not replace native qualification. - -From the checkout root, the final portable checks **passed**: - -- `yarn workspace @fluentui-react-native/desktop-driver format` -- `yarn workspace @fluentui-react-native/desktop-driver lint` -- `yarn workspace @fluentui-react-native/desktop-driver build` -- `yarn oxfmt --check` and `git diff --check` -- `yarn markdown-link-check -c .github\markdown-link-check-config.json ` - for all ten changed Markdown files, including both patch changesets: - 72 links checked. - -One permitted session check at **2026-11-12T17:21:31Z** reconfirmed remote -session 1 as Disconnected, locked, LogonUI running, and without an input -desktop. No native or GUI retry followed. - -The implementation is committed as -`cd036a2d06ae450f9c26ac0837c4fe140877ffb9`. Its committed -`yarn change:check` **passed**. A post-commit process audit also passed: -`FURN-AUDIT: no FURN desktop processes found`; campaign status reported no -campaign or holder. The branch has not been pushed. - -The post-review clean-cache repository aggregate and final native/physical -reruns are still pending. Outstanding qualification stays tracked in -[remaining work](remaining-work.md) / #4441; no issue is closed and no -final-helper physical-success claim is made here. - -## Windows Fabric and Win32 Paper — popup liveness review fixes — 2026-11-12 UTC - -**Observed source base:** `96b91661e2886aa4a22bd3f08b485e51d11c90d7` -plus the then-uncommitted review fixes below. This section supplements the -earlier Windows evidence; it does not rerun or relabel that evidence. The -earlier section's final helper was the pre-review `b47212dc…` artifact. -No macOS code, ledger section, or qualification changed. - -### Admission, liveness, and publication - -Registered untitled popups now bypass UIA content discovery when their lease, -process, thread/class, primary identity, and owned-content native predicates -still match. A failed content read or temporary re-layout therefore cannot -remove a live, admitted popup from `windows`. New candidates share one -1,000 ms deadline per listing; UIA waits use its remaining time, late reads -stop before another transaction, and unprobed candidates retry later. -Exhausting discovery does not skip registered windows later in the listing. -Publication diagnostics now include admitted live popups as `listed` as well -as `tracked`, without another content probe. - -**Test-first authoring:** the gated fixture regressions and native self-test -were added before production edits. The self-test injects false, throwing, -and unavailable content probes, an exhausted shared deadline, a later retry, -and listed/tracked publication assertions. The fixture checks hung-provider -popup retention/rectangle reads and bounded new-popup discovery/recovery; -its existing admission case also checks publication. Native red/green -execution is **Not run**, not inferred as passed: the desktop remains paused. - -### Compile-only and portable validation - -The helper-only build used the shared W3 lock, no campaign, and no launch. -The following build command is shown with its cache-root argument relative to -the driver package; the recorded invocation resolved that path absolutely: - -```powershell -yarn desktop-driver build-driver --platform windows --cache-root ..\..\..\artifacts\desktop\w3-review\native-cache -``` - -**Passed** x64 release compilation, including the new native self-test. -Lock acquisition was at `2026-11-12T17:45:53Z`, and the wrapper released -successfully with exit 0. No helper, fixture, REX, packaged app, or Metro was -launched. This is compilation, not native runtime qualification. - -- Artifact: `eb91ef8a4164080bcae787aacce3fd5aa5556d8c3d0b22d0cd0c661a50934066` -- Source digest: `89cc81fbb72cb8a79852e5d4d90ce96d8b2982283086236b860289f2664ac557` -- Build fingerprint: `7b38faa5c0eeba1ace7441d613a3f0c126915f78eff58c4a32466825c4e11854` -- Build log: `artifacts\desktop\w3-review\helper-build.log` - -From `packages\agentic\desktop-driver`, with native opt-in unset: - -```powershell -Remove-Item Env:FURN_NATIVE_DRIVER_TEST -ErrorAction SilentlyContinue -yarn test --runInBand win32Fixture nativeDriver -yarn test --runInBand -``` - -**Passed:** focused portable checks, eight tests; 20 native tests **skipped**. -The full portable suite passed 73 suites / 1,201 tests, with eight native -suites / 95 tests **skipped**, not passed. Logs are -`tests-authored-portable.log`, `focused-portable-final.log`, and -`portable-tests.log` beneath `artifacts\desktop\w3-review`. - -From the checkout root, **passed**: - -```powershell -yarn workspace @fluentui-react-native/desktop-driver format -yarn workspace @fluentui-react-native/desktop-driver lint -yarn workspace @fluentui-react-native/desktop-driver build -``` - -The README, protocol, helper/fixture references, and patch changeset describe -the same policy. FDR1 fields, commands, and `native/protocol.json` remain -unchanged. Root `yarn oxfmt --check`, `git diff --check`, and -`yarn markdown-link-check -c .github\markdown-link-check-config.json ` -**passed** for all seven changed Markdown files (66 links). -The review-fix implementation is committed as -`4d39582f2292ff6436dafaba646ee63cb9bdc3cc`. -Its committed `yarn change:check` **passed**; the log is -`artifacts\desktop\w3-review\change-check.log`. -The process audit passed with -`FURN-AUDIT: no FURN desktop processes found`; no W3 campaign was started, -and shared campaign status reported no campaign or holder. No push was made. - -### Remaining session-gated runs - -One permitted check at **2026-11-12T17:52:11Z** observed remote session 1 as -Disconnected, locked, LogonUI running, with no input desktop. No GUI or native -contract was attempted; no reconnect or session move was attempted. - -Still **Not run (session disconnected / W3 desktop grant required)**: - -- Final-helper Fabric main/secondary semantic case. -- Fabric and Paper secondary physical clicks, under `-RequirePhysical`. -- Final native selector below, including the new failed/hung-probe, - shared-deadline/retry, and publication regressions plus the native self-test. - -```powershell -$env:FURN_NATIVE_DRIVER_TEST = '1' -yarn test --runInBand operations.contract win32Fixture.contract nativeDriver -``` - -The clean-cache repository aggregate is **Not run** in this review-fix -segment; W1 holds the buildci-only grant. It remains required before -publication. The native cases remain in [remaining work](remaining-work.md) -and #4441; no new native pass, physical success, or issue closure is claimed. - -## Windows Fabric and Win32 Paper — default-environment publication validation — 2026-11-12 UTC - -**Observed tested source:** `cad25f8c53855c30f2c45f3ce6655f668d09bdc9`. -The coordinator authorized buildci-only use of the shared W3 lock, followed -by draft publication. No app, native contract, helper runtime, or physical -input was launched in this segment. Native opt-in remained unset. TEMP and -concurrency used their defaults; no workaround, guidance, or timeout edits -were made. - -The exact root validation command, inside the W3 machine-wide lock: - -```powershell -cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci" -``` - -### Failure comparison and unchanged-source retry - -- **Failed, first branch attempt:** lock acquired at `18:33:59Z`, released - at `18:36:20Z`; 136 tasks succeeded and Storybook desktop tests failed. - Four checks failed: the WDIO worker's 7,000 ms deadline, two 10,000 ms - diagnostic timeouts, and the server-integration stop result. -- **Observed cleanup:** Lage exited before its desktop-test descendants. - No independent WDIO command had been started outside the lock. The - descendants finished naturally; their recorded root PID was already gone - when checked. A subsequent process audit exited 0. -- **Failed on clean baseline too:** the same default-environment command - ran on isolated detached `505a2187b0a2451aed7081c9260b343d55742f93`. - It reproduced the WDIO deadline and server-stop readiness race; baseline - diagnostics passed. There were 138 successful tasks and one failed task. - All baseline children drained before the lock released at `18:45:14Z`. - The disposable checkout remained clean and was removed afterward. -- Storybook source/config, desktop-test source/config, scripts, lockfile, - and Lage configuration were byte-unchanged between the baseline and this - branch. **Inferred:** these load-sensitive failures were not caused by - the Windows activation/popup changes; no unrelated source fix was made. -- **Passed, unchanged-source branch retry:** the lock was acquired at - `18:48:10Z`, validation started at `18:48:11Z`, and all **178 tasks passed** - at `18:50:53Z` (zero failed, aborted, pending, or skipped Lage tasks). - Owned descendant drain completed before lock release. This is a task-level - aggregate result, not a claim that opt-in native cases executed. - -The local ignored supervisor `run-buildci-and-drain.ps1` runs the exact -command above and keeps the wrapper alive until owned Node descendants -finish. It does not alter the test environment or send input. Logs are -retained beneath `artifacts\desktop\w3-buildci`: -`buildci-default.log`, `baseline-buildci-default.log`, and -`buildci-default-rerun.log`. - -At `2026-11-12T18:52:16Z`, the post-run audit exited 0: -`FURN-AUDIT: no FURN desktop processes found`. Shared state reported no -campaign or holder. W3 started no campaign; its build-only lock was released. - -### Publication scope and remaining native work - -The final local Fabric semantic case, both final-helper secondary physical -clicks, and final native contracts remain **Not run**. The earlier disconnect -prevented them; after the session became Active/unlocked again, the desktop -was reserved for W1, not a W3 native campaign. No reconnect, session move, -display change, or native retry occurred here. The current review helper -remains compile-only as recorded above. - -The default-environment aggregate requirement is now satisfied. Draft -publication does not close #4441 or replace its outstanding native, 200%, -mixed-DPI, negative-origin, or separate #4491 console-Ctrl+C qualification. - -## Windows Fabric and Win32 Paper — hosted popup compatibility follow-up — 2026-11-12 UTC - -**Observed published revision:** `4208ab36d641bcfe3675836634f18dee2072ef49`, -draft PR #4498, workflow run `43284991535` (attempt 1). -**Tested portable correction:** `28c7690085342a3bc2ee114b2c4c2032d7d050c4`. -No native source or FDR1 schema changed in this follow-up. - -### Hosted evidence at the published revision - -- **Passed:** Windows Storybook's opt-in package command, - `yarn workspace @fluentui-react-native/desktop-driver test --runInBand` - with `FURN_NATIVE_DRIVER_TEST=1`: - 79 suites / 1,251 tests; two suites / 45 tests **skipped**. The Windows - fixture, operations, native-driver self-test, and other matching-OS native - suites actually ran. This includes the registered-popup hung-provider, - deferred-discovery, and listed/tracked publication regressions, rather than - merely compiling them. These are hosted fixture results, not the remaining - local real-application or physical-secondary reruns. -- **Passed:** Windows Storybook accessible smoke: 255 executed / 255 passed, - no failures, skips, unsupported cases, or retries. Internal navigation - diagnostics were retained; this count is the final selected-test summary, - not a claim that every diagnostic navigation attempt succeeded. -- **Passed:** Win32 Storybook physical smoke: 288 passed and one - unsupported/quarantined case; no failures or retries. The helper source - digest was `89cc81fbb72cb8a79852e5d4d90ce96d8b2982283086236b860289f2664ac557`, - matching the review helper. -- **Passed:** Windows legacy E2E, JS PR, both Win32 Storybook jobs, publishing, - links, changesets, and CodeQL. Other hosted jobs were still completing - during this follow-up; no whole-workflow pass is claimed. -- **Failed, deterministic:** Win32 legacy E2E executed 201 tests: 191 passed, - ten menu tests failed only at the final “An assert popped up” safety check. - Their menu-open/state checks had passed. The existing guard treated every - second window as an assertion dialog. All 20 retained failure captures - (including suite retries) showed the secondary accessibility root as `menu`. - Correctly registering native menu windows exposed that consumer assumption; - no flake rerun was requested. - -Logs and the bounded downloaded E2E artifact are retained under -`artifacts\desktop\w3-buildci`: `ci-windows-storybook.log`, -`ci-win32-physical.log`, `ci-win32-job.log`, and `ci-win32-e2e`. - -### Test-first correction and boundaries - -`DesktopSessionClient.getWindowTree` now reads a currently listed session -window without selecting or activating it. The server refreshes authoritative -window membership and rejects unknown/closed IDs before tree inspection. -Native focus, selected-window identity, and input state are unchanged. - -The Windows/Win32 legacy guard exempts only a single `menu` root without -nested dialogs or windows. Unrecognized secondary roots still fail; unreadable -windows reject rather than report no assertion. A menu cannot hide another -unexpected window. macOS, non-native runners, Android, and the native-task -library bypass keep their earlier behavior. Diagnostic window events are -not used as authority. The public route is documented in -[the WebDriver contract](protocol.md); it is a full tree query, not a budgeted -agent operation. - -From `apps\E2E`, `yarn test` first produced **two expected failures** for a -legitimate menu and an unreadable secondary. From `packages\agentic\desktop-driver`, -the two focused new routing cases first failed because the read-only route -did not exist. Logs: `e2e-assert-red.log`, `window-tree-red.log`. - -| Exact command | Result | -| ------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------- | -| `yarn test --runInBand desktopDriver.test DesktopDriverClient.test` (desktop-driver) | **Passed:** 52 tests, including non-activation and unknown/closed-window refusal. | -| `yarn test --runInBand` (desktop-driver, native opt-in unset) | **Passed:** 73 suites / 1,203 tests; eight suites / 95 native tests **skipped**. | -| `yarn workspace @fluentui-react-native/e2e-testing test` | **Passed:** all 11 adapter tests, including menus, possible assertions, failed reads, and unchanged other-runner behavior. | -| `yarn workspace @fluentui-react-native/desktop-driver build` | **Passed**. | -| `yarn workspace @fluentui-react-native/e2e-testing build` | **Passed**. | -| `yarn workspace @fluentui-react-native/desktop-driver lint` | **Passed**. | -| `yarn workspace @fluentui-react-native/e2e-testing lint` | **Passed** after marking the mocked loopback URL as test-only. | -| `yarn workspace @fluentui-react-native/desktop-driver format` | **Passed**. | -| `yarn oxfmt --check` | **Passed**. | -| `git diff --check` | **Passed**. | -| `yarn markdown-link-check -c .github\markdown-link-check-config.json ` | **Passed:** E2E README and public protocol, 25 links. | -| `yarn change:check` | **Passed** after committing the implementation and patch changeset `quiet-menu-windows.md`. | - -The new correction is committed locally and awaits coordinator review before -push. Its clean-cache aggregate and affected hosted E2E rerun are **not run -yet**; the earlier 178-task aggregate does not qualify this new source. -W1 holds the live desktop campaign. W3 has not launched an app, native helper, -fixture, WDIO suite, or physical action in this portable correction segment. -The final local Fabric semantic case, both final-helper secondary physical -clicks, and requested local native selector remain **not run**. Hosted -fixture passes do not substitute for those real-app/physical cases. - -The shared process audit returned 2: it listed W5's test tree and the W3 -formatter running concurrently with the audit. Both recorded W3 formatter -PIDs were confirmed exited afterward. No W3 native runtime or campaign is -left running; a globally clean audit is not claimed while peer work is active. - -## Windows Fabric and Win32 Paper — closing-menu review correction — 2026-11-12 UTC - -**Reviewed base:** `ad8b67c9720a00d7d8df4d27c2ca7436c78d6198`. -**Tested implementation:** `36efc5a88dc507c626f2d29ac487cfc37e9e4bf3`. -The coordinator's review accepted the scoped, non-activating tree query and -conservative menu classification, and identified a closing-popup race in the -legacy E2E adapter. - -The adapter now skips a tree read only when its typed WebDriver error is -`no such window`, then continues checking other handles. Before reporting an -empty, non-menu, or dialog-containing tree as an assertion, it refreshes the -authoritative window list and confirms that handle is still present. -Persistent unexpected windows still fail the check. Transport, provider, -and stale-element errors still reject; no diagnostic event or input fallback -is used. - -**Observed test-first:** `yarn test` in `apps\E2E` initially failed two new -cases: a `no such window` read rejected, and an empty closing-menu tree -reported an assertion. The retained log is -`artifacts\desktop\w3-buildci\popup-race-red.log`. -After the fix, all 14 adapter tests passed, including a vanished menu followed -by a real unexpected window, empty/non-menu roots followed by disappearance, -persistent unexpected roots, and unreadable still-listed windows. - -| Exact command | Result | -| --------------------------------------------------------------------------------------------------------------- | --------------------------------------------- | -| `yarn workspace @fluentui-react-native/e2e-testing build` | **Passed**. | -| `yarn workspace @fluentui-react-native/e2e-testing test` | **Passed:** 14 tests, zero skips or failures. | -| `yarn workspace @fluentui-react-native/e2e-testing lint` | **Passed**. | -| `yarn oxfmt --write apps\E2E\src\common\nativeDesktop.ts apps\E2E\desktop-adapters.test.cjs apps\E2E\README.md` | **Passed**. | -| `yarn oxfmt --check` | **Passed**. | -| `git diff --check` | **Passed**. | -| `yarn markdown-link-check -c .github\markdown-link-check-config.json apps\E2E\README.md` | **Passed:** 19 links. | -| `yarn change:check` | **Passed** after the implementation commit. | - -This correction changes only the private E2E adapter, its focused tests, and -related documentation; no native helper or public protocol changed. The -existing driver patch changeset remains committed. -Per coordinator instruction, the locked clean-cache aggregate, native -contracts, GUI/physical runs, and push are **not run** in this segment. -W1 retains the desktop campaign. W3 awaits its explicit grant for the final -Fabric semantic case, both secondary physical clicks, and final native -contracts/probe tests, then the clean-cache aggregate and publication. - -## Windows Fabric and Win32 Paper — final sweep and native close-race correction — 2026-11-12 UTC - -**App qualification revision:** `3957cd6aec1cf83c8dc62e4d56403fe942505aa4`. -**Close-race implementation:** `a53451513eb84f5ff62dc7763bd419521c617d5b`. -The session had returned to Active/unlocked before this campaign. All native -commands below used the W3 campaign and serialized lock with -`-RequirePhysical`; both preflight and postflight reported an eligible -**remote** session with the Default input desktop. No session change was -reported. - -The registered Fabric native app was reused only after checking its install -identity and verifying no native app-source differences from its build -revision. Metro and bundles came from this worktree. Paper used the exact -REX executable with SHA-256 -`3ecd1f58690b029eccf06838dc23411f08b1f3765b99950091f8246aedd74725`. -Apps, Metro, helpers, and fixtures were closed between runs. - -The following results used helper source -`89cc81fbb72cb8a79852e5d4d90ce96d8b2982283086236b860289f2664ac557`, -artifact `eb91ef8a4164080bcae787aacce3fd5aa5556d8c3d0b22d0cd0c661a50934066`. -Run artifacts are under `artifacts\desktop\w3-qualification\runs\`. - -| Endpoint / case | Observed result | Run ID | -| ------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------- | -| Fabric main/secondary semantic action | **Passed:** honest `activation-unsupported` refusal with effect `none`; no fabricated receipt, fallback, or app change. | `98d0ae8f-e84b-43c1-b332-7136dd676f23` | -| Fabric secondary physical click | **Passed:** separately registered window, complete 11-node sample, physical-pointer receipt and public `focus: confirmed`, then retirement. | `537cc479-b3d3-4ef0-93b2-d39c35177653` | -| Paper secondary physical click | **Passed:** exact owned tool window, complete eight-node sample, physical-pointer delivery and retirement. The legacy story has no native-command outcome, so this is not proof of the RN focus command. | `e29ff4aa-62f5-4f53-8a5e-450a5e799466` | - -### Observed failure and test-first correction - -With `$env:FURN_NATIVE_DRIVER_TEST='1'`, the requested package command -`yarn test --runInBand operations.contract win32Fixture.contract nativeDriver` -**failed**: 33 passed, one failed, two macOS-only cases skipped. The tool-content -fixture failed with `no-such-window`, not a session or permission refusal. -Its log remains `artifacts\desktop\w3-buildci\final-native-contracts.log`. -This failed sweep is not claimed as a final native pass. - -A bounded diagnostic replay reproduced the failure on its first fixture. -The command journal isolated it to `windows` immediately after the close -button's successful accessibility click; the destroy event confirmed the -secondary had closed. Native enumeration can return a window that closes -before `WindowSummary` reads its geometry. This is separate from semantic -dispatch and the E2E adapter's closing-tree race. - -The correction omits a geometry failure only when its code is -`no-such-window` **and** native liveness confirms the handle has gone. A -still-live handle's error and every unrelated error propagate. Direct -commands for the omitted window still refuse. No input fallback, -cross-process admission, or wire-schema change was added. - -A deterministic native self-test closes a secondary exactly between -enumeration and geometry serialization. Before the guard, this command -**failed** its new live-window-listing assertion: - -```powershell -yarn test --runInBand nativeDriver --testNamePattern 'handshakes with the Windows helper' -``` - -After the guard, the following command **passed**, two native cases with 26 -unselected cases skipped. It includes the deterministic self-test and the -original real-fixture registration/action/closure scenario: - -```powershell -yarn test --runInBand nativeDriver win32Fixture.contract --testNamePattern 'handshakes with the Windows helper|registers only owned untitled tool roots' -``` - -The self-test also verifies direct closed-window refusal, propagation of an -error for a live handle, and propagation of an unrelated error even when its -window closes. The unchanged diagnostic replay then **passed six isolated -fixture lifecycles**, including close polling and subsequent main-window -actions. Retained logs under `artifacts\desktop\w3-buildci` are -`trace-tool-contract-replay.log`, `closed-window-listing-red.log`, -`closed-window-listing-green.log`, and `trace-tool-contract-green.log`. - -The fixed helper has source -`7119f42fb101452990b7bea13daf685e90a67ad75840fba60968a5efe42f62d1`, -artifact `e0e6fc988fc9214b91e02bd0909c099b203d4f9bfd89c45db8f8c9f9397a05e0`, -and build fingerprint -`afca889e68ccd07b9c1874b6d46b0c2a3a3f635e3e7b09c1b43eca20b9fb6f9c`. -The package `yarn build` and `yarn lint`, root `yarn oxfmt --check`, -`git diff --check`, and Markdown link checks for the native provider README -and protocol **passed**. Patch changeset: `brave-popup-listings.md`. - -The new implementation needs coordinator review before push. The complete -native selector, Fabric semantic case, and both secondary physical clicks -must be rerun against this **new** helper; the earlier app results above do -not qualify its new source digest. The new-source clean-cache aggregate and -hosted CI are **not run** yet. The 200% DPI, mixed-DPI, and negative-origin -gaps remain unchanged. Both campaigns ended, and the final audit reported -`FURN-AUDIT: no FURN desktop processes found`. - -## Windows Fabric and Win32 Paper — reviewed-helper final qualification — 2026-11-12 UTC - -**Tested revision:** `6c626602cf1eedd10425c8a63d794b1e12b1ff95`, clean worktree -for every real-app run. The coordinator's Opus review of the close-race -correction reported no findings. W3 received a new desktop grant after W2's -recovery verification; the session was **remote**, Active/unlocked, with -the Default input desktop and `physicalEligible: true`. - -All native and real-app commands below ran inside the W3 campaign and -serialized machine lock with `-RequirePhysical`. Every wrapper's before/after -session check stayed eligible. No hosted input skip or semantic action is -counted as physical evidence. - -The source-verified helper cache selected source -`7119f42fb101452990b7bea13daf685e90a67ad75840fba60968a5efe42f62d1`, -artifact `e0e6fc988fc9214b91e02bd0909c099b203d4f9bfd89c45db8f8c9f9397a05e0`, -and build fingerprint -`afca889e68ccd07b9c1874b6d46b0c2a3a3f635e3e7b09c1b43eca20b9fb6f9c`. -The app supervisor used prebuilt-only policy, not an implicit helper build. -The exact native app identity was checked before each run. The registered -Fabric app had no native source differences from its build revision; -Metro and bundles came from this worktree. Paper used the exact REX executable -with SHA-256 -`3ecd1f58690b029eccf06838dc23411f08b1f3765b99950091f8246aedd74725`. -The plain Win32 fixture executable had SHA-256 -`9e082f5a9a50e1a6f93f8d91a0c01f8e880820030a7e9d8b5c1917993fc8bc13`. -Apps, Metro, helpers, and fixtures were cleaned up between runs. - -The native package command below ran from `packages\agentic\desktop-driver` -with `$env:FURN_NATIVE_DRIVER_TEST='1'`. Endpoint supervisor commands ran -from the worktree root; their source and immutable run artifacts remain -under ignored `artifacts\desktop\w3-qualification`. - -| Exact command | Observed result | -| -------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| `yarn test --runInBand operations.contract win32Fixture.contract nativeDriver --verbose` | **Passed:** three suites, 34 tests; two macOS-only cases **skipped**. The previously failing tool-content closure case passed. The log is `artifacts\desktop\w3-buildci\final-native-contracts-fixed.log`. | -| `node artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only --activation-diagnostics` | **Passed:** main Pressable/Button and modern secondary refuse honestly with `activation-unsupported`, effect `none`, no delivery receipt, quarantine, physical fallback, or application change. Run `8f6d457c-c0d3-4ab5-9204-6dbf66383dbd`. | -| `node artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only --physical-secondary` | **Passed:** a distinct secondary is registered, its 11-node sample is complete, the click is applied through `physical-pointer`, the public `focus: confirmed` effect is observed, and its handle retires. Run `94c0b321-cebd-4e9b-99e2-3404bbbd80c3`. | -| `node artifacts\desktop\w3-qualification\run-endpoint.mjs win32 --secondary-only --physical-secondary` | **Passed for native delivery:** the exact owned tool secondary is registered, its eight-node sample is complete, a physical-pointer click is applied, and its handle retires. Run `f46d008a-d310-4e6e-98fd-e85941879edd`. The legacy story has no native-command outcome, so this is not proof of the RN focus command. | - -The full native selector executed the review-specific coverage, not just -compiled it: failed-probe retention and shared-deadline/retry self-tests; -the real hung-popup listing and bounded new-popup discovery/recovery cases; -truthful listed/tracked publication; the close-after-enumeration self-test -with live-error and unrelated-error propagation; and both legacy-only -activation refusals. Legitimate Invoke/Toggle and the existing observation, -capture, mutation settlement, containment, cancellation, and retention -contracts also passed. Cancellation completed in 32 ms; retained elements -stayed at 20 after 500 fixture cycles. The earlier failed sweep and -deterministic red/green evidence remain recorded above. - -### Final-source aggregate and hygiene - -After ending the native campaign and obtaining a clean audit, the exact root -command `cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` ran under -the W3 machine lock with **default TEMP and concurrency**, native opt-in -unset, and no policy, timeout, or guidance workaround. It **passed on the -first attempt**: 178 successful tasks, zero failed, aborted, pending, or -skipped tasks. Native tests remain opt-in in this aggregate and are qualified -by the separate 34-test run above. - -The aggregate ran from 22:30:40 to 22:33:27 UTC. The operational supervisor -retained the lock until its owned descendants drained; drain completed -before lock release. Its retained log is -`artifacts\desktop\w3-buildci\final-buildci-fixed.log`. No readiness-flake -rerun was needed. No code changed after the reviewed and qualified source; -this publication update changes only Windows evidence/tracking. - -The campaign ended. Both its post-campaign audit and the post-aggregate audit -reported `FURN-AUDIT: no FURN desktop processes found`. No W3 app, helper, -fixture, Metro, or WDIO process survived. - -The current-source local reruns are complete. Hosted results for the -publication head will be recorded in draft PR #4498, separately from -earlier-head hosted evidence. The 200% DPI, mixed-DPI, and negative-origin -cases remain **not run** for the previously recorded authorization/topology -reasons; console Ctrl+C remains separate work in #4491. - -## 2026-11-12 UTC macOS physical-input postconditions (F16) - -Part of [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441) -and [#4435](https://github.com/microsoft/fluentui-react-native/issues/4435). -This section is independent of the F15 ledger update in -[#4493](https://github.com/microsoft/fluentui-react-native/pull/4493). -The implementation commits are `6aa7a68ba`, `5baf9f735`, and `e0af46b38`, -based on `505a2187b`; they include menu-independent selection, the LiteUI -viewport fix, and bounded pointer acknowledgment. The separate renderer -prerequisite was first committed as `1492afe57`; the independently verified -backport branch starts at `505a2187b`, with patch commit `1eaa60a44` and -self-contained ledger commit `9393fe896`. Normal merge `adde47dfd` makes -that branch an ancestor of E. Its three-dot diff contains only E-owned -files, not the backport, manifest, lockfile, or Native Core documents. -All original implementation commits, including `1492afe57`, are retained. -The phase-1 -signed helper source digest is -`26c8f8e9060d0e10b26d6b7cab0ced556a172b29c1c14aac2429ce3e0773ebea`. -Local evidence is under `artifacts/desktop/f16-macos-physical/`, with consumer -reports under `apps/storybook/artifacts/desktop//`. - -### Findings and fixes - -- **Physical click: observed preview overflow and a separate identity defect.** - A controlled 480 × 302 Paper window reproduced interception. Raw - application-scoped and system-wide AX hit tests found the Controls addon - button over the checkbox, not the checkbox. F15's Accordion screenshots - show the same small-window, clipped-preview layout. Accepting any - hit-tested ancestor window would therefore dispatch input to the wrong - control. The guarded full physical run - `8f464c02-6e09-4bbc-8007-059310e01c93` then exposed the same class at - 932 × 782: Input FocusTreatment's entry field started at window-relative - y = 31.5 while its preview started at y = 76, and the actual hit was the - Tone combobox. LiteUI's fixed-height, centered decorator placed oversized - stories behind native chrome. Its macOS decorator now uses the - existing Win32 pattern: a scrollable viewport with naturally growing - content, centering only when there is spare space. Windows retains the - prior non-scrolling container; the review added an explicit platform - entrypoint and a regression that rejects the extra Windows scroller. - No hit-test ancestor is - accepted as the target and no story assertion changed. The unchanged - oversized Input callback now passes, including all native focus and - geometry assertions. At normal geometry the real descriptor's MCP - Checkbox click, checked state, and physical receipt pass. Separately, the native - `externalWindowSnapshot` incorrectly assigned an external hit the target - window's identity, preventing the agent operation's foreground activation. - Only hits proven to belong to another process now have distinct opaque - window identities. Same-process covering windows and unproven owners retain - the target's no-activation obstruction scope: raising through a popup could - dismiss or reorder it. Review regressions failed for both owner cases - before that correction; agent coverage verifies `obscured`, effect `none`, - and no activation or pointer input. Two attempts - to place a standalone external fixture in the foreground did not establish - the fixture precondition; they are not external-occlusion qualification. - At 480 × 302 the Controls addon still covers the Checkbox's native center, - even with the scrollable decorator; the driver correctly rejects that - unusable preview with effect `none`. The fix does not promise to click - through chrome or implicitly resize a caller's window. -- **Keyboard clear: false-success root and ignored shortcut observed; - menu-route explanation inferred.** The helper returned after posting - Command-A/Delete, without checking native focus, selection, or the empty - value. The first guarded full Paper physical run exposed the original - Input callback's ignored Command-A: focus and foreground were established, - the Select All menu item was enabled, but the selected range remained - `{ location: 3, length: 0 }`. Explicit Unicode Command-A, physical pointer - focus, and holding the chord until acknowledgment did not correct it. - Physical Command-Up followed by Command-Shift-Down selected the complete - value in the same field and cleared it. Keyboard clear now uses those - document-navigation chords once, without the application's Select All menu - route; the precise Paper menu-dispatch cause is not established. - It confirms focus, waits for complete selection when - `AXSelectedTextRange` is exposed, posts Delete once, and verifies an empty - value. Unreadable values refuse before keys; unacknowledged selection, - retained text, and cancellation retain the posted input's partial effect. - There is no key replay, semantic fallback, or story assertion change. -- **Foreground activation: inferred from F15 evidence, observed in the - regression.** F15 recorded a successful AXFrontmost setter with - `frontmost=false`; the old fallback ran only for setter errors. Activation - now checks the postcondition and requests workspace activation once when - the acknowledged AX write did not establish foreground state. The final - foreground/main-window checks remain mandatory. No Accordion story or - assertion was changed. -- **Pointer acknowledgment: an exposed timing gap, with inferred connection - to the full-run failure.** Run `c741b8d5-1a4b-49a0-b2f8-24ba25e8e8c1` - stopped at Button Default because the pointer was not at its requested - coordinate; the same callback passed on a fresh isolated run. Code checked - the asynchronous move only once, after 16 ms. A recorded-input regression - reproduced delayed OS observation, and another exposed lost cancellation - during that acknowledgment. Position is now read for a bounded roughly - 200 ms, without reposting the motion or pressing a button until verified. - The specific full-run delay/interference cause is not established. -- **Fabric renderer prerequisite: failing reader observed, concurrent writer - inferred.** Full physical and accessible qualification exposed a crash in - the unchanged PGB-006 window-resize callback. A fresh DerivedData/native - rebuild still failed. The owned-process AddressSanitizer report recorded - main-thread `BUS` in `folly::dynamic::type()` while `ShadowNode::clone` - copied `nativeProps_DEPRECATED` during AppKit layout. The installed fork - had an unguarded cloning reader and two UIManager writers. The - minimal upstream MIT backport documented in - `@fluentui-react-native/native-core/references/rnmacos-patch.md` gives - them the family-owned mutex from React Native commit - `923e7ddcaf4d2de45ca7dedafbb84c6b87dbffc4`, releasing it before parsing - and shadow-tree commits. The concurrent writer is inferred from source - and the upstream correction, not established by ThreadSanitizer. - The backport is required for the complete Fabric physical lane's PGB-006 - resize case, not for the three original input reproductions. - -### Native and focused qualification - -The responsible process tree had Accessibility, PostEvent, and Screen -Recording authority; permission diagnostics requested no prompts. Native, -GUI, physical-input, and aggregate commands use the shared native lock. -The helper is standalone: CocoaPods-module and Office bridging-header -qualification do not apply. - -- **Passed:** `swift build -c release -Xswiftc -warnings-as-errors` with a - worktree-local scratch root; helper self-test **461/461**. The first regression - run failed seven new checks; the external-hit regression then failed - independently before its fix. The menu-independent selection regression - failed **458/459** before its fix. Pointer acknowledgment and cancellation - regressed **459/461** before their fix. Logs: `native-regressions-red.log`, - `native-hit-regression-red.log`, `document-selection-red.log`, and - `document-selection-green.log`, `pointer-ack-red.log`, and - `pointer-ack-green.log`. -- **Passed:** driver `yarn format`, `yarn lint`, `yarn build`, and focused - `yarn test act.test sessionInput --runInBand`: **24 passed, 62 skipped** - (the selector also discovers opt-in platform contracts). -- **Passed:** complete driver suite with `FURN_NATIVE_DRIVER_TEST=1` and - stable signing, **75 suites / 1,240 tests passed; 6 suites / 47 tests - skipped**, including the macOS native build, framing, authority, and - incarnation contracts. The final complete run took **151.668 seconds** - (`native-package-contracts-post-backport-repeat.log`); an earlier complete - run also passed (`native-package-contracts-final.log`). Skipped contracts - are not native qualification. -- **Failed, final native-suite repeat:** **1,238 passed, 2 failed, - 47 skipped** (`native-package-contracts-post-backport.log`). Both failures - are the existing `nativeDriver.test.ts` macOS helper build/sign/reuse - contracts exceeding their unchanged **240-second** test deadlines. Other - contracts passed. No native-source change accompanied this repeat, and - the timeout cause is not established; the earlier passing suite does not - turn this failed repeat into a pass. -- **Passed after timeout isolation:** unchanged - `yarn test nativeDriver.test --runInBand` with native opt-in and signing, - **10 passed, 1 skipped**, in **106.588 seconds** - (`native-package-contracts-focused-after-timeout.log`), followed by the - complete passing suite above. No test deadline, source, or assertion was - relaxed. These passes do not establish the earlier elapsed-time cause. -- **Passed:** runtime `yarn format`, `yarn lint`, `yarn build`, and - `yarn test --runInBand`: **25/25**, four suites. The added LiteUI viewport - contract failed first, alongside two passing Win32 cases - (`liteui-viewport-red.log`), then all three focused cases passed. -- **Passed, Paper:** bundle and native build after the viewport change; - unchanged oversized Input FocusTreatment callback **1/1**, run - `70b6f0cd-b569-4d80-9a11-8495aa9fc68c` - (`paper-viewport-fixed-callbacks-treatment-qualification.json`). -- **Passed, Paper:** real MCP descriptor Checkbox physical click, checked - state, `physical-click: 1`; required-keyboard clear verified empty with - `keyboard: 2` and no semantic-clear receipt - (`paper-native-props-final-qualification.json`, final helper digest above). -- **Passed, Fabric:** the same real-descriptor MCP click, checked state, and - physical receipt, plus required-keyboard clear with verified typed/empty - values and `keyboard: 2`, without semantic fallback - (`fabric-native-props-final-qualification.json`, source `1492afe57`). -- **Passed, Paper:** the original Input callback, without retries or changed - assertions, run `7a7a9962-1633-4060-b4a6-aa5fdaeb5939`, **1/1** - (`paper-document-fixed-callbacks-input-qualification.json`). The earlier - full physical run `99a6f266-3f9d-471e-8ca8-75d879b5205b` failed honestly - with partial effect before Delete; the held-chord experiment also failed. - Neither result is counted as a pass. -- **Passed, Paper:** both unmodified Accordion callbacks through - `yarn storybook test --macos --paper --test '*ACC-001*' --input physical --reporter json`, - **2/2**, run `9cc7f6b4-63e2-484e-98a1-a57bab79174b` - (`paper-native-props-final-callbacks-qualification.json`). The unchanged - oversized Input callback also passed **1/1**, run - `da79154b-fce2-4137-b324-1b870e9133a7` - (`paper-native-props-final-callbacks-treatment-qualification.json`). -- **Passed, Fabric:** both unchanged Accordion callbacks, **2/2**, run - `e2dc76d4-40c4-46ff-a2e1-238777623093`; oversized Input FocusTreatment - callback **1/1**, run `db6b223f-cc45-465e-8952-89383d5c5c2e` - (`fabric-native-props-final-callbacks-qualification.json` and - `fabric-native-props-final-callbacks-treatment-qualification.json`). -- **Passed, Paper accessible smoke:** final helper, **259 passed, 3 declared - unsupported**, run `e638b907-df55-412a-b1ae-fdebd3df522a`, with no failed, - skipped, or quarantined cases. Declared native projection gaps are not - passes. -- **Passed, Paper physical smoke:** isolated full rerun, **343 passed, - 3 declared unsupported**, run `b0ad0949-786d-4550-a620-402a4f4609ba`, - including the original Input, focus-treatment, and Accordion callbacks. - There were no failed, skipped, quarantined, retried, flaky, or incomplete - cases (`paper-final-physical-repeat-lanes-qualification.json`). -- **Failed, Paper physical smoke:** run - `3067945a-9c0f-460a-8895-8c8b0387fc01` stopped during Radio story - navigation, before callbacks: **82 passed, 3 failed, 138 unfinished work - items**. The first context change did not settle; subsequent mutations - correctly refused the quarantined target. A fresh accessible run passed - the same catalog; this does not establish a physical-input root cause or - count as successful physical qualification. The isolated full physical - rerun above passed without changing production code, policy, assertions, - or callback retries; the earlier failure remains recorded. -- An instrumented MCP attempt with pre-click menu/AX probing stopped before - dispatch with `automation-failed`, not an obstruction. Final descriptor - checks omit that extra diagnostic interposition and pass. Instrumentation - causality is not established; the failed attempt remains in - `paper-pointer-fixed-qualification.json`. -- **Failed before the backport, Fabric full smokes:** physical run - `7bfe3fc7-bcbc-4275-ac5b-b457b66af7d7` reached **295 passed, 1 failed, - 3 declared unsupported**; accessible run - `aebe42da-fcf5-4482-8eec-3e28401ea50c` reached **246 passed, 1 failed, - 3 declared unsupported**. Both lost the application during PGB-006's - window resize, and ownership checks refused stale-window evidence. - The same unchanged callback passed **1/1** in isolation, run - `7daeeea1-5042-4dd8-9efa-0b227b33a9a1` - (`fabric-callbacks-progress-current-qualification.json`). The isolated - pass does not qualify either full lane. -- **Failed before the backport, clean Fabric resize:** fresh DerivedData and - a native rebuild retained AddressSanitizer but still lost the application - in the unchanged forced PGB-006 callback, run - `5314fc3d-da45-403e-b9bd-e187f0a9284c`. The later owned-process report - `fabric-progress-asan-report.86816` pinpointed the native-props copy. - Diagnostic stdout/stderr redirection changed the result; those passes - are not evidence of a fix. -- **Passed after the backport, clean Fabric resize:** deleted only this - worktree's Pods and DerivedData, regenerated dependencies, bundled, and - built the native app. The unchanged forced callback passed **1/1**, run - `61d2dce4-5980-49b9-a727-412da47fca71`, with AddressSanitizer still linked. - Native Core's Apple module/Office-style bridging-header probe passed - **3/3** (`native-props-clean-build-green.log`). This probe is not an - Office application build. -- **Passed after the backport, complete Fabric physical smoke:** - `yarn storybook smoke --macos --mode stories-and-tests --input physical --reporter json`, - **343 passed, 3 declared unsupported**, run - `3cb2f6af-69fd-44c1-8fb0-495607e37ca0`. -- **Passed after the backport, complete Fabric accessible smoke:** - `yarn storybook smoke --macos --mode stories-and-tests --input accessible --reporter json`, - **260 passed, 3 declared unsupported**, run - `e8ef9d6a-e3a0-4164-b692-de319aaae545`. Both final Fabric lanes had zero - failed, skipped, quarantined, incomplete, retried, or flaky cases and no - owned process survivors - (`fabric-native-props-final-lanes-qualification.json`). Their ordinary - launch path retained AddressSanitizer; no diagnostic output redirection - or callback assertion change was used to make them pass. -- **Passed after the backport, complete Paper physical smoke:** - `yarn storybook smoke --macos --paper --mode stories-and-tests --input physical --reporter json`, - **343 passed, 3 declared unsupported**, run - `064e399f-2bd0-4913-9996-f2cb3b5aead6`. The application smoke exited - successfully with zero failed, skipped, quarantined, incomplete, retried, - or flaky cases. The local wrapper then failed its additional cleanup - assertion: a substring-only process detector counted the queued - standalone-build `bash` and `lockf` commands because their arguments - mentioned the future native image. Both recorded PIDs are wrappers, - not app/helper/Metro survivors. The original evidence is retained in - `paper-native-props-final-lanes-qualification.json`, with the exact - classification in `paper-native-props-final-cleanup-classification.json`. - The ignored harness now checks executable identity; no product code, - assertion, or smoke result was changed, and no physical input was replayed - to conceal this harness failure. -- **Passed after the backport, complete Paper accessible smoke:** - `yarn storybook smoke --macos --paper --mode stories-and-tests --input accessible --reporter json`, - **259 passed, 3 declared unsupported**, run - `4882b35f-20a4-442a-a097-5fb5b572000e`. This separate complete run used the - corrected harness and had zero failed, skipped, quarantined, incomplete, - retried, or flaky cases and no owned survivors - (`paper-native-props-accessible-final-lanes-qualification.json`). - -### Hosted policy - -The new advisory `macos-storybook-physical` Fabric job has no disabled input -features, runs the explicit physical lane, and uploads its own artifacts. -It is not in the aggregate `pr` job's `needs`. Existing Fabric, Paper, and -legacy macOS compatibility jobs retain their `physicalClick` mask: fixing -external-hit identity and passing managed local input do not qualify hosted -foreground behavior or remove genuine clipped/covered controls. - -### Phase-1 repository checks (plain gate not accepted) - -- **Failed, initial aggregate:** the required locked - `yarn clean-lage-cache && yarn lage buildci` first stopped on this ledger's - new sibling-package Markdown link. It was corrected to the - package-qualified code span required by R-PKGDOC-005, and - `yarn check-agent-instructions` passed. -- **Failed, ordinary parallel aggregate retries:** the unchanged - `boundedCommand.test.ts` escalation fixture received `SIGTERM` rather - than its expected `SIGKILL`. Its child installs the ignored-signal - handler after Node startup, with no readiness handshake before the - 300 ms deadline; a startup/deadline race under aggregate load is inferred, - not directly timed. The focused serial check passed **7/7**, and the - complete opt-in driver suite above also passed unchanged. No source, - timeout, or assertion in this area was modified - (`buildci-final-retry.log`, `buildci-final-complete.log`, and - `bounded-command-focused-final.log`). -- **Passed, serial diagnostic only:** `TF_BUILD=1 yarn lage buildci`, using - the repository's existing serial-Jest mode in `scripts/src/tasks/jest.ts`, - completed **18 tasks with 160 cached/skipped tasks**, zero pending or - aborted tasks, and exit code 0 (`buildci-final-serial.log`). There was - no second cache clear. Serial execution does not change the selected - assertions or establish the earlier timing cause. This is not acceptance - of the PR's plain, parallel `yarn lage buildci` gate; that gate failed twice - locally as recorded above. -- **Passed:** changed Markdown link checks for all **11 files**, root - `yarn oxfmt --check`, `yarn check-agent-instructions`, and - `git diff --check`, plus `yarn change:check`. All source changesets were - committed before the changeset check. -- **Passed, cleanup:** the locked exact-worktree argv and Node CWD scan at - **18:14:43 UTC** found no E-owned app, Metro, helper, compiler, or fixture - survivors (`process-cleanup-final.json`). This includes the standalone - backport child worktree, excludes other owners, and terminates none of - their processes. The audit command completed and released its lock. - Diagnostic launch wrappers and copied dependency/backport scratch were - removed; qualification evidence and both Git worktrees remain. - -### Local disposition and remaining limits - -F15's three input findings are resolved locally on this branch at usable -window geometry: both renderers have real-descriptor physical receipts, -verified keyboard clear, and passing unchanged Accordion callbacks. -Genuine chrome overlap in a 480 × 302 window still refuses input safely; -this is not a promise to click covered content. The external-occlusion -fixture did not establish its foreground precondition, so that native -scenario remains unqualified. - -Windows and Win32 native hosts were not run locally. Windows' shared -non-scrolling layout has been restored to the baseline; the new scrollable -entrypoint is macOS-only, and Win32's existing viewport is unchanged. -Hosted macOS foreground behavior, macOS FluentTester, and other native -application hosts were not run locally. Native Core's module/bridging-header probes -are compile checks, not Office application qualification. The separate -private evidence at `artifacts/pr-evidence/rnmacos-backport.md` records the -immutable install and fresh native build of the bottom branch without E's -driver/runtime sources; session evidence is not shipped in Native Core's -published reference. - -### Review remediation and requalification - -The review fixes are committed separately: `5bdfc133f` keeps Native Core's -published backport reference durable, `2a166010e` scopes the viewport to the -macOS platform entrypoint, and `f34c0686a` restores same-process occlusion -refusal. The lower branch was merged normally, without rewriting the original -commits. The shared Windows decorator is byte-identical to `505a2187b`; -its regression preserves the existing non-scroll parent around story scrollers. -macOS and Win32 retain their scrollable viewport tests. Windows and Win32 -native hosts remain **not run**, not newly qualified by these React tests. - -The new native regressions failed **461/463** before the owner-scope fix: -same-process and unproven hit owners incorrectly permitted activation. -They now retain the target's no-activation scope, while the actual foreign-PID -case retains a distinct identity. The agent regression verifies `obscured`, -rejected mutation, effect `none`, no activation, and no pointer dispatch. -The Windows layout regression failed on the unwanted extra scroller before -the platform split. No story callback, assertion, or timeout was weakened. -Red/green and package-check evidence is under -`artifacts/desktop/f16-macos-physical/review-r2/`. - -All app runs below use production source `f34c0686a` and the newly provisioned -signed helper source digest -`dd3f28bb27bea9349bbf6e26955b8fdaf34c88eeb37b6baa7e0e1fd99552b93b`. -`desktop-driver doctor --platform macos --permissions` passed with Accessibility, -PostEvent, and Screen Recording preflight true and no prompt requested. -GUI resolution was prebuilt-only and physical runs had no disabled input -features. Native and GUI commands ran under the shared lock. - -| Check | Result | -| ----------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------- | -| Strict release Swift build with `-Xswiftc -warnings-as-errors`, external scratch root; helper `--self-test` | Passed; **463/463** native checks | -| Runtime `yarn format`, `yarn lint`, `yarn build`, `yarn test --runInBand` | Passed; **26/26** tests | -| Driver `yarn format`, `yarn lint`, `yarn build`, focused `yarn test src/operations/act.test.ts --runInBand` | Passed; **10/10** agent tests | -| Driver `FURN_NATIVE_DRIVER_TEST=1 yarn test --runInBand`, stable signing, default TMPDIR | Passed; **75 suites / 1,241 tests**, **6 suites / 47 tests skipped** for platform boundaries | -| Real-descriptor MCP Checkbox click on Fabric and Paper | Passed; checked state verified and **one physical-click receipt** each | -| Required keyboard clear on Fabric and Paper | Passed; typed value verified, then empty verified; keyboard receipts, no semantic clear | -| Unchanged Accordion callbacks on Fabric and Paper | Passed; **2/2** each | - -The standalone Swift helper has no CocoaPods-module or Office bridging-header -integration; those checks are **not applicable** to this helper-only revision. -The renderer backport and its previously qualified native-host boundary are -unchanged. - -From `apps/storybook`, each renderer passed bundle, renderer-specific prep and -native build, then both complete smoke lanes: - -| Renderer | Command | Result | Run | -| -------- | -------------------------------------------------------------------------------------------------- | -------------------------------------- | -------------------------------------- | -| Fabric | `yarn storybook smoke --macos --mode stories-and-tests --input physical --reporter json` | **343 passed, 3 declared unsupported** | `2a9eb215-a86f-4f8d-ad18-df6ac473d738` | -| Fabric | `yarn storybook smoke --macos --mode stories-and-tests --input accessible --reporter json` | **260 passed, 3 declared unsupported** | `82a66e7e-6bef-4f28-a9b2-fa95468564f2` | -| Paper | `yarn storybook smoke --macos --paper --mode stories-and-tests --input physical --reporter json` | **343 passed, 3 declared unsupported** | `34bb7859-f333-47a6-a88e-9208e43a4b85` | -| Paper | `yarn storybook smoke --macos --paper --mode stories-and-tests --input accessible --reporter json` | **259 passed, 3 declared unsupported** | `87bc44d4-a335-49bc-ba2d-6dde7a79bcc0` | - -All four lanes had zero failed, skipped, quarantined, incomplete, retried, or -flaky cases. Their canonical reports remain under -`apps/storybook/artifacts/desktop//`. The wrapper reports -`fabric-review-r2-lanes-qualification.json` and -`paper-review-r2-lanes-qualification.json`, alongside the real-descriptor and -Accordion qualification reports, record source identity and empty owned -survivor lists between runs. Existing hosted compatibility masks remain -unchanged; these local results do not qualify hosted foreground behavior. - -### Review plain-gate comparison and cleanup - -- **Failed, required plain gate:** the locked - `yarn clean-lage-cache && yarn lage buildci`, with `TF_BUILD` unset, - completed **171 successful tasks and one failed task** in **34.84 s**. - The unchanged `boundedCommand.test.ts` escalation fixture expected - `SIGKILL` but received `SIGTERM` (`review-r2/buildci-plain.log`). - The serial diagnostic above is not a substitute for this failed gate. -- **Passed, comparison setup:** a clean, detached worktree at the fetched - `origin/main`, `505a2187b0a2451aed7081c9260b343d55742f93`, - passed `yarn install --immutable` and `yarn build`. Installation followed - an initial missing-dependency failure; that attempt is retained rather - than counted as validation. -- **Failed, first untouched-main plain gate:** the same locked - cache-cleared command completed **110 successful tasks and one failed - task** in **41.65 s**. Its `wdioRunner.test.ts` worker exceeded the - **7,000 ms** fixture deadline; the Driver test task, including - `boundedCommand.test.ts`, passed. This first run did not reproduce the - escalation failure (`review-r2/main-comparison.log`). -- **Failed, warm-transform untouched-main plain gate:** one further - cache-cleared plain run, retaining the Jest transform cache to match E's - prior package-test state, completed **138 successful tasks and one - failed task** in **38.11 s**. It reproduced the same escalation fixture's - **expected `SIGKILL`, received `SIGTERM`** failure on unmodified main - (`review-r2/main-comparison-warm.log`). Both compared files were - byte-identical between main and E: implementation SHA-256 - `5066540d6aa245eeeb089f609b3cb0b6b342c836583d594d4d68674bb0b44e43`, - fixture SHA-256 - `a26cfbecbe6ec56a2b5b53ec4da26f3bff30de80b6e35aad597672514329fdd8`. - This establishes that the failure also occurs on the baseline, not its - elapsed-time cause; the startup/deadline race remains inferred. -- Both comparison gates used **Node v24.18.0, Yarn 4.18.0**, default macOS - TMPDIR, `TF_BUILD` and native-test overrides unset, and no serial or - worker-count override. No bounded-command source, fixture, assertion, - or deadline was changed. The plain gate remains **failed**, not accepted. -- **Passed, cleanup:** the locked exact-worktree argv/CWD audit at - **19:51:34 UTC** found no E-owned app, Metro, helper, compiler, or fixture - survivors, including the recorded failed WDIO fixture root. No live - candidate had an unreadable CWD, and no native helper remained. - No process was terminated. The clean, owned detached main worktree was - removed normally at **19:53:09 UTC**; the lower backport worktree and all - comparison logs were retained. Both commands completed and released their - locks. Evidence is in - `review-r2/process-audit-before-main-removal.json` and - `review-r2/main-comparison-summary.json`, beneath the review artifact root. - -## 2026-11-12 hosted physical-window follow-up for #4500 - -Hosted job `131453025148` in workflow run `43295615562`, at `867119f1d`, -failed both unchanged Accordion keyboard callbacks. The retained target tree -and screenshot show a 480 × 302 window with the preview under Storybook's -Actions chrome. The driver correctly refused `obscured` before interaction; -this is not an occlusion-detection defect. The previously persisted local -932 × 782 frame concealed the hosted geometry. - -The harness-only policy in `d949e3c20`, refined by `cc96a3a9d`, establishes -and freshly verifies a **1024 × 768 logical-point** window for macOS -`--input physical` traversal and standalone attached callbacks. It uses the -existing window-rectangle API, queries AppKit's visible screen frames without -permission prompts, chooses a fitting screen, and centers the window with -its integral origin clamped inside that frame. No fitting display or an -unconfirmed rectangle fails before callbacks. There is one resize per -preparation, not a resize retry. Accessible, compatibility, Windows and Win32 -lanes do not query or request this geometry. Native occlusion checks, input -mechanisms, runtime layout, story assertions, and test deadlines are unchanged. - -### Regression and native evidence - -- **Observed baseline:** forcing the owned Fabric window to 480 × 302 and - using compatibility `--click-mode physical` reproduced both Actions - obstructions, run `f4e7461b-9079-4984-b4f3-ea2591c4d84d`. - The application subsequently grew the frame to 515 × 341; the original - wrapper's additional exact-480 assertion failed and is retained, not - credited as a passing qualification. The two callback failures themselves - match the hosted obstruction. -- **Observed edge-position refusal:** the first origin-preserving clamp - requested 1024 × 768 at the visible screen's lower boundary, but the - host returned 1024 × 767. Preparation correctly failed before selecting - tests, run `b4c5479a-4c71-457f-9c6b-0ecae93a019f`. This motivated centering - away from native screen-edge constraints, not accepting a smaller window. -- **Passed, test-first package checks:** the initial missing-policy - regressions failed seven cases; the centering regression then failed - three before its fix. `yarn test storyTraversal DesktopStorybookCli` - passed **86/86**; package `yarn format`, `yarn lint`, `yarn build`, and - ordinary parallel `yarn test` passed, with **353 passed / 4 skipped**. - Cases cover physical-only preparation before traversal/callbacks, - negative-origin and fitting secondary displays, off-screen persisted - frames, undersized displays, rejected geometry, and owned-session cleanup. -- **Passed, signed native preflight:** Accessibility and PostEvent were - true; no prompts or privacy changes were requested. The production - screen query returned a visible logical frame of - `{x:0,y:33,width:1728,height:999}` during preflight. -- **Passed, unchanged helper at `d949e3c20`:** fresh - `swift build -c release -Xswiftc -warnings-as-errors` with external scratch - and `--self-test` passed **463/463**; the signed Driver - `FURN_NATIVE_DRIVER_TEST=1 yarn test --runInBand` passed **75 suites / - 1,241 tests**, with **6 suites / 47 tests skipped** for platform boundaries. - The native source digest remains - `dd3f28bb27bea9349bbf6e26955b8fdaf34c88eeb37b6baa7e0e1fd99552b93b` - at `cc96a3a9d`; the later change is confined to the Node harness, tests, - and documentation. Standalone-helper module/bridging-header checks are - not applicable. - -Detailed attempts, red/green package logs, preflight, and native checks are -retained under -`artifacts/desktop/f16-macos-physical/hosted-4500-20261112T214852Z/`. -No missing-lease harness attempt or malformed diagnostic command is credited -as product validation. - -### Current-head real-app qualification and remaining limits - -The GUI checks below use production Node source -`cc96a3a9dade18f373546f2a72edbf20a41e5461`, the same signed helper, and -prebuilt-only helper resolution with no disabled input features. All native, -GUI, physical-input and aggregate commands used the shared native lock. - -| Check | Result | Run | -| ---------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------- | -------------------------------------- | -| Fabric bundle, renderer prep and native build | Passed; consuming native/runtime sources unchanged by the later Node-only centering commit | `da613def-aace-4786-a5ad-396abcd06110` | -| Fabric unchanged Accordion callbacks, forced off-screen 480 × 302 window | Passed, **2/2**; final window **1024 × 768**, completely inside the visible display | `6ef69ad4-8207-4ed1-97e6-a05be4ba7d9e` | -| `yarn storybook smoke --macos --mode stories-and-tests --input physical --reporter json` | Passed, **343 passed / 3 declared unsupported** | `cfdeb316-5a15-45ca-9b8b-b034ce312142` | -| Paper bundle, renderer prep and native build | Passed | `a915fd88-3518-4266-94a1-6f78bb66a6cc` | -| Paper forced-small-window callback preparation | Failed before resize or callback selection: **screen locked**; owned app, helper and Metro cleaned | `716bf45c-c7a0-47e1-891c-4a62d4510792` | -| Paper complete physical smoke | **Not run:** locked console; do not attribute the earlier `f34c0686a` Paper passes to this Node-policy revision | None | -| Accessible lanes, Windows and Win32 native hosts | **Not run:** physical-only macOS harness change; unit regressions retain their no-query/no-resize behavior | None | - -The passing forced-small Fabric run read back -`{x:352,y:150,width:1024,height:768}` inside the freshly queried visible -`{x:0,y:33,width:1728,height:1001}` frame. Both passing Fabric runs had zero -failed, skipped, quarantined, incomplete, retried or flaky cases and empty -owned-survivor lists. Canonical run records remain under -`apps/storybook/artifacts/desktop//`; the adjacent qualification -wrappers preserve source identity, frame before/after, and cleanup. - -Paper's attempt at **23:00 UTC** could not attach to its owned application. -Nonprompting doctors at **23:06** and **23:11 UTC** confirmed -`screenLocked:true` while Accessibility and PostEvent remained `true`. -No input was sent, no unlock or privacy change was attempted, and no -current-head Paper GUI success is claimed. Its fresh bundle/prep/native -build succeeded independently of this console limitation. - -### Follow-up repository gate and cleanup - -- **Failed, required plain gate:** one locked - `yarn clean-lage-cache && yarn lage buildci`, with `TF_BUILD` and native - test overrides unset, failed the unchanged - `packages/agentic/storybook-desktop/src/testing/wdioRunner.test.ts` - fixture's **7,000 ms worker deadline**. - The Storybook test task reported **352 passed / 1 failed / 4 skipped**. - No timeout, fixture, runner, or assertion was changed. The retained - untouched-main cold comparison in the preceding section already shows - this same worker-deadline failure; no new main comparison or serial - acceptance is claimed for this follow-up. -- **Passed, focused diagnostic:** package `yarn test wdioRunner` passed - **1/1**, in addition to the complete **353 passed / 4 skipped** package - run before native qualification. This does not accept the failed - aggregate gate. Logs are `plain-buildci.log` and - `wdio-runner-focused-after-gate.log` beneath the follow-up evidence root. -- **Passed, cleanup:** after aggregate cancellation, **46** still-draining - owned test-process IDs were recorded; all settled naturally before the - bounded follow-up check, without termination. The final locked, - exact-worktree argv/CWD audit at **23:11:33 UTC** found no owned app, - Metro, helper, compiler, test worker or fixture survivors and no - unattributed live candidate with an unreadable CWD - (`final-native-process-audit.json`). The audit command completed and - released its lock; no additional GUI runs were attempted while locked. -- **Passed:** changed Markdown link checks, root `yarn oxfmt --check`, - `yarn check-agent-instructions`, `git diff --check`, and - `yarn change:check` after committing the changeset. - -### Environment-selected physical smoke review follow-up - -A focused review found that traversal received the raw optional smoke input, -while the manifest and callback paths already resolved `DESKTOP_INPUT`. -Consequently, `DESKTOP_INPUT=physical` without `--input` omitted window -preparation. Source `eaab0b6781332ee8ef5054e54787033c6817d96e` now passes -the existing `resolveWdioOptions` merge's effective input to traversal. -The same resolver preserves explicit-flag precedence and input/click-mode -validation. No geometry policy, native implementation, callback assertion, -or accessible/compatibility/Windows sizing rule changed. - -- **Passed, test-first:** the new environment-only lifecycle case failed - before the fix (**1 failed / 76 passed**). The focused command - `yarn test src/cli/DesktopStorybookCli.test.ts src/cli/storyTraversal.test.ts src/config/wdio.test.ts src/cli/inputLane.test.ts` - then passed **114/114**, including environment-only physical selection, - explicit accessible precedence, existing validation, and no-resize lanes. -- **Passed, package:** `yarn format`, `yarn lint`, `yarn build`, and ordinary - parallel `yarn test`; **23 suites / 355 passed / 4 skipped**. - The existing patch changeset now names environment-selected smoke input; - committed-changeset `yarn change:check` passed. -- **Failed, single required plain aggregate:** locked - `yarn clean-lage-cache && yarn lage buildci`, with serial/native/input - overrides unset, reproduced the unchanged - `src/native/boundedCommand.test.ts` escalation fixture's **expected - `SIGKILL`, received `SIGTERM`** failure. The earlier baseline comparison - remains recorded above. No fixture, deadline, serial override, aggregate - retry, or second cache clear was used to conceal this known failure. -- **Passed, cleanup:** all **50** still-settling aggregate processes drained - naturally under the same lock. The exact-worktree argv/CWD audit at - **23:30:20 UTC** found no owned survivors or unreadable live candidates; - cleanup issued no termination and the command released its lock. - -This environment-only fix is **portable-qualified, not newly native-qualified**. -The explicit-`--input physical` Fabric evidence above belongs to -`cc96a3a9dade18f373546f2a72edbf20a41e5461`, not this revision. The console -remained declared locked; no GUI/physical run, privacy change, prompt, or -unlock attempt was made. Detailed red/green, package, aggregate and cleanup -records are in -`artifacts/desktop/f16-macos-physical/review-env-input-20261112T232312Z/`. - -### Hosted visible-frame clamp and native requalification - -Hosted macOS physical job `131573011450` on `b66140d7b7b04cb33d8b6910c6f41a16651af713` failed before selecting stories because no visible screen fit the 1024 × 768 target. On November 13, the user chose to clamp the physical-lane window size to the available visible frame. The follow-up keeps the 1024 × 768 maximum, adds a one-point inset on each edge, chooses the largest clamped area (overlap breaks ties), verifies the exact requested rectangle, and includes target/requested/actual/visible geometry in failures. Accessible, compatibility, Windows and Win32 lanes remain unchanged. - -- **Passed, native Fabric at `b66140d7b`:** `DESKTOP_INPUT=physical yarn storybook smoke --macos --mode stories-and-tests --reporter json`; **343 passed, 3 declared unsupported, 0 failed**. The signed helper source digest was `dd3f28bb27bea9349bbf6e26955b8fdaf34c88eeb37b6baa7e0e1fd99552b93b`; the fresh app build included the RNmacOS native-props patch. The visible frame was 1728 × 999, so the 1024 × 768 target was unchanged. -- **Paper native retry at the same head:** the first full env-only physical smoke passed 254, failed 7 and reported 1 unsupported; six failures were obscured by the unrelated Captive Network Assistant window, and one reported an unsupported native hit-test operation. No process was signalled. After that OS window disappeared, the repeated full smoke passed **343, 3 declared unsupported, 0 failed**. -- **Passed, current clamp working tree on Paper:** env-only physical Accordion smoke, **225/225**, no unsupported or failed cases. Accessible Button callbacks on the prior `b661` source passed **1/1 on both Fabric and Paper** without a physical window preset; existing unit tests assert accessible/compatibility/Windows traversal does not resize. -- **Not completed, current clamp working tree on Fabric:** `bundle --macos` passed, but the subsequent `prep --macos` stopped in `pod install` with `undefined method 'fast_generate' for module JSON`; no build or smoke followed. Native source and app configuration were not changed by the clamp. -- The hosted failure is the reason for the current clamp change, not a story or hit-test failure. The post-clamp hosted run is pending publication and CI. - -Detailed native receipts are in -`artifacts/desktop/f16-macos-physical/native-b66140d7-20261113T000145Z/`, -including the initial Paper failure, the passing rerun, current Paper smoke, -and failed Fabric preparation. - -## 2026-11-12: bounded refresh timing observations - -Part of #4435. Story selection can return allowlisted, service-local timing -observations and a canonical native-verified PID/start-time pair. They are -response diagnostics, not authenticated marker fields, context identity, or -authority to clear quarantine. Request/run/epoch matching and fresh native -marker verification remain unchanged; provider payloads and private values -are not copied into these observations. - -The instrumented baseline ran eight independently launched accessible cohorts -per macOS renderer, each with three passing ListboxItem callbacks and one -declared macOS-unsupported callback. Each passing callback selected a fresh -preview/context generation; keyed remounts are inferred from rendering -policy, not independently counted. Every cohort retained one verified app -incarnation and one authenticated bridge connection during its callbacks. -JS initialization/HMR was not observed independently and cannot be inferred -from hello counts. Fabric accepted one hello and Paper two on the same socket. -No activation command or main/focused-window change was observed in these -baseline cohorts; Paper recorded one application deactivation. Historical -per-case truncation counts are withdrawn: a process-lifetime history flag -and globally allocated sequence gaps did not establish per-case event loss. -A zero advisory count -is not visual qualification. -The Storybook evidence ledger, -`packages/agentic/storybook-desktop/references/current-status.md`, records -selection, timing, variability, and limitations. - -This timing implementation changes no Swift source, physical-input -activation, foreground checks, hit-testing, focus, or window-switch policy. -It does not qualify those paths or remove provider-local checks. Windows -Fabric and Win32 Paper were **not run locally** for this change. - -The direct-selection Fabric candidate preserved identical coverage, fresh -per-case generations, one app incarnation/connection per lifecycle, and no -within-lifecycle app replacements across eight paired-by-index repetitions. -Runner p50/p95 changed from 7939.07/8845.37 to 6680.04/6995.31 ms; navigation -p50 changed from 753.13 to 274.20 ms. Paper's equivalent eight-plus-eight -comparison changed runner p50/p95 from 7915.99/8755.09 to 7218.78/7811.59 ms -and navigation p50 from 756.77 to 340.59 ms, with identical coverage and -lifecycle invariants. Each side/renderer passed 24 callbacks plus eight -declared unsupported. Sequential before/after blocks, cold startup -differences, and the small observational cohort limit attribution; -native-verification p95 worsened on both renderers (Fabric 168.00→546.11 ms; -Paper 170.07→582.90 ms). There is no universal tail-latency, reliability, -startup, or flashing claim. - -Window-scoped synchronized recordings on both sides/renderers contained four -encoded frames each, followed by held pixels. The unvalidated whole-content -uniformity detector returned zero candidates; persistent chrome and -window-only capture prevent interpreting this as zero blank-preview or -activation-flash episodes. No flash reduction was established. - -Focused candidate qualification on both macOS renderers passed the complete -223-story render traversal plus five accessible callbacks (228 passes per -renderer), and a separate sparse physical selection with 15 callbacks -(238 passes per renderer). Button, Dropdown, native Callout, and modern -Callout paths retained fresh per-case contexts/workers and passed popup -isolation/dismiss/reopen checks. These are not full authored-input suites, -MCP descriptor qualification, native cancellation/fault injection, or -qualification of another branch's physical-input corrections. - -The unchanged standalone macOS helper passed a release build with -`-Xswiftc -warnings-as-errors` and all 448 self-test checks. Focused -`FURN_NATIVE_DRIVER_TEST=1` contracts passed 87 tests, with 16 skipped -(Windows contracts not run locally). An earlier local aggregate passed with `TF_BUILD=1` serializing package Jest -runs and with project-local fixture overrides; this is not PR CI mode. -The PR workflow uses parallel `yarn lage buildci` without those overrides. -The Storybook evidence ledger records exact commands, failed attempts, -package counts, and the subsequent plain-command review qualification. - -### Activation trace and deferred flashing change - -The W3C click path activates before its server-side hit test. The operation -click path conditionally raises an externally obscured target and rechecks; -in-window obstruction remains a no-effect refusal. The native physical-click -provider independently activates under its physical-input scope, rereads the -target, and verifies ownership of the click point. These are separate -postconditions, not interchangeable duplicate calls. Removing the server -activation can move hit-testing before foreground establishment; removing the -provider activation assumes ownership cannot change between commands and -weakens callers that use the provider independently. - -macOS activation sets main/raise attributes even when already frontmost. -The menu test helper `inWindowWith` also calls activating `switchToWindow` -while polling each candidate window and restoring its owner. Both are -plausible flashing contributors, not established visual causes in the -observational ListboxItem cohort. - -No activation optimization is shipped here. Concurrent physical-input -corrections retain foreground/main-window postconditions, add a workspace -fallback when acknowledged AX activation is ineffective, and keep genuine -occlusion refused. An unqualified dedupe must not fork those semantics. -Zero activation counters in this read-only cohort do not meet the optimization -gate. The compatible experiment and its qualification conditions are tracked -in [remaining work](remaining-work.md#macos-activation-and-popup-polling). - -## 2026-11-12: refresh-diagnostic review corrections - -Part of #4435. The Storybook report now separates the native process-lifetime -`eventsTruncated` flag into `processEventsTruncated` from a possible -per-case `eventsTruncated` warning. A case warning requires truncated -history whose earliest retained event is newer than the case cursor. -Globally allocated sequence gaps are not loss evidence; the unsupported -command-truncation inference was removed. Historical 16/17-of-24 -comparisons are withdrawn because the original bounded reports cannot -reconstruct per-case loss. The duplicate mount-generation field was also -removed: fresh preview generations infer remounting but do not count mounts -independently. No native producer or activation behavior changed. - -The earlier `TF_BUILD=1` aggregate was local serialization, not PR CI mode. -Under the shared native lock, plain -`yarn clean-lage-cache && yarn lage buildci` with no environment overrides -**failed in both candidate runs** on the `boundedCommand` termination-signal -assertion; `privateStorage` passed in both. The identical -command on independently installed, detached base `aadc2f359` **failed** -that same assertion and a `privateStorage` timeout. The base worktree was -removed after terminating its verified orphaned Jest worker. Those -pre-existing tests and implementations were not changed. - -Review regressions **passed 112 tests** across four Storybook suites. -macOS Paper manifest/bundle and full-index accessible smoke **passed** -(223 rendered stories plus five callbacks, three extractions/two cache -hits). A harness PATH failure before selection was corrected by invoking -it through Yarn, without a product change. macOS Fabric and physical -lanes were **not rerun for this follow-up**; Windows Fabric and Win32 Paper -were **not run locally**. The earlier paired timings and window recordings -were not repeated, and no new performance or flashing claim is made. - -## 2026-11-12: macOS exact application identity backend - -Source `76a2be59abfde3a7bcb0abd8d86a2052e7378274`, macOS 27.2.1 arm64, -standalone AppKit fixture; no React renderer. Exact kernel start timestamps, -causal direct-child/LaunchServices capture, exact-primary cleanup and stable -per-user exact-app exclusion are implemented. Both macOS launch kinds remain -uncontained. Unknown or stale identity never authorizes termination. -The generic lease/claim policy lives in `desktop-test`, not in the helper. - -Under the exclusive native-validation lock, with `FURN_NATIVE_DRIVER_TEST=1`, -stable signing identity `FURN Desktop Driver Development`, project-local -temporary storage and an isolated helper cache: - -- At source `fc1df4bda42004ad86378d2acc6701427dda1836`, - `yarn workspace @fluentui-react-native/desktop-driver test --runInBand` - passed 1,241, failed 0, skipped 47; 75 suites passed, 6 skipped. - Artifact: `artifacts/desktop/f10m/driver-native-stable.log`. - This includes signed build/rebuild and causal-identity native self-tests. - An earlier build-ID mismatch was caused by editing native inputs while the - run was queued; this successful rerun used frozen source. -- At source `76a2be59abfde3a7bcb0abd8d86a2052e7378274`, - `yarn workspace @fluentui-react-native/desktop-driver test --runInBand applicationIncarnation serviceIdentity` - passed 69, failed 0, skipped 14; 4 suites passed, 2 skipped. - Artifact: `artifacts/desktop/f10m/native-identity-final.log`. - The intervening fixture deployment-target/schema-bound correction was - separately exercised by the harness native safety cases. -- The harness's native OS-owner and rejected-launch cases passed 3 with - 5 skipped. Actual AX calls returned `apiDisabled` (-25211), but exact - direct-child and LaunchServices primary cleanup and claim-resource records - were proved. Both recorded PIDs and all own-worktree fixture/helper - processes were absent afterward. The corrected fixture's Mach-O minimum - is macOS 11.0, matching the helper. - -Exact helper artifact/digest, commands, PIDs, failures, cleanup and evidence -locators are in `packages/agentic/desktop-test/references/current-status.md`, -section "2026-11-12: F10-M neutral application ownership". -Positive application lifecycle/external preservation and cross-service -application exclusion remain blocked by missing Accessibility permission; -the physical-input mutex is not qualified, and PostEvent preflight is false. -macOS Fabric and Paper signing/TCC-responsibility qualification remain -separate, unrun gates on this branch. No TCC/privacy settings were changed. -Primary exit is not descendant containment; Windows/Win32 native application -claim/survivor qualification remains not run for this change. - -### 2026-11-12 review follow-up - -At source `62e6552028ea8fb81e76d430379321145fab709c`, the signed helper -was forcibly rebuilt and -`yarn workspace @fluentui-react-native/desktop-driver test --runInBand nativeDriver.test applicationIncarnation serviceIdentity` -passed 79, failed 0, skipped 15; 5 suites passed, 2 skipped. The lock, signing, -native opt-in and project-local storage policy above were retained. -The native self-test exercises actual flock contention, preserves a real child -after injected EWOULDBLOCK and proves exact cleanup of a real launched child -after injected EBADF. Pre-spawn failures report `not-started`; a lost dispatched -launch response remains `unknown`. Default-TMPDIR transport regressions passed -59/59. - -The matching AppKit safety subset passed 4, failed 0, skipped 5, including -pre-spawn/LaunchServices no-app rejection and claim reuse. Its helper artifact is -`1d68fae8df504432900e30e5b2a6714029762f5dd690cd953a8ef07b1f5da4f3`, -source digest -`2c8ac56b2b824b05b1cb10d358050dfb643b44eb6e209508d53eaa0f2744b0f8`. -Raw evidence is -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-JMEGuu/evidence.json`; -commands and exact primary identities are in the harness ledger. -Positive AX and renderer qualification remain blocked/not run as above. - -The final review-fix `yarn clean-lage-cache && yarn lage buildci --concurrency 4` -passed under the exclusive lock at clean source -`e74043b0e6eb5b452e9023bc60c70ba6c9540600`: 178 successful tasks, -0 skipped/pending/aborted/failed. Project-local TMPDIR/Git-ceiling isolation and -the declared `TF_BUILD=1` in-band test task were used, with native-only -signing/cache/opt-in overrides removed. Log: -`artifacts/desktop/f10m/review-buildci-final.log`. -The harness ledger records the exact invocation. This aggregate is portable -repository validation, not additional positive native qualification. - -### 2026-11-12 Storybook identity-observation regression - -CI run `43148275374` at `0bd0ad6f7` failed both macOS Storybook renderers: -the normal driver session held the exact-application flock, while bridge-epoch -recovery created another helper and tried a competing controlling `attach`. -The passing macOS E2E artifact has 62 tests using session-owned `launch`; -it does not exercise that Storybook identity callback. - -Fix source `5e360734e9b85bd4732aabc47c7654bb7211c3d3` adds -`NativeDesktopHost.inspectApplication` and the matching macOS native command. -It shares descriptor/legacy-lease/window validation and checks the exact -kernel incarnation before and after discovery, without creating a lease, -acquiring application/input locks or registering windows/elements. Storybook -uses the session's helper for this observation. Real launch/attach exclusion -is unchanged; foreign control remains `application-lock-busy`. -Windows/Win32 retain their existing temporary identity attachment, and an old -macOS helper never falls back to controlling attachment. - -Default-TMPDIR adapter regressions passed 11; composed Fabric/Paper Storybook -regressions passed 5; claim/ownership regressions passed 29. The full portable -driver suite passed 1,236 with 66 skipped, and all three owning packages passed -format/build/lint and their full suites. Broad fixtures require project-local -TMPDIR under this task's filesystem restriction; only the focused selections -are default-TMPDIR evidence. - -Under the exclusive native-validation lock, the stable-signed helper was -forcibly rebuilt. Native identity/termination/backend contracts passed 79 with -15 skipped, including new inspection-refusal self-tests; AppKit safety passed -4 with 5 skipped. Helper artifact: -`982f3cfa696bbde7f9868186ae9365d59014fd656687e5cbec2de74863280ed3`, -source digest -`6bafaf5835d9cd22410992466a8d8f86086ff6b3b3efd182b3a48922f698527c`. -The noninteractive doctor still reports Accessibility/PostEvent preflight false. -Both rejected-launch primaries and all own-worktree fixture/helper processes -were absent after the run. No privacy changes were made. - -The harness ledger records exact commands, identities and evidence locators. -Raw evidence is under `artifacts/desktop/f10m/ci-regression/`. -Positive AX-dependent ownership/observation contracts and actual macOS -Storybook Fabric/Paper CI requalification remain pending, not passed; the -other native acceptance gates above are unchanged. - -The final clean-cache aggregate at `a7c37b86a` passed **178 tasks**, with 0 -skipped, failed, pending or aborted, under the exclusive native-validation -lock: `yarn clean-lage-cache && yarn lage buildci --concurrency 1`. -It used `TF_BUILD=1`, project-local TMPDIR and the matching Git ceiling, with -native-only overrides removed. The harness ledger records the full invocation. -The first concurrency-4 attempt failed after 112 successful tasks with abrupt -partial Storybook Jest output and no cause diagnostic. The unchanged isolated -Storybook suite passed 344 with 4 skipped under that exact environment before -the serialized aggregate retry. The first failure's cause is unestablished; -no source or test waiver was applied. Logs are `buildci.log`, -`storybook-buildci-environment.log` and `buildci-final.log` in the same artifact -directory. Runtime source remains `5e360734e`; this portable aggregate does -not qualify the still-pending positive native or renderer gates. - -### 2026-11-12 AX-enabled application-ownership qualification - -**Observed, superseding the permission-blocked results above.** Source is -`9395737943d7e1402a40a67a41dc605ef0afb24b` plus the live-discovery fix in -this delivery. Host: macOS 27.2.1 arm64. The signed, nonprompting doctor -reported Accessibility, PostEvent and ScreenCapture preflight `true`, unlocked -and on-console, with `promptRequested=false`. No privacy settings were changed. -Helper artifact: -`1429ce4663795bad31f545086ef80bffb83e9b9533287964991c5527388b2f6c`; -source digest: -`c17944691081d30a320cf6e4393a5bb0ce6cece8dd71295a400dbc5e6a53b3cb`. -The signer is `FURN Desktop Driver Development`, with designated requirement -`identifier "com.microsoft.fluentui-react-native.desktop-driver" and certificate leaf = H"ab592605f8619fcaa952b4a2525ae6d41296fdd7"`. - -The initial positive AppKit run passed 8 and failed 1: the session's window -existed, but its helper could not rediscover the newly launched app. -The new warmed-discovery regression failed before the fix (1 failed, -9 skipped). A read-only native diagnostic observed the child missing from -`NSWorkspace.shared.runningApplications` while direct PID and static bundle -queries saw it. The standalone helper does not pump the main run loop needed -to refresh that cached workspace list. Descriptor discovery now enumerates -current-user kernel PIDs and constructs fresh `NSRunningApplication` objects, -retaining exact bundle/executable/window, uniqueness and incarnation checks. -Enumeration failure refuses discovery; launch causality is unchanged. - -Under `$HOME/Library/Caches/furn-overnight/with-native-lock.sh`, -with the default TMPDIR unchanged: - -```sh -env FURN_NATIVE_DRIVER_TEST=1 FURN_NATIVE_PHYSICAL_INPUT_TEST=1 \ - FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ - FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ - yarn workspace @fluentui-react-native/desktop-test test --runInBand \ - applicationOwnership.contract applicationOwnership.test ownership.test \ - --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-native-harness" -``` - -**Passed 39, failed 0, skipped 0** across three files: 10 native ownership -tests and 29 portable regressions. Native cases cover supervisor lifetime -across two sessions, external preservation, independent-service exclusion, -lease release, fresh read-only discovery while another helper controls the app, -duplicate-bundle refusal, stale identity, causal primary cleanup, cancellation -and the physical-input mutex. The mutex uses only bounded Shift down/up, not -pointer input. AppKit evidence: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-1V0SrR/evidence.json`. -Raw red/green logs, the diagnostic and verified experimental-survivor cleanup -are in `artifacts/desktop/f10m/ax-qualification-20261112/`. - -The harness ledger's matching dated section owns the renderer commands, full-suite counts, -cleanup and remaining acceptance gates. AppKit evidence does not establish -descendant containment, interrupted native publication, automatic recovery, -or Windows/Win32 packaged-survivor behavior. - -**Observed, real renderers:** Fabric and Paper each passed all 9 qualification -checks, with 0 failed/skipped and 0 cleanup failures. Each run used the same -stable-signed helper above and independently rechecked nonprompting -Accessibility/PostEvent preflight. Neutral external owners preserved the app; -foreign native control returned `application-lock-busy`; a replacement -attached after lease release. Read-only inspection succeeded while the -controlling lease was held and retained its window handles. Supervisor-owned -launch survived two borrowed sessions, published exact application/owner -identity in schema 3 and exited at supervisor disposal. The transitional -Storybook CLI's outer claim remains `unregistered`; registered target/instance -checks apply to the neutral claims, not a fictional completed F9 migration. - -| Renderer | Exact smoke command from `apps/storybook` | Passed / failed / skipped / unsupported | Artifact | -| ------------ | --------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------- | -------------------------------------------------------------------------------- | -| macOS Fabric | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook --reporter json smoke --macos --mode stories-and-tests --input accessible` | 224 / 0 / 0 / 0; retries 0 | `apps/storybook/artifacts/desktop/616bf2e8-a675-408f-bd56-2e09982ea969/run.json` | -| macOS Paper | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook --reporter json smoke --macos --paper --mode stories-and-tests --input accessible` | 224 / 0 / 0 / 0; retries 0 | `apps/storybook/artifacts/desktop/19358539-eed2-49e9-9c9a-79b5e7db4e9f/run.json` | - -Each smoke traversed 223 catalog entries and executed the one manifest-selected -Button semantics callback. One physical callback was excluded by input policy, -not passed or skipped. The physical-input mutex was separately qualified by -the AppKit test; these smokes are not pointer/capture qualification. -The exact run-owned and supervisor-owned apps exited, all reported -driver/channel/Metro ports closed and all test claims were released between -runs and at the end. Renderer evidence: -`artifacts/desktop/f10m/ax-qualification-20261112/{fabric,paper}/evidence.json`. -The harness ledger records the declared bundle/prep/run procedure, installed -Ruby JSON selection and failed qualifier attempts with their exact cleanup. - -**Not run/unqualified:** native descendant containment, interrupted native -publication, supervisor-death survivors, Windows/Win32 packaged-survivor -journaling, repeated-run reliability and clean-machine/PPPC/revoked-grant -behavior. Effective TCC acceptance is scoped to the observed `Rex Server.app` -responsibility tree, not an independently durable helper-owned grant. -Keep draft; ordinary hosted Windows driver/smoke CI is not sufficient for -the unrun harness survivor gate. - -Final declared package format checks, TypeScript builds and lint passed for -desktop-driver, desktop-test and storybook-desktop. The full signed, -native-enabled driver suite passed **1,255**, failed **0**, skipped **47**. -The separately scoped portable suites passed driver **1,236** (66 skipped), -harness **409** (11 skipped), and Storybook **344** (4 skipped), with no failures. -All used the default TMPDIR. Skips are not native passes; the signed focused -ownership run separately passed all 39, including all 10 AppKit tests. - -A broad mixed-environment harness attempt failed 3 (416 passed, 1 skipped): -two portable fixtures replace HOME, hiding the signing keychain while -inheriting the native signing override, and one LaunchServices acquisition -reported an unverifiable incarnation and preserved the process. Its cause is -unestablished; this is not reliable-launch qualification. That exact -run-owned fixture was operator-cleaned only after native identity/path/time -checks. The discovery-only regression now uses externally started children -with native identities, independent of LaunchServices ownership acquisition; -no causal guard was relaxed. The signed native selection and the portable -suites then passed separately. The harness ledger records exact commands, -failed-attempt evidence and the remaining causal-acquisition reliability gate. - -**Observed final repository validation:** the single final clean-cache -aggregate, `yarn clean-lage-cache && yarn lage buildci --concurrency 1`, -passed **178 tasks**, failed **0**, skipped **0**, pending **0**, aborted **0** -in 6m 50.47s under the native lock. It used `TF_BUILD=1`, the default TMPDIR, -and explicitly removed the five native/signing overrides listed in the -harness ledger. This portable aggregate supplements, not replaces, the -separate signed native and renderer results. Full invocation and output: -`artifacts/desktop/f10m/ax-qualification-20261112/buildci-final.log`. - -**Observed final cleanup:** after the aggregate, signed native probes -reported all **31** recorded exact process incarnations exited. Scoped -command-line and current-working-directory inspection found **0** worktree -workload processes; all three reported driver/channel/Metro ports refused -connections. The shared workload claim and all `interactive.lock` files in -eight native/renderer/failed-attempt artifact roots were absent. The -locked verifier exited successfully, and a subsequent locked `true` -reacquired and released the machine-wide lock. No privacy changes or -unknown-process termination were used. Evidence and operational verifier: -`artifacts/desktop/f10m/ax-qualification-20261112/cleanup-final.json`, -`cleanup-final.log` and `verify-cleanup.mjs` in that directory. -The experimental diagnostic/cleanup scratch sources were removed; their -JSON results and logs remain ignored evidence. - -Documentation closeout passed all four changed Markdown link checks -(12 / 6 / 15 / 0 links), `yarn lage test-links --verbose --grouped` -(90 tasks, 0 failed/skipped) and repository-wide `yarn oxfmt --check` -(3,159 files). The harness ledger records each exact link-check command. -Final logs in the same artifact root: `changed-markdown-links-final.log`, -`test-links-final.log` and `root-format-final.log`. - -### Rebased macOS publication/survivor contracts — 2026-11-12 - -The macOS ownership layer is rebased onto F6 main -`041ac8b9894aeaa7e1aa273baa2ece4e46298b94`; test/fixture qualification -commit `d5a7530a9f2d4cb78721a36c1109b5c36fbdc5a1` does not change -production helper behavior. Under the shared native lock on macOS 27.2.1 -arm64, the stable-signed native driver suite passed **1,256, failed 0, -skipped 47**: - -```sh -FURN_NATIVE_DRIVER_TEST=1 \ -FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ -FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ -yarn workspace @fluentui-react-native/desktop-driver test --runInBand \ - --cacheDirectory "$PWD/artifacts/desktop/f10m/rebased-qualification-20261112/jest-driver-native" -``` - -Package format, lint and build passed. Nonprompting doctor reported effective -Accessibility/PostEvent/ScreenCapture permissions and an unlocked console; -no privacy changes were made. Helper artifact/source digest are unchanged -from the previous signed qualification. - -The harness's focused suite passed **13 native AppKit cases plus 29 portable -cases, no failures or skips**. New native evidence covers failed acquired-record -publication, owner death before/after acquired-record publication, retained -workload fences, real application/known-descendant survivors and exact operator -cleanup. The descendant remains alive after the primary exits, consistent with -both macOS launch kinds remaining explicitly `uncontained`. This is not -automatic recovery or arbitrary descendant containment. Stale native start -identities and inaccessible foreign process identity still fail closed. - -Exact commands, process identities, limitations and the independent renderer -and aggregate outcomes are recorded in -`packages/agentic/desktop-test/references/current-status.md`, section -"Rebased F10-M publication and survivor qualification". -Evidence root: `artifacts/desktop/f10m/rebased-qualification-20261112/`; -AppKit identities: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-kffNMN/evidence.json`. - -The rebased real Storybook consumer passes at -`ca88c289c0fbd79b03dbd43b9bd85f0b21e7e800` independently passed all -**9 ownership checks on Fabric and all 9 on Paper**, with no cleanup failures. -Each accessible smoke passed **224/224**, with no failures/skips/unsupported -cases/retries; one physical callback was excluded before selection. -Neutral supervisor/session lifetime, exact external preservation, independent -application exclusion, read-only inspection and exact cleanup all passed. -This supplements the unchanged helper's native contracts and does not -qualify arbitrary descendants or deferred Storybook lifecycle integration. -Exact commands and report IDs are in the harness ledger. - -**Aggregate qualification remains failed:** one plain cache-cleared -`yarn clean-lage-cache && yarn lage buildci` encountered an unreproduced -desktop-test pack dry-run failure. Two unchanged plain `yarn lage buildci` -retries then reproduced the pre-existing `boundedCommand` parallel-startup -SIGKILL/SIGTERM assertion; the first also hit an unmodified private-storage -timeout, which passed in the last attempt. The focused serial diagnostic -passed 73 with 10 skipped. No environment/concurrency override or unrelated -test fix was used; this does not establish a green aggregate. - -Separate Markdown validation passed 90 link-check tasks. Final locked -cleanup proved **24 exact incarnations exited**, **3 ports closed**, zero -worktree processes and no shared/fixture claims. Four test workers briefly -outlived the failed aggregate but all exited naturally before cleanup -inspection. The final verifier and lock reacquisition passed. The harness -ledger preserves each exact command and every failed attempt. - -### 2026-11-12 review follow-up: bounded qualification fixtures - -Source commit `0fb6cd47fccc4f6210797b25085e7b56aa8ca4fe` bounds the -ready crash-owner fixture to a 120-second wait followed by exit 1 and adds -four lines requesting AppKit fixture termination on the main queue after -300 seconds. The `/bin/sleep 300` descendant was already bounded. Its focused -portable regression exercises the same wait with a short deadline and no -native resources. Production helper/ownership behavior, exact-identity -cleanup and durable claim fences are unchanged. - -The portable regression failed with `ETIMEDOUT` before the fix and passed -afterward. The focused harness suite passed 30 with no skips. Both owning -packages' format, lint and build passed; full portable suites passed driver -1,237 with 66 skipped and harness 410 with 14 skipped. - -**Native build passed, runtime blocked:** under the shared native lock, -`node packages/agentic/desktop-driver/native/macos/fixtures/build.cjs artifacts/desktop/f10m/retirement-followup-20261112/fixture` -compiled and strictly verified the stable-signed fixture. Nonprompting -doctor reported `screenLocked=true`, despite effective permission preflights. -No application was launched. The earlier native and Fabric/Paper passes are -pre-fix evidence present at `16558610ff2bd14775430f86c738c12947fe81ee`, -not qualification of the new retirement deadlines. - -The single fresh `yarn clean-lage-cache && yarn lage buildci` failed only the -unchanged `boundedCommand.test.ts:31` SIGKILL/SIGTERM assertion, with driver -1,236 passed, 1 failed, 66 skipped. No environment/concurrency override or -retry was used. Separate links passed 90 tasks. Final locked audits found -zero worktree workloads/unreadable candidates; no signals were issued. - -The harness ledger's "Review follow-up: bounded fixture retirement" section -records exact commands and limitations. Evidence is retained under -`artifacts/desktop/f10m/retirement-followup-20261112/`. - -## Windows Fabric and Win32 Paper — F10-W application ownership — 2026-11-12 UTC - -**Source and dependency:** Windows implementation commit -`6e881e1a05fef735c1b9bacd7d123285f47269ff`, branch -`user/jasonvmo/desktop-windows-app-ownership`, stacked on F10-M #4492 at -exactly `6489386956c0485cb91b478a8f4adc4828f4178b`. The intended draft PR -base is `user/jasonvmo/desktop-macos-app-ownership`; its independently owned -branch must not be rewritten. This implements the same neutral -`applicationOwnership` host/resource model, schema 3 claims, native OS owner -identity, `processIdentity`, `inspectApplication`, and launch settlement -metadata, rather than a parallel Windows contract. Application naming uses -the existing neutral publication seam; no additional F11 #4448/#4491-owned -file changes were made relative to the pinned base. - -**Observed, portable Windows host:** the driver's declared format, lint and -TypeScript build passed. With native opt-in unset, the full driver suite -passed **74 suites / 1,234 tests**, with **9 suites / 93 tests skipped**. -The focused `NativeDesktopHost.applicationIdentity` suite passed **31 tests**, -covering malformed/missing Windows image/session/AUMID/start evidence, -read-only inspection and trusted recovery response/input validation. -Test-first failures were observed before implementing the Windows boundary; -those fake-helper tests do not establish native exclusion or cleanup. - -Commands ran from the worktree root; `$W2` denotes that root. `TEMP`/`TMP` -were scoped to its ignored `artifacts\desktop\f10w\scratch` directory and -`FURN_NATIVE_DRIVER_TEST` was unset. - -| Command | Result | -| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-driver format` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver lint` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver build` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver test --runInBand --cacheDirectory "$W2\artifacts\desktop\f10w\jest-driver"` | Passed; native opt-ins skipped | -| `yarn workspace @fluentui-react-native/desktop-driver test --runInBand NativeDesktopHost.applicationIdentity --cacheDirectory "$W2\artifacts\desktop\f10w\jest-driver"` | Passed 31 tests | -| `yarn check-agent-instructions` | Passed | -| `yarn oxfmt --check` | Passed repository-wide, 3,163 files | - -Ignored run evidence: `artifacts/desktop/f10w/portable-20261112/driver-test.log`, -`driver-build.log`, `driver-lint.log`, and `changed-markdown-links.log`. -Every changed Markdown file's link check passed. The root formatter -deliberately excludes `.changeset/*.md`; the committed patch changeset -`windows-exact-application-ownership.md` covers driver, harness and Storybook. - -**Observed, compile only:** the locked direct `buildWindowsDriver` Debug/x64 -invocation (`buildId: f10w-compile-02`, staging -`artifacts/desktop/f10w/b2`) passed in **35.09 seconds, 0 warnings / 0 errors**. -All seven changed C++ source/header files match the successful build's source -snapshot byte-for-byte. The first compile failed on a self-test `std::wstring` -argument, corrected to `.c_str()` before the successful build. Native build -log: `artifacts/desktop/f10w/b2/build.log`; output: -`artifacts/desktop/f10w/b2/output/furn-desktop-driver-host.exe`. -No helper self-test, fixture or application was executed by this compile. - -**Implementation, not native qualification:** Windows leases and inspection -retain canonical image, seven-digit native creation time, session and packaged -AUMID. Exact pins reject a one-tick mismatch and ambiguous candidates. A -per-incarnation nonwaiting semaphore excludes controlling helpers; a losing -packaged activation receives no cleanup authority over another controller. -Unpackaged launches retain suspended-create / Job Object containment. -Packaged activation stays explicitly `uncontained`: a post-activation job -cannot remove its already-running race and may encounter existing job -restrictions. Exact primary-process cleanup is not arbitrary-descendant proof. - -| Native acceptance case | Windows Fabric packaged | Win32 Paper REX | Reason / required next evidence | -| ---------------------------------------------------------- | ----------------------- | --------------- | ---------------------------------------------------------------------------- | -| Exact AUMID / image / REX incarnation, start and session | Not run | Not run | Awaiting the lane's granted desktop campaign | -| Native one-tick/session/image rejection and ambiguity | Not run | Not run | Authored native fixture/consumer contracts; portable tests do not prove this | -| Cross-service exact-application exclusion | Not run | Not run | Must use independent real helpers | -| Job containment / explicitly fenced packaged activation | Not run | Not run | Runtime process/lease evidence required | -| Physical input mutex | Not run | Not run | Requires physical opt-in and connected/unlocked session gate | -| Acquired survivor journaling and next-run exact recovery | Not run | Not run | Requires identity-verified helper/owner fault injection | -| Five consecutive launches/teardowns without identity leaks | Not run | Not run | Required separately on both real endpoints | - -Native session kind is **not sampled**: no GUI/physical campaign has run. -Planned provider command, under the exclusive lock: -`$env:FURN_NATIVE_DRIVER_TEST='1'; yarn workspace @fluentui-react-native/desktop-driver test --runInBand applicationOwnership.contract`. -The consumer runbook is -`packages/agentic/desktop-test/references/native-qualification.md`, -“Neutral Windows application ownership”; it additionally exercises three -paired packaged activations and physical input across both endpoint helpers. -The pre-campaign audit observed `FURN-AUDIT: no FURN desktop processes found`; -no W2 campaign was started. These are preparation results, not native or -hosted-CI qualification. Part of #4447 and #4435. - -Preparation closeout on the implementation SHA above: the post-commit -focused driver rerun passed **31 tests**. `yarn change:check` passed; -`yarn lage test-links --verbose --grouped` passed **90 tasks, 0 -failed/skipped**. The two appended Windows ledgers' link checks passed -(driver 6 links, harness 15), and `yarn oxfmt --check` again passed -3,163 files. Only Windows evidence text followed those checks; no native or -runtime source changed. Additional ignored provenance under -`artifacts/desktop/f10w/portable-20261112/`: -`tested-source-revision.txt`, `compile-source-evidence.json`, -`driver-focused-committed.log`, `test-links.log`, and `root-format.log`. - -## Windows Fabric and Win32 Paper — F10-W native qualification — 2026-11-12 UTC - -**Observed source:** application qualification ran at -`8343dc0c48a2de01475454c892e7a295523caa75`; helper/self-test correction -`4a4b87f37b574820b93e342b68721d931be07c55` follows implementation -`6e881e1a05fef735c1b9bacd7d123285f47269ff`. The branch remains -`user/jasonvmo/desktop-windows-app-ownership`, pinned on #4492 at -`6489386956c0485cb91b478a8f4adc4828f4178b`. The coordinator handoff -`W2-base-decision.md` records **STACK**: W2 implements the inherited neutral -application host/resource model, schema-3 native owner claims, exact process pins, -inspection and settlement messages, rather than a second Windows model. -Application naming required no new F11-owned file edits. The eventual draft PR -base remains `user/jasonvmo/desktop-macos-app-ownership`; later dependency -movement does not authorize a rebase, merge or dependency-branch push. - -**Observed host and identity:** Windows build 26300 x64, Node 26.13.0, -active/unlocked RDP (`sessionKind: remote`, `windows-session:1`). Every native -command used the exclusive W2 lock and `-RequirePhysical`; before/after gates -stayed eligible. The actual endpoint helper's SHA256/artifact identity is -`863cc2dbfc8f5792d22c8606376d226affbf8655fe4ded62ba78c23a29c516c7`, -source digest -`ea76e931dc4cdebc87e0621a7d730f81faabce5ef902063a0512cd7c2b100f2d`, -Debug/x64. Resolution was prebuilt-only. - -The registered Fabric build was reused from the coordinator-specified -`furn-81` worktree, built at `bdeb0b291`: no Windows app/native-core Windows -source differs from that build. Its generated `RnwNewArch` and `UseFabric` -properties are true. All five ordinary native identities matched the registered -manifest's exact image and AUMID -`40411fc5-8e92-4d46-b68d-b62df44b1366_syyqehzwccsyy!App`. -Metro and both JavaScript bundles came from W2. Win32 used the canonical -prebuilt REX `ReactTest.exe` and W2's Paper bundle; Win32 native app build is -**Unsupported**, not a skipped build. - -**Observed provider result:** `nativeDriver`'s selected Windows build/reuse/ -resolution/handshake contract and all five new `applicationOwnership` fixture -cases passed: **2 suites / 6 tests passed**, 10 unrelated filtered cases skipped. -This includes the native C++ self-test and the actual physical mutex fixture. -The first self-test run failed because an existing release-count assertion -kept a pointer into a destroyed temporary JSON result. Commit `4a4b87f37` -retains the result for the assertion; the native rerun passed. This is a -self-test lifetime fix, not a change to release behavior. - -From `packages/agentic/desktop-driver`, the exact selected test command was: - -```powershell -yarn test --runInBand --cacheDirectory "$W2\artifacts\desktop\f10w\jest-driver" --runTestsByPath src\native\nativeDriver.test.ts src\native\windows\applicationOwnership.contract.test.ts --testNamePattern 'Windows exact application ownership|builds, reuses, resolves, and handshakes with the Windows helper' --json --outputFile "$W2\artifacts\desktop\f10w\native-20261112\driver-native-fixed-results.json" -``` - -Inherited settings were `FURN_NATIVE_DRIVER_TEST=1`, -`FURN_NATIVE_PHYSICAL_INPUT_TEST=1`, native cache -`$W2\artifacts\desktop\f10w\n`, configuration `debug`, build policy `never`, -and project-local `TEMP`/`TMP`. From `apps/storybook`, -`yarn desktop-driver doctor --platform windows --configuration debug --cache-root "$W2\artifacts\desktop\f10w\n"` -**passed** against the selected artifact. An initial doctor without the explicit -configuration selected its CLI default `release` and correctly reported no -verified prebuilt; that setup attempt was not a native application result. - -The separate real-endpoint consumer suite passed **9 tests / 0 failed / -0 skipped**: - -| Case | Windows Fabric packaged | Win32 Paper REX | -| --------------------------------------------------------------------- | ------------------------------------------------------------------------ | -------------------------------------------------------- | -| Exact AUMID/image/REX incarnation, native start and session | Passed | Passed | -| Start/session/image mismatch; independent exact-controller exclusion | Passed | Passed | -| Multiple actual matching processes/windows refuse unpinned inspection | Passed, three pairs | Passed, three pairs | -| Containment or explicit exact-identity fencing | Passed, `uncontained` and journaled | Passed, `job-object` | -| Physical Shift contention, effect-none cancellation and transfer | Passed | Passed | -| Live-owner fence, acquired journal and next-run recovery | Passed, packaged app survived helper crash then exited by exact recovery | Passed, Job Object drained then the claim was reconciled | -| Repeated launches and settled teardown | Passed 5/5, five distinct incarnations | Passed 5/5, five distinct incarnations | -| Close one concurrent winner without ending the other | Passed 3/3 pairs | Passed 3/3 pairs | - -Both packaged contenders started distinct processes in every pair. Same exact -incarnation exclusion was independently proved by pinned attach, but a -**shared-PID packaged activation race was not induced**. Actual kernel PID -recycling was **Not run**; the real fixture's one-100-ns-tick mismatch and both -endpoints' stale-start refusal passed. Arbitrary uncontained descendants and -mixed Storybook-graph recovery remain outside this bounded qualification. - -Ignored evidence root: `artifacts/desktop/f10w/native-20261112/`. -`driver-native.log` retains the initial failure; -`driver-native-fixed.log` and `driver-native-fixed-results.json` retain the -green provider run. `qualification-summary.json`, `selected-helper.json`, -`application-build-provenance.json` and `consumer-evidence-final.json` record -the tested source, binary hash, registration, bundles and per-case identities. -The real consumer run is `windows-application-ownership-GiW4x1`; outer command -report `4ed98f8f-8ebb-49ca-ac90-2ee6caece7dc` is a lifecycle report, not the -test-count source. The harness ledger records the full consumer commands. - -**Observed closeout:** all W2 campaigns ended; before/after audits were clean: -`FURN-AUDIT: no FURN desktop processes found`. No helper, fixture, Metro, REX -or packaged app from W2 survived. Review, the final one-time clean-cache -`buildci`, push, draft PR and hosted CI remain **Not run** at this review -handoff. The three-package patch changeset remains -`windows-exact-application-ownership.md`. Part of #4447 and #4435. - -## Windows Fabric and Win32 Paper — recovery review follow-up — 2026-11-12 UTC - -**Scope and provenance:** the review of `6489386956c0..6bdcde61f` found that -trusted recovery incorrectly rejected an application's reused PID. The fix is -commit `8b2fb7e67dc5864180c7863ac94fc3c0c3c9d6e3`, still stacked on the -pinned #4492 dependency `6489386956c0485cb91b478a8f4adc4828f4178b`. -No rebase, dependency-branch write, push, or PR occurred. - -**Implemented, not yet runtime-qualified:** recovery first queries native -creation time with `PROCESS_QUERY_LIMITED_INFORMATION`. A different time -settles the recorded incarnation as `exited`, without application control, -termination rights, or signals to its live replacement. An unreadable start, -including access denied, remains fenced. A matching start still requires -complete identity and exclusive control before exact cleanup. Inspection and -attach continue rejecting stale pins. This supersedes the earlier ledger's -recovery-specific stale-start refusal; that earlier native run does **not** -qualify the corrected recovery semantics. - -**Observed validation:** - -- **Passed:** package `yarn format`, `yarn lint`, and `yarn build`. -- **Passed:** `yarn workspace @fluentui-react-native/desktop-driver test --runInBand` - with native opt-in unset: **1,234 passed, 93 skipped**. Native skips are - not native qualification. Expanded arguments/results are retained under - `artifacts/desktop/f10w/review-20261112/driver-tests.log` and - `driver-results.json`. -- **Passed:** compile-only Windows x64 Debug `build-driver --platform windows ---configuration debug --cache-root artifacts/desktop/f10w/n --force ---reporter json`, through the declared package CLI and machine lock. It - waited for W3's campaign, acquired the lock at `2026-11-12T16:50:25Z`, - launched/registered no application, and exited zero. - Helper source digest: - `3558e8597f6cfe1ee3f4e38dca2137710b80d193a5c30bf946d54065bf8bc153`; - executable SHA-256: - `97f0072ce22449bc5c87c05de259d3e4c46eca666e6135977432b30fa16053b2`. - Full invocation/artifact receipt: ignored `helper-build.log` in that evidence - directory. Compilation is not native runtime or renderer qualification. - -**Not run on this corrected source:** native self-test, the expanded fixture -recovery case, and Windows Fabric/Win32 Paper consumer recovery. The opt-in -fixture now expects a stale recorded start to return `exited` while the live -controlled replacement survives; matching-start session/executable/AUMID -mismatches and another controller must still fail closed. Consumer Native RED -against the retained pre-fix helper and provider/consumer GREEN against the -rebuilt helper await the next granted W2 campaign, under the physical-session -gate where applicable. -No W2 campaign or session kind is claimed for this compile/portable phase. -Final clean-cache `buildci` and publication remain **Not run**. Part of #4447 -and #4435. - -**Observed closeout:** root `yarn oxfmt --check` (**3,163 files**), -`yarn check-agent-instructions`, `yarn change:check`, `git diff --check`, and -link checks for every edited Markdown reference and both ledgers **Passed**. -The first concurrently requested audit saw only the portable formatter's two -Node processes; the idle rerun was clean: -`FURN-AUDIT: no FURN desktop processes found`. No W2 desktop campaign was -started. The old and rebuilt helper files both match their recorded SHA-256; -the old binary remains available for the expected-failing consumer run. - -## Windows Fabric and Win32 Paper — default-environment buildci — 2026-11-12 UTC - -**Observed:** the root `yarn clean-lage-cache` followed by `yarn lage buildci` -passed at `68d63ccfbf008b963e59772f57aca8ac8f9ff05f`: **178 successful -tasks, 0 failed/skipped**, in 2m 32.26s. Both commands ran under the W2 machine -lock without `-RequirePhysical`, with default TEMP and concurrency and no -environment or repository-guidance adjustments. - -The first attempt at that same revision failed the Storybook server-integration -child-exit assertion and the inline-worker aggregate's 120s timeout. -Before considering a fix, the same full default-environment gate was run at -pinned base `6489386956c0485cb91b478a8f4adc4828f4178b`: **Passed**, 178/0. -The unchanged W2 rerun then passed. No source fix, timeout increase, dependency -install, or concurrency override was made; baseline failure was not reproduced. - -Ignored receipts are under `artifacts/desktop/f10w/review-20261112/`: -`buildci-default.log`, `buildci-default-base.log`, -`buildci-default-rerun.log`, `buildci-default-summary.json`, and -`buildci-default-final-audit.log`. The final audit **Passed**: -`FURN-AUDIT: no FURN desktop processes found`. -No app or native contract was launched. The desktop remained disconnected; -both endpoints' corrected native recovery qualification remains **Not run** -pending an active/unlocked session and a separate grant. No push occurred. -Part of #4447 and #4435. - -## Windows Fabric and Win32 Paper — recovery native re-verification — 2026-11-12 UTC - -**Observed source:** native implementation -`dbaf558939db31b4a3935ad68b1eb20d25d90d3e`, stacked on pinned #4492 -`6489386956c0485cb91b478a8f4adc4828f4178b`. The helper and public host API -were unchanged in this phase. The harness consumer needed a test-only fix: -`recoverApplication` checks native `{ state: 'exited' }` internally and resolves -`Promise`, so the consumer must expect `undefined`. That one-line -assertion correction is committed with this evidence and needs review before -publication; no production, protocol or F11 file changed. - -**Observed granted native session:** active/unlocked **remote** session 1, -input desktop `Default`, with eligible before/after checks for every native -wrapper. Both W2 campaigns used the machine lock and `-RequirePhysical`; no -exit 75 or session change occurred. The final consumer wrapper ran -`21:54:54Z–21:57:02Z`. Fabric registration was checked before each phase and -remained the `furn-81` Debug/x64 build at `bdeb0b291`, with no relevant native -source difference and executable SHA-256 -`df9cb3559843fcfb0158ee4ab5edbae78890c4e231ebd895cbb636c24b18cd45`. -The exact AUMID remained -`40411fc5-8e92-4d46-b68d-b62df44b1366_syyqehzwccsyy!App`. -REX Paper used the canonical worktree image `0.81.1-6828073195`, exact launch -arguments, a fresh Win32 bundle and owned worktree Metro. A Win32 native-app -build remains **Unsupported**. - -**Observed native RED/GREEN:** the retained pre-review helper's source was -`ea76e931dc4cdebc87e0621a7d730f81faabce5ef902063a0512cd7c2b100f2d`, -executable SHA-256 -`863cc2dbfc8f5792d22c8606376d226affbf8655fe4ded62ba78c23a29c516c7`. -It produced **4 expected failures / 0 passed / 7 filtered skips** across the -two endpoints' stale-start and crashed-recoverer cases: -`recoverApplication` rejected with `identity-start-mismatch`. - -The selected review-fixed consumer helper had source -`3558e8597f6cfe1ee3f4e38dca2137710b80d193a5c30bf946d54065bf8bc153` -and executable SHA-256 -`97f0072ce22449bc5c87c05de259d3e4c46eca666e6135977432b30fa16053b2`. -Its first consumer run had **9 passed / 2 failed**, solely because the test -expected the wire object rather than the host's void resolution. -After correcting only that assertion, all **11 consumers passed / 0 failed / -0 skipped**. Native recovery settles a different recorded start without -control/termination access or replacement signals; matching-start -session/executable/AUMID mismatches still fail closed. Unreadable native starts -remain fenced; no real protected-process access-denial case was induced here. - -**Exact provider command:** from the driver package, inside the granted -physical-gated machine lock, with `FURN_NATIVE_DRIVER_TEST=1`, -`FURN_NATIVE_PHYSICAL_INPUT_TEST=1`, Debug configuration and worktree-local -scratch/cache. `$W2` denotes the assigned worktree; `$E` is its ignored -`artifacts\desktop\f10w\review-native-20261112` directory. -The direct-helper override is unset for the cache-resolution contract. - -```powershell -yarn test --runInBand ` - --cacheDirectory "$W2\artifacts\desktop\f10w\jest-driver" ` - --runTestsByPath src\native\nativeDriver.test.ts src\native\windows\applicationOwnership.contract.test.ts ` - --testNamePattern 'Windows exact application ownership|builds, reuses, resolves, and handshakes with the Windows helper' ` - --json --outputFile "$E\driver-green-results.json" -``` - -| Provider/self-test case | Outcome | -| ------------------------------------------------------------------------------------------------ | --------------- | -| Build, reuse, resolve, handshake and Windows helper self-test | **Passed** | -| Noncontrolling inspection and exclusion of another independent helper | **Passed** | -| Reused recorded start settles without signalling the live replacement; other mismatches rejected | **Passed** | -| Multiple matching process/window candidates rejected | **Passed** | -| Five contained exact-identity launch/close cycles | **Passed**, 5/5 | -| Independent applications serialize through the physical input mutex | **Passed** | - -Final provider result: **6 passed / 0 failed / 10 unrelated filtered skips**. -The first invocation mistakenly set a direct-helper path, so the five ownership -cases passed but the isolated cache test resolved that explicit artifact -instead of its temporary build. Removing the conflicting invocation override, -without changing source, produced the full passing result. Both receipts are -preserved; this was not a resolver fix or an additional production defect. - -**Observed real consumer results:** both endpoints passed five distinct -PID/start incarnations, exact independent-service exclusion, native mismatch -refusals, crashed-recoverer barrier retirement, crashed-owner journal -reconciliation and three concurrent pairs. Fabric consistently reported -`uncontained` and exact survivor cleanup; REX reported `job-object` and -reconciled its job drain. The cross-endpoint physical-mutex case passed with -cancellation, release and transfer. All recorded application sessions were -`windows-session:1`. The crash fixtures verified each owned child's native -identity before literal-PID termination; the new dead-recoverer PIDs were -**54440** and **55808**. Live replacements survived stale-record retirement. -The harness's matching section records all 11 cases and exact consumer -commands. - -Ignored evidence root: `artifacts\desktop\f10w\review-native-20261112`. -`driver-green-results.json`, `red-consumer-results.json`, -`green-initial-consumer-results.json`, `green-consumer-results.json`, -`green-consumer-evidence.json`, `consumer-assertion-fix.diff` and -`qualification-summary-final.json` retain the failures, corrected results, -digests and actual test-source provenance. Final consumer evidence: -`windows-application-ownership-3BLzLV`; lifecycle report -`32f3d6be-c4b0-4792-aeb6-a5d7df165fc0` has no test counts. - -**Observed portable/aggregate closeout:** harness format/lint/build and -**453** portable tests **Passed** (**28** native skips). Existing driver -`NativeDesktopHost.applicationIdentity` contracts **Passed**, **31/31**. -Locked default-environment `yarn clean-lage-cache` then `yarn lage buildci` -initially **Failed** the reported intermittent Storybook `serverIntegration` -readiness assertion, `A stopping command cannot publish readiness` -(`176` successful tasks, one failed). The cause remains unproven. -After the fail-fast test children settled and a clean audit, the sole unchanged -full rerun **Passed**, **178 successful / 0 failed / 0 skipped**, 2m 38.65s. -No TEMP, concurrency, timeout, dependency or guidance workaround was applied. -Both W2 campaigns ended and the final audit **Passed**: -`FURN-AUDIT: no FURN desktop processes found`. - -**Publication/remaining limits:** no push or PR-body update occurred because -the test-only correction needs review. Initial PR #4494 CI at `dbaf558939` -was observed green: run `43282149276`, all **19/19** checks successful. -Windows job `131405956387` selected the ownership provider suite -(`PASS applicationOwnership.contract.test.ts`; **4** nonphysical ownership -cases selected, physical case gated off; full driver **1,281 passed / -46 skipped**). The Win32 jobs did not select that ownership step, and the -actual-endpoint consumer opt-in was absent: **0/11** consumers hosted. -These are historical hosted counts, not new CI for the local test correction -or substitutes for this physical campaign. F11 claim naming is deferred and -untouched. Actual kernel PID recycling and shared-PID activation were -**Not induced**; all concurrent winners were distinct. Arbitrary uncontained -descendants and mixed-graph recovery remain unqualified. The existing -`windows-exact-application-ownership.md` patch changeset remains applicable. -Part of #4447 and #4435. - -**Observed reference checks:** Markdown links in both edited ledgers -**Passed** (6 driver links, 15 harness links). Root `yarn oxfmt --check` -(3,163 files), `yarn check-agent-instructions`, `yarn change:check`, and -`git diff --check` **Passed**. Their receipts are in the same ignored evidence -directory. - -## F15 non-AX macOS supporting evidence — 2026-11-12 - -Source `505a2187b0a2451aed7081c9260b343d55742f93`; macOS 27.2.1 -(26B131), arm64 (`sw_vers`). App: `FurnDesktopDriverHost.app` and -process-identity contract children, not Storybook. Renderer: N/A. -Feasibility was recorded before execution in the machine-local ignored -artifact `artifacts/desktop/f15-20261112/feasibility.md`. - -All native commands ran serially through the machine-wide native lock. -From the repository root, the shared environment was: - -```sh -export FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" -export FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f15-20261112/native-cache" -``` - -| Command / selection | Passed / failed / skipped | Cleanup / artifact locator | -| ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver --reporter json build-driver --platform macos` | Passed build and signature verification; 0 tests selected | Build exited; `artifacts/desktop/f15-20261112/helper-build.json` and `helper-build.stderr` | -| `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver --reporter json doctor --platform macos --permissions` | Passed diagnostic invocation; 0 tests selected. Accessibility=false, PostEvent=false, ScreenCapture=true; no prompt requested | Diagnostic helper exited; `artifacts/desktop/f15-20261112/doctor.json` and `doctor.stderr` | -| Signed helper `--self-test`, resolved from `helper-build.json`; exact invocation below | 448 / 0 / 0 checks, including 8 strict-clear refusal/negotiation assertions | Helper exited; `artifacts/desktop/f15-20261112/signed-self-test.log` and `signed-self-test.stderr` | -| `FURN_NATIVE_DRIVER_TEST=1 yarn workspace @fluentui-react-native/desktop-driver test --runInBand nativeDriver.test applicationIncarnation.contract serviceIdentity.contract` | 27 / 0 / 15 tests; 3 passed suites, 2 platform-skipped suites; default TMPDIR | Test/helper children exited; `artifacts/desktop/f15-20261112/native-contracts.log` | - -The stable-signed self-test invocation retained the selected artifact as data, -with a 120-second subprocess timeout: - -```sh -python3 -c 'import json,subprocess; from pathlib import Path; root=Path("artifacts/desktop/f15-20261112"); artifact=json.loads((root/"helper-build.json").read_text()); print("helper="+artifact["artifactId"]); result=subprocess.run([artifact["executablePath"],"--self-test"],stdout=(root/"signed-self-test.log").open("w"),stderr=(root/"signed-self-test.stderr").open("w"),timeout=120); print("signed-self-test exit="+str(result.returncode)); raise SystemExit(result.returncode)' -``` - -Stable helper artifact: -`bdc07c7c4333eeef6cf5da683128591da0564cf978cf813335b5c63494dc7903`; -source digest: -`b1d5233de326faf886c8d6310c2844c6857e4936686dfd493ff898f9e5ec6b9a`. -Its designated requirement was -`identifier "com.microsoft.fluentui-react-native.desktop-driver" and certificate leaf = H"ab592605f8619fcaa952b4a2525ae6d41296fdd7"`. -The doctor reported the same artifact and signing requirement, an unlocked -console session, and `promptRequested=false`. Its `ready=true` describes -artifact availability, not permission readiness. - -Strict-clear coverage in this initial pass is limited to four invalid-mechanism -refusals, semantic/keyboard missing-element refusals with no effect, and -probe/hello negotiation. Successful real-element clears, Fabric/Paper -Storybook descriptor lifecycle and smoke, physical input, and Paper MCP -interaction measurements were blocked on Accessibility/PostEvent during that -pass and were not launched or credited as passes. The later permission recheck -below resolves that prerequisite, not the unexecuted acceptance cases. -No privacy settings changed. -Process inspection after the contracts found no surviving test/helper -processes from this worktree. This supporting evidence is **not macOS -requalification**, native positive strict-clear qualification, or a reliability -window; the existing F15 acceptance gaps remain open. - -### F15 permission-enabled follow-up — 2026-11-12 - -**Observed, partial qualification, not macOS requalification.** Product source: -`141b88b4b81a1bd9a0472de533681f9274d8cb00` (main plus this PR's -documentation); macOS 27.2.1 (26B131), arm64. No sibling-branch evidence is -included. The pre-execution feasibility note and ignored machine-local -artifacts are under `artifacts/desktop/f15-20261112T115411Z/`. -The signing identity/cache environment above was reused. Every native, -GUI and physical-input command ran through -`$HOME/Library/Caches/furn-overnight/with-native-lock.sh`; -unit tests retained the default TMPDIR. No privacy settings changed and no -permission prompts were requested. - -| Check / exact command from the repository root | Observed result / passed, failed, skipped | Artifact beneath the follow-up directory | -| -------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver --reporter json doctor --platform macos --permissions` | Passed diagnostic: Accessibility=true, PostEvent=true, ScreenCapture=true; unlocked, onConsole=true, promptRequested=false; 0 tests | `doctor.json`, `doctor.stderr` | -| `FURN_NATIVE_DRIVER_TEST=1 yarn workspace @fluentui-react-native/desktop-driver test --runInBand` | Passed: 1,240 / 0 / 47 tests; 75 passed, 6 platform-skipped suites; includes native build/identity contracts and portable fake-host tests | `native-tests.log` | -| `node artifacts/desktop/f15-20261112T115411Z/qualify.mjs fabric strict` | Passed: 3 / 0 / 0 manual checks: real-element keyboard clear/input, semantic clear, protected args refusal | `fabric-strict-qualification.json`, `fabric-strict.log` | -| `node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper` | Failed: 14 / 4 / 0 manual assertions; includes a dependent semantic-clear precondition failure and unsupported Checkbox activation, not four independent product defects | `paper-qualification.json`, `paper-qualification.log` | -| `node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper input` | Failed overall: 4 / 1 / 0 manual checks; semantic clear, independent physical-input delivery, args refusal and MCP type/check/observe passed; keyboard clear failed | `paper-input-qualification.json`, `paper-input.log` | - -The doctor resolved the same stable helper artifact and source digest as the -initial pass. **Inferred:** this preflight is authority for the current -responsible process tree (Rex Server.app), not proof of helper-owned durable -TCC continuity. - -The real element was Storybook `components-input--default`, -`agentic-storybook-input`. The physical session verified typed text, cleared a -nonempty value to empty with the required keyboard mechanism, then exercised -element-relative pointer down/up, an arrow key and one wheel event. -Its input receipt was `{ keyboard: 4, pointer: 1, wheel: 1 }`. -A subsequent accessible session verified the nonempty value and cleared it to -empty with exactly `{ "semantic-clear": 1 }`, no keyboard fallback. -**Observed:** wheel delivery succeeded; its visual scrolling effect was not -asserted. These are manual real-element checks, not synthetic self-test -assertions. The strict run stopped the exact owning CLI with SIGTERM (exit 0), -then terminated and verified exit of its own app. - -Paper's first keyboard clear returned success but changed the verified -synthetic value `F15` to `F1`, not empty. Its next semantic check incorrectly -required the original `F15`, so that assertion failed **before semantic -clear was attempted**. The separate `paper input` probe made the semantic -precondition independent and polled the value after typing and clearing, -without retrying either mutation. **Observed:** keyboard clear still left -two characters after 100 reads separated by 50-ms waits; its receipt was -`{ keyboard: 2 }` for typing and clear. Semantic clear then passed from a -nonempty value to empty with exactly `{ "semantic-clear": 1 }`. -A separate Paper physical session verified text and delivered pointer, -arrow-key and wheel actions with -`{ keyboard: 2, pointer: 1, wheel: 1 }`; visual scrolling was not asserted. -Thus real-element keyboard clear passed on Fabric, not Paper; semantic -clear passed on both. Receipt delivery is not proof of successful clearing. - -Fabric descriptor follow-up: -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs fabric lifecycle` -recorded **10 passed, 1 failed, 0 skipped manual checks**; -`fabric preservation` added **1 passed, 0 failed, 0 skipped**. -The protected descriptor/native helper was real throughout. -CREATE/DELETE response-loss and delayed-open injections wrapped the public -typed client method boundary; they were not TCP fault proxies or forced -native cleanup failures. Grant revocation, terminal drain/reacquisition, -same-port replacement fencing, read cancellation and target-only PNG -resource reads passed. A separate PID/start-time/executable comparison proved -the exact externally attached app survived session and service close. -Artifacts and consumer commands are detailed in the harness ledger's -"F15 permission-enabled consumer follow-up". - -The Fabric and Paper full `stories-and-tests` smokes each failed: -**223 passed, 2 failed, 0 skipped**, 225 selected/executed per renderer, -no retries/flaky entries. -The 223 passes were navigation; later authored callbacks remained not run. -This does not invalidate the separate real-input delivery receipts, but it -does leave successful physical control activation unqualified. - -#### Open F15 native findings - -**Observed, unresolved; no product fixes in this docs-only PR.** - -- `packages/agentic/desktop-driver/src/operations/act.ts`, `physicalClick` - / `obscuredRejection`: through either real renderer's MCP descriptor, select - `components-checkbox--default`, find `agentic-storybook-checkbox`, then - `desktop_act` with one `click` step citing its unique ref/sample. - The sequence returns `stopped` at step 0, `element click intercepted`, - reason `obscured`, mutation `rejected` / effect `none`, not quarantined. - The refusal measured 1,187.5 ms, 107 text tokens, 438 UTF-8 bytes in the - lifecycle follow-up; this is **not a successful click measurement**. - Evidence: `fabric-lifecycle-qualification.json`, `physical-click` measurement. - Paper reproduced the same refusal in 1,172.4 ms, 106 tokens, 438 bytes - (`paper-qualification.json`). Paper semantic `activate` was separately - **Unsupported** by this element's provider (743.9 ms, 79 tokens, - 329 bytes); that response is not a successful alternative activation. -- `packages/agentic/desktop-driver/native/macos/Sources/DesktopDriverHost/Driver.swift`, - `clear`, and `Input.swift`, `pressCommandADelete`: reproduce with - `node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper input`, - using the instance-aware Paper app. Create a physical session, select - `components-input--default`, find `agentic-storybook-input`, type and - verify `F15`, then clear with required mechanism `keyboard`. The helper - returns success and records keyboard input, but the value remains - nonempty through bounded read polling. Accessible semantic clear succeeds - independently. Evidence: `paper-input-qualification.json`; exact local - harness `paper-input-harness.mjs`, SHA-256 - `183d2ef5f943c8d384598dc1f5976611f9df72c5bc8481ef83345d39fc0fc278`. - **Inferred routing:** macOS driver keyboard-clear/focus owner with the - Paper Input provider owner; the root cause is not established. -- `packages/agentic/desktop-driver/native/macos/Sources/DesktopDriverHost/Driver.swift` - (activation) and - `packages/agentic/components/src/components/accordion/accordion.stories.tsx` - (`KeyboardDisclosure`): reproduce with - `yarn storybook smoke --macos --mode stories-and-tests --reporter json` - from `apps/storybook`. Its two keyboard callbacks fail in the initial - `focusByTab` click: own-window obstruction, then failure to make the target - foreground. Report: - `apps/storybook/artifacts/desktop/8e1bfc4a-3037-42af-8d52-47739db9065d/run.json`. - The matching Paper smoke (`--paper`) reproduced own-window obstruction - in both callbacks: - `apps/storybook/artifacts/desktop/9f8bb135-907f-419c-8c23-d543fc201230/run.json`. - -**Inferred routing:** macOS desktop-driver hit-testing/geometry and activation -owners, with the Accordion story/focus-test owner for the callback -reproduction. Whether the cause is provider geometry, app state or an -environmental foreground transition is not established; no root cause or -component defect is asserted. The coordinator should route these open -findings under #4441/#4435. Semantic activation must not replace the missing -physical-click evidence. - -#### F15 remaining acceptance and cleanup - -**Not run on this source:** forced native failed-cleanup claim retention; -startup/shutdown survivor and channel/helper/Metro crash reconciliation; -unknown-effect story quarantine/trusted recovery; post-dispatch mutation -cancellation; correlated readiness/render faults; lost-context fresh-preview -fencing and authenticated app-relaunch recovery; registered secondary -windows; anonymous/foreign HTTP/WebSocket protected ingress; launch-capable -discovery and explicit legacy-lane separation. Read cancellation and normal -shutdown are narrower than these cases. Fabric's idle-channel interval was -not measured in this follow-up; Paper passed a single 22-second interval. -The reliability window, durable TCC continuity, and clients beyond the MCP -SDK are **Not run**; Windows and Win32 are **Not run** on this host. -Failed smokes, successful Checkbox/Accordion physical activation, remaining -authored callbacks and Paper keyboard clear remain unqualified. Permissions -are no longer the blocker. This does not meet #4441's macOS acceptance. - -**Observed cleanup, passed:** the isolated -`node artifacts/desktop/f15-20261112T115411Z/audit-cleanup.mjs` command, -under the native lock, found no worktree processes or listeners on -39860/23935/28029, no default descriptor and no per-user workload claim at -2026-11-12T12:20:26Z (`cleanup-audit.json`). Every launched app had been -terminated after exact-incarnation verification. The wrapper exited 0; -subsequent `lsof` found no holder of the machine-wide `native.lock`. - -## F11 scoped runner handoff — 2026-11-12 - -Source `5c9d0191e` implements the reviewed generic attach-only phase: attenuated -runner/controller grants, bounded busy by default, operator-authorized -preemption, and target reservation only after in-flight acquisition, -session deletion, and controller-subtree drain settle. Revocation callbacks -alone are not a drain barrier. Failed cleanup remains fenced and retryable; -unknown runner input is rejected before granting or reserving authority. -Settled drains restore the 4096 live/draining grant capacity; only the last -4096 settled cleanup proofs are retained, never pending or failed proof. -Busy/failed attempts issue no runner grant, and shared descriptor/workload -holders require relinquish instead of subtree preemption. Retired requests -still check live authority after asynchronous target probes. -Driver package format, lint, build, and tests passed (1,233 passed, -66 skipped). The default-`TMPDIR` focused authority/delegation/handoff selection -passed 51 cases, including 4,101 drained delegation cycles under one live parent. - -Native selection at `87bc8b710` on macOS 27.2.1 arm64 passed 27 identity -and backend cases, with 15 Windows cases skipped. The stable-signed helper -doctor was ready but reported Accessibility/PostEvent `false`. The real -private-descriptor AppKit consumer then failed session creation with no -supported click input mode. Native handoff and worker-crash drain remain -unqualified; no permissions were changed. Cleanup preserved the externally -owned fixture through service shutdown, removed both descriptors and -released the claim before its spawning test ended that exact process. -The Accessibility-blocked native contract was not repeated for review fixes. - -Exact commands, helper/fixture revisions, the initial stale-fixture failure, -one corrected-fixture retry, cleanup evidence and artifact locators are in -`@fluentui-react-native/desktop-test/references/current-status.md`. -Windows/Win32 native handoff is not selected by existing PR smoke/E2E jobs -and remains not run. Storybook integration, retained launches, and -cross-connection resume/replay are not implemented by this phase. - -## Windows Fabric and Win32 Paper — F11 handoff qualification — 2026-11-12 UTC - -**Observed:** source `c9753208d0f5602d28510d84ededf3f2fc3c3855`, -Windows x64, active/unlocked RDP (`sessionKind: remote`). Every native command -ran inside the W1 campaign through `with-native-lock.ps1 -Lane W1`; native -contracts and live endpoint runs used `-RequirePhysical`. Both session gates -remained eligible. No physical input was delivered by the handoff probes; -these results are not physical-input qualification. - -From this package, with `FURN_NATIVE_DRIVER_TEST=1`: - -```powershell -yarn test --runInBand serviceIdentity.contract applicationIncarnation.contract nativeDriver --json --outputFile \artifacts\desktop\f11\w1\native-contracts.json -yarn test --runInBand protectedService.contract --json --outputFile \artifacts\desktop\f11\w1\protected-service.json -``` - -The first selection **passed 23, skipped 19, failed zero**. All 15 Windows -cases skipped in the earlier macOS selection executed and passed here: - -| Windows case | Result | -| ------------------------------------------------------------------------- | ------ | -| Service identity: describes the publishing supervisor process | Passed | -| Service identity: refuses nonexistent/exited/foreign-principal processes | Passed | -| Service identity: verifies only the process owning the loopback listener | Passed | -| Service identity: refuses changed start time, image or session | Passed | -| Service identity: compares executable paths case-insensitively | Passed | -| Service identity: refuses an unserved address | Passed | -| Service identity: refuses an exited process | Passed | -| Service identity: refuses `http://localhost:49152` | Passed | -| Service identity: refuses `http://192.0.2.1:49152` | Passed | -| Service identity: refuses `https://127.0.0.1:49152` | Passed | -| Service identity: refuses `http://127.0.0.1` | Passed | -| Service identity: refuses `http://127.0.0.1:49152/status` | Passed | -| Service identity: keeps IPv6 and IPv4 listeners apart | Passed | -| Application incarnation: proves exact-process exit, not a weaker identity | Passed | -| Native driver: builds/reuses/resolves/handshakes the Windows helper | Passed | - -The other eight passing cases are portable native-driver source/build-policy -checks. All 19 skips are macOS-only: 16 service-identity cases, one -application-incarnation case and two helper-build/signing cases. They are -**skipped**, not Windows passes. Every exact test title, source file, outcome, -skip reason and mapping to the 15 earlier skips is retained in -`artifacts\desktop\f11\w1\case-mapping.json`; the raw Jest results and -`native-contracts.log` are alongside it. - -`protectedService.contract` was included because it exercises changed -session-owner admission, busy, revocation and drain paths with a real helper. -Three cases **passed**: grant/owner isolation, revoked-owner handoff, and -quarantine retention without exit proof. The recovering-quarantine case -**failed before its assertions, environment-blocked**: the project-local -scratch directory's ancestry is refused by Windows private storage -(`reason: ancestor`). A native diagnostic reproduced that refusal and opened -the default private store successfully (`storage-probe.json`). No ACL was -changed. Both the Node storage adapter and native storage implementation are -unchanged from this branch's main base. **Inferred:** this is storage -placement, not an F11 regression. Recovery in that additional contract remains -unqualified by this run. Other native fixture suites were not repeated: -F11 changes no native protocol/provider code, and live RN endpoints below -exercise the changed generic handoff. - -**Observed, live native:** Windows Fabric packaged Storybook and Win32 Paper -REX each passed default busy, explicit MCP relinquish, operator preemption, -normal worker drain, killed-worker drain, and real lost-acquisition-response -drain. A fresh MCP session succeeded after each runner; stale references, -observation cursors and captured-resource links were rejected. Preemption -invalidated the old binding (`service-unavailable` at the MCP adapter), kept -MCP busy while the runner held the reservation, then allowed a fresh child. - -| Endpoint | Local evidence root | Terminal runner runs: normal / preempt / crash | -| -------------- | ----------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Windows Fabric | `artifacts\desktop\f11\w1\windows-3tYoml` | `a3bf02e2-df22-4a63-92f8-686057ef2a80` passed / `24d38403-b4e5-49d6-9e0f-d1d123e4fb17` passed / `8f702133-e18a-41f0-b05d-ce5fcb5cff1a` failed as injected | -| Win32 Paper | `artifacts\desktop\f11\w1\win32-BefbqE` | `d79fd672-df2d-4943-8b40-c017a8978380` passed / `f1f17781-08aa-4cac-a6d8-34cb13bd749b` passed / `a09a8c20-1380-4933-a7a3-9c2d786352c3` failed as injected | - -The retained reproduction is `node artifacts\desktop\f11\w1\qualify.cjs -windows` or `... win32`, from the worktree root under the W1 physical/session -lock. It composes the real `DesktopSupervisor`, verified native descriptors, -MCP SDK/server and WDIO/Jasmine, mirroring the macOS fixture procedure rather -than using the unsupported Storybook handoff. Apps are externally launched -through the public native host with exact AUMID/REX identity. Windows Metro -uses the app's declared `yarn start --no-interactive --port 8081`, from this -worktree; Win32 uses `yarn storybook bundle --win32`. The registered Fabric -build at `bdeb0b291` was reused: the requested native-source diff contains only -native-core changelog/version metadata, no Windows native source changes. - -Helper artifact: `d8ee05e4796fb3aa22f7b5dbf481bd8cdfb0c20b6d5d24676d689dc7900d5b3c`. -Both external applications remained the same running process incarnation -through service shutdown; only their exact recorded launcher then closed them. -Both descriptors and workload claims were removed. All `run.json` files are -terminal, including the deliberately failed crash runs (one selected, -zero completed, one incomplete; cleanup succeeded). The service lifecycle -runs passed. A scan of 45 retained text/config/report files found no bearer -credential. Final audit: `FURN-AUDIT: no FURN desktop processes found`; -campaign ended. - -**Observed, portable:** uncached `yarn workspace -@fluentui-react-native/desktop-driver test --runInBand` passed 1,211, skipped -88, failed zero with native opt-ins unset. `yarn build` passed. Windows -`yarn desktop-driver doctor --platform windows` and `--platform win32` -each **passed** with `ready: true`; adding `--permissions` is **unsupported** -(exit 1), not a permission pass. -The existing three-package `scoped-runner-handoff` patch changeset is retained. - -**Remaining automation gap:** desktop-test's `nativeRunner.contract` is -darwin-gated and its fixture uses `Contents/MacOS/fixture`; Windows execution -selects one skipped case even with the native opt-in. The CI native-driver -step runs driver contracts, not this generic MCP/WDIO consumer. A Windows -fixture path and explicit consumer selection are proposed to the coordinator; -no test/source or CI code was changed by this qualification lane. - -**Evidence validation:** `yarn markdown-link-check -c -.github\markdown-link-check-config.json -packages\agentic\desktop-driver\references\current-status.md` passed -(six links). Root `yarn oxfmt --check`, `yarn change:check` and -`git diff --check` passed. No clean-cache `buildci` rerun was required for -these evidence-only markdown edits. - -## Windows hosted native-runner fixture provisioning — 2026-11-12 UTC - -**Observed:** hosted PR run `43279226602` at branch source -`23f974a727ca2627690304cbecee0fa96cb6de83` failed before the selected native -MCP/WDIO consumer in Windows Storybook job `131395742955` and Win32 -accessible Storybook job `131395742289`. Both helpers passed `doctor`; the -Windows job additionally passed 1,254 driver tests with 45 capability/platform -skips. These are not nativeRunner passes: its ownership modes and accessibility -input were **Not run** in both jobs. - -The fixture provisioner's built ESM import failed with `ERR_MODULE_NOT_FOUND` -for its extensionless `NativeDesktopHost` dependency. The test-only source also -used CommonJS `__dirname`. The repair uses `.js` specifiers and the existing -package-root anchor, preserving the fixture source directory for both Jest and -built Node consumers. The C++ fixture, helper protocol, cache policy and native -entrypoints are unchanged. - -**Observed, portable:** tested repair source -`0e2dcb26ca156133cb577fc81869b23cebf1df3a`. From this package, -`yarn test nativeDriver -t 'loads Windows fixture provisioning from built ESM' ---runInBand` was **Failed as expected** before the repair with the exact hosted -import error, then **Passed** after building the correction. It imports the -actual built module in an ordinary Node subprocess, without Jest's resolver, -native compilation or application launch. - -`yarn format`, `yarn lint`, `yarn build` and `yarn test --runInBand` all -**Passed**; the complete portable driver suite passed 1,212 and skipped 88. -The root `yarn clean-lage-cache && yarn lage buildci`, serialized through -`with-native-lock.ps1 -Lane W1`, **Passed** with inherited temporary-directory -settings and default concurrency: 178 successful tasks, zero cache hits or -failures, 2m 52.86s. No native opt-in or physical-session requirement was used. -The first post-run audit reported another lane's portable Yarn cache-clean -process, not a process owned by this repair; it was not terminated. - -Receipts, exact commands, job logs and doctor artifacts are retained under -`artifacts\desktop\f11\w1-fix\push-ci`. Hosted requalification of the repaired -provisioner is **Not run** at this receipt. Local nativeRunner and physical -console Ctrl+C remain **Not run** because the desktop session is disconnected -and no desktop campaign is granted. Portable loading is not native evidence. - -## Windows native fixture consumers and console cleanup — 2026-11-12 UTC - -**Observed at `b1bbb26df880252a184ea5a3d2e1848bd811e83f`:** explicit -`yarn desktop-driver build-driver --platform windows` and -`yarn desktop-driver doctor --platform windows` **Passed**, the latter -with `ready: true`. Built ESM fixture provisioning succeeded without Jest. -The C++ fixture and provider/protocol sources are unchanged; verified helper -artifact `d8ee05e4796fb3aa22f7b5dbf481bd8cdfb0c20b6d5d24676d689dc7900d5b3c` -was used. - -With `FURN_NATIVE_DRIVER_TEST=1`, the locked root command -`yarn workspace @fluentui-react-native/desktop-driver test -serviceIdentity.contract applicationIncarnation.contract nativeDriver ---runInBand --json --outputFile -\artifacts\desktop\f11\w1-native\native-driver-jest.json` **Passed**: -24 tests passed, zero failed, 19 macOS-only cases skipped. This mixed -selection includes the new ordinary-Node ESM import regression as well as -the Windows native identity/backend contracts. No skip is a native pass. - -The newly enabled Windows desktop-test consumer also **Passed** all five -outer modes against the plain Win32 fixture: handoff, operator preemption, -worker-crash recovery, lost successful runner-grant response and capture. -The four non-crash modes each executed two inner specs, including native -accessibility input, with zero skips. Crash recovery intentionally produced -a terminal failed WDIO run before reuse. Every mode confirmed native deletion, -descriptor withdrawal and workload-claim release while preserving the -external fixture; only the scenario's own launcher then ended it. -This qualifies generic native transport/ownership, not React Native rendering -or Storybook runner handoff. - -Separately, physical console Ctrl+C **Passed** for real Storybook -`desktop-test serve --agent` on both renderers in an active unlocked RDP -session (`sessionKind: remote`). Exact-console foreground/isolation checks -preceded four real `SendInput` keyboard events; the observer received console -`SIGINT`, and no semantic signal generation was used. Both apps had completed -authenticated Button story/native-marker and MCP observation checks; the -MCP session was closed before interrupting the long-lived service. - -| Endpoint | Terminal passed lifecycle | CLI / Yarn / observer exit | -| --------------------------- | -------------------------------------- | -------------------------- | -| Windows Fabric packaged app | `75c40800-571c-4d8f-9fa1-debf876f0664` | 0 / 0 / 0 | -| Win32 Paper REX | `9a5c618f-d54d-4599-b36f-e3fd5b5436b9` | 0 / 0 / 0 | - -The service withdrew its descriptor, released its claim and preserved the -same running app incarnation. Its recorded external launcher then closed -that app. Metro, channel and helper cleanup completed without survivors. -Fabric reused the verified registered `bdeb0b291` native build because its -Windows native sources are unchanged; the app config was inherited with an -explicit verified AUMID/title, not a title-only target. Paper used the prebuilt -REX host and a fresh bundle carrying the serving instance/manifest identity; -a Paper native build remains **Unsupported**. - -All live commands used the granted W1 campaign and -`with-native-lock.ps1 -Lane W1 -RequirePhysical`. Exact reproduction, -session/lock receipts, native identities, run IDs and per-mode results are in -`artifacts\desktop\f11\w1-native`; physical roots are -`physical-windows-UmfSj6` and `physical-win32-sssEyg`. -The owning harness reference -`packages/agentic/desktop-test/references/current-status.md` records its -commands and the historical Ctrl+C regression proof separately. - -Workflow `43284867711` reached the Windows native consumer in jobs -`131415841258` and `131415842551` after the built-import correction. -Both ran three successful ownership modes, including two accessibility specs, -but their lost-response case failed on an exact-message assertion before -WDIO. Source `b1bbb26d` repairs that test composition in desktop-test and -passes it natively; it does not change the client's legitimate transport-error -wrapping. Corrected-head hosted coverage is pending publication here. - -The campaign is ended. Broad audit exit 2 names only four foreign portable -W5 Jest/Yarn processes with no listeners; zero W1 native/GUI processes remain, -and no foreign process was stopped. Native helper builds, JavaScript-only -checks and capability skips are not substituted for the live results above. - -**Portable closeout:** three fresh locked default-environment -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` attempts -**Failed**, each only on the unchanged Storybook integration test -`hosts the Storybook channel and Stage 1 smoke-test target in one process` -with `A stopping command cannot publish readiness.`. The test and readiness -source match main and review base `04d8d8b13`; no unrelated source, -timeout, TEMP/Git-ceiling or concurrency workaround was added. The owning -package's full default-concurrency `yarn workspace -@fluentui-react-native/storybook-desktop test` **Passed** 340 with three -skips. This is not a passing aggregate receipt for the current source. -Details and the unchanged-source object IDs are in the harness ledger and -`artifacts\desktop\f11\w1-native\aggregate-readiness-source-comparison.json`. -The final attempt began with a clean audit; the failure is not attributed -to foreign work. Owned fail-fast portable workers exited naturally; the -foreign owner removed its abandoned processes. Final audit **Passed**: -`FURN-AUDIT: no FURN desktop processes found`. Ledger format, instruction -and link checks passed. - -## Windows default-aggregate follow-up — 2026-11-12 UTC - -**Observed at `2bf750f9e`:** final locked default -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` **Passed** -178 tasks, zero failed, at 20:52:21–20:55:29 UTC. No source, deadline, -TEMP/Git-ceiling or concurrency change was made. A detached main baseline -first failed the driver cancellation/reuse test (`unique` expected, -`inconclusive` received); its unchanged default repeat passed all 178. -That different failure does not reproduce the earlier Storybook readiness -error. Exact causality of that error remains unproven. - -Native-store inspection found no live descriptors/claims. Only the two -campaign-owned bridge epochs were compare-and-removed at 20:46:53 UTC, -after exact recorded application/start-time matching and verified exit -proof; eleven unrelated recovery records were retained. No process, -registered app, firewall or loopback setting was changed. The scratch -worktree was removed, and final broad process audit **Passed**, no -FURN processes found. The owning harness receipt in -`packages/agentic/desktop-test/references/current-status.md` records state -locations/timestamps, both baseline results, the head receipt and the -inference boundary. - -Corrected-source hosted workflow `43297198219` also **Passed**, zero -failed jobs. Windows `131458546719` and Win32 accessible `131458546805` -each proved four ownership cases passed / one capture-policy skip / -zero failures; handoff, preemption and lost-response each executed both -inner specs with `accessibilityClick: passed`. Those hosted results and -the preceding physical campaign remain distinct from this portable -aggregate follow-up. - -## F11 rebased macOS positive qualification — 2026-11-12 UTC - -**Observed at `5dbadb53b10cd81ccf1e15739e213f0046d93323`:** the original -`11820d6330f4ea591bed400193b13e793970b5bd` branch was rebased onto F6/main -`041ac8b9894aeaa7e1aa273baa2ece4e46298b94`, preserving every Windows addition -and both provenance/handoff ledgers. No runtime/native-provider fix was needed. - -Under the shared native lock, stable-signed noninteractive doctor reported -Accessibility, PostEvent and ScreenCapture **true**, no prompt, connected -unlocked console. The actual AppKit private-descriptor MCP/WDIO contract -**passed all five modes, zero skips**: handoff, operator preemption, worker -crash, lost successful grant response, and target capture. All four non-crash -modes executed/passed both inner ownership/accessibility-input specs without -skips. The deliberately failed crash worker was drain-confirmed before reuse. -Every mode preserved its exact external fixture through service shutdown, -withdrew both descriptors, released its claim, then ended only its own fixture. -The earlier macOS permission block is historical. - -The locked native `serviceIdentity.contract applicationIncarnation.contract -nativeDriver --runInBand` selection passed **28 / 15 Windows skips** (including -the portable built-ESM regression). Portable full driver tests passed -**1,235 / 66 skips**; format, lint and build passed. Exact commands, immutable -helper/fixture identities, per-mode run IDs, cleanup and evidence locators are -in `packages/agentic/desktop-test/references/current-status.md`, under -"F11 rebased macOS positive qualification — 2026-11-12 UTC". -This qualifies generic attach-only handoff, not RN rendering, Storybook/F9, -F10 retained launches, physical input or replay. Windows evidence above is -preserved historical evidence, not a rerun of the rewritten source. -Plain locked aggregate acceptance remains **failed**: the unchanged main -`boundedCommand.test.ts` expected `SIGKILL`, received `SIGTERM` on both -post-documentation-correction attempts. The first also timed out in private -storage; that passed on repeat. Focused diagnostics passed 73 / 10 skips, -but do not replace aggregate acceptance. No environment/concurrency or -unrelated source workaround was used; see the harness receipt for every result. +# Current status + +This document describes the capabilities, limits and native qualification state of the application-neutral desktop driver. +Done means the stated implementation exists and is covered by deterministic tests; it does not mean every native endpoint is qualified. +Parent issue: [#4435](https://github.com/microsoft/fluentui-react-native/issues/4435), which records dependencies and parallel waves for the follow-up work. +See [architecture](architecture.md) for boundaries and [remaining work](remaining-work.md) for the delivery-gate inventory cited by [#4345](https://github.com/microsoft/fluentui-react-native/issues/4345) and [#4348](https://github.com/microsoft/fluentui-react-native/issues/4348). +The approved direction is a stronger `desktop-test` workload host with native authority here and MCP as a client; no coordinator package or daemon is added, and management and run schemas stay in the harness. + +## Status overview + +The driver, its W3C service, typed clients and agent operations are implemented and covered by deterministic tests. +All four endpoints (macOS Fabric, macOS Paper, Windows Fabric, Win32 Paper) have bounded physical-input Storybook smoke passes at the revisions listed under Native qualification. +Failure-path, descendant, reliability and release-matrix qualification remain open on every endpoint. + +- **Service and clients:** Done. The W3C service and typed and WDIO clients accept registered targets only. See [Service and clients](#service-and-clients). Tracking: [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450). + +- **Sessions and input:** Qualified (macOS Fabric and Paper signed ownership contracts; physical-input smokes on all four endpoints). One reservation per target, exact application exclusion and physical-input postconditions are implemented; failure-path qualification is open. See [Sessions and input](#sessions-and-input). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441), [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349). + +- **Discovery, grants and authority:** Done. Verified private discovery and attenuated connection principals are covered by deterministic contracts. See [Discovery, grants and authority](#discovery-grants-and-authority). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Mutation settlement and recovery:** Done. Native mutations settle before advancing and uncertain changes quarantine. See [Mutation settlement and recovery](#mutation-settlement-and-recovery). Tracking: [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451). + +- **Observation and actions:** Done. Bounded observation, find, action, check, sequence and named accessibility actions. See [Observation and actions](#observation-and-actions). + +- **Capture and artifacts:** Done. Artifacts are bounded and session-bound; Windows and Win32 capture qualification is open. See [Capture and artifacts](#capture-and-artifacts). Tracking: [#4348](https://github.com/microsoft/fluentui-react-native/issues/4348). + +- **Storybook story control:** In progress. Catalog, select and reset are an experimental integration; native cancellation and fault injection are unqualified. See [Storybook story control](#storybook-story-control). Tracking: [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Crash containment and continuity:** In progress. The ownership ledger and reclaim live in desktop-test, and application exit reports relaunch-required without restoring state. See [Crash containment and continuity](#crash-containment-and-continuity). Tracking: [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451), [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Recording, replay and visual comparison:** Deferred. There is no recording export, event tail or visual comparison. See [Recording, replay and visual comparison](#recording-replay-and-visual-comparison). + +- **Platform scope:** Unsupported for Intel macOS and distributed or remote control. See [Platform scope](#platform-scope). Tracking: [#4345](https://github.com/microsoft/fluentui-react-native/issues/4345). + +## Service and clients + +### Behavior and limits + +- The W3C service and the typed and WDIO clients accept registered targets only. +- Browser-inapplicable native commands are explicitly unsupported. +- The driver has no Storybook, React or React Native dependency; package-boundary tests reject forbidden imports and dependencies. +- The optional story provider is injected. + +### Open work + +- [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450): management facet and transport. + +## Sessions and input + +### Behavior and limits + +- One reservation exists per registered target in one service. +- Each session's commands and all physical input are serialized, including across helper processes. +- Exact macOS application exclusion and the physical-input mutex passed signed AppKit contracts. +- Windows behavior is unchanged by the macOS ownership work. +- Failed-cleanup claims stay fenced; there is no automatic reclaim. +- Packaged Windows launches are `uncontained` and fenced; packaged activation receives no cleanup authority over another controller. +- The Windows helper's Job Object (suspended-create) ends an unpackaged application it launched when the helper dies. +- Windows ownership retains canonical image, native creation time, session and packaged identity; a changed start, session, image or ambiguous match fails closed, and an unreadable native start stays fenced. +- Native recovery settles a different recorded start without control or termination access. +- A Win32 native-app build is unsupported. + +### macOS physical input + +- Physical input verifies postconditions (pointer acknowledgment, keyboard clear and focus) and refuses when the target is obscured; it does not click covered content. +- The physical lane clamps its window to the visible frame (maximum 1024 x 768 with a one-point inset) and verifies the exact requested rectangle. +- A genuinely overlapped 480 x 302 window refuses physical input safely. +- A locked console blocks physical-input lanes; no unlock or privacy change is attempted. +- Keyboard clear confirms focus, selects the whole value with document-navigation chords once, posts Delete once and verifies an empty value; unreadable values refuse before any key, and there is no key replay or semantic fallback. +- Pointer motion is acknowledged for a bounded wait (about 200 ms) before any button press. +- Foreground activation checks its postcondition and requests workspace activation once when the acknowledged write did not establish it. +- Only hits proven to belong to another process get a distinct window identity; same-process covering windows and unproven owners refuse as `obscured`, effect `none`, with no activation, because raising through a popup could dismiss it. +- The physical Fabric lane needs the rnmacos native-props backport documented in `@fluentui-react-native/native-core/references/rnmacos-patch.md`. +- Exact macOS identity uses kernel start timestamps, causal direct-child or LaunchServices capture and exact-primary cleanup; both launch kinds are `uncontained`, and unknown or stale identity never authorizes termination. +- Primary exit is not descendant containment. +- Strict clear requires an explicit mechanism: invalid mechanisms and missing elements are refused with no effect. +- Doctor `ready` describes artifact availability, not permission readiness. +- Hosted macOS physical input runs in an advisory CI lane ([#4500](https://github.com/microsoft/fluentui-react-native/pull/4500)); hosted skips are not passes. + +### Windows secondary windows and activation + +- Windows secondary and popup windows register as distinct targets with bounded discovery and recovery; their handles retire when closed. +- Modern Pressable and Button controls refuse semantic activation honestly with `activation-unsupported`, effect `none`, without a physical fallback. +- Legitimate Invoke and Toggle activation works. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): live requalification of renderer and failure paths across the four endpoints, including checkbox and Accordion physical activation and Paper keyboard clear. + +- [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349): required native CI, including managed macOS physical-input lanes. + +## Discovery, grants and authority + +### Behavior and limits + +- Native identity is verified before any credential is transmitted. +- All protected driver routes authenticate. +- Storybook agent manifests additionally disable alternate channel mutations in the integration adapter. +- Driver, workload and recovery scopes are distinct. +- Delegation is fixed and driver-only over HTTP; child grants only attenuate targets and scopes. +- Revocation fences the subtree synchronously and drain is awaitable and retryable. +- Runner handoff grants are attenuated runner or controller grants with bounded busy by default and operator-authorized preemption. +- A target reservation is granted only after in-flight acquisition, session deletion and controller-subtree drain settle; a revocation callback alone is not a drain barrier. +- Failed cleanup stays fenced and retryable; unknown runner input is rejected before any grant or reservation. +- Settled drains restore the 4096 live and draining grant capacity; only the last 4096 settled cleanup proofs are retained, never pending or failed proof. +- Shared descriptor or workload holders must relinquish; they are not preempted as a subtree. +- Storybook integration, retained launches and cross-connection resume or replay are not implemented by handoff ([#4491](https://github.com/microsoft/fluentui-react-native/pull/4491)). +- Descriptor-path native requalification has not been run for the attenuation work. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): descriptor-path requalification. + +## Mutation settlement and recovery + +### Behavior and limits + +- Native mutations settle before the queue advances. +- Uncertain story changes quarantine instead of assuming a later render is safe. +- Generic and Storybook native paths share the harness factory's private settlement storage and exact-incarnation probes. +- Direct library embedders must pass `recovery`; a restart is not proof. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): integrated survivor, interrupted-publication and recovery parity for generic native applications. + +## Observation and actions + +### Behavior and limits + +- Observation enforces provider capability and budget checks and reports completeness and truncation explicitly. +- Reference generation is session-local; reopen after application replacement. +- A sequence has up to eight steps; request-id outcomes prevent replay within their retained session history. +- Nothing replays across service replacement. +- Named accessibility actions invoke one uniquely named action through agent operations and WDIO (`desktopPerformAccessibilityAction`). +- The W3C named-action route is fenced against stale Storybook preview context like click, clear and value. + +### Open work + +- None tracked beyond the endpoint qualification in [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +## Capture and artifacts + +### Behavior and limits + +- Artifacts are session-bound, bounded and hashed; pixels are opt-in. +- Screenshots are evidence, not visual comparison. +- Hosted Windows policy disables physical input and can reject capture with `0x80070057`; these skips stay explicit and are not passes. +- Hosted physical-input and capture skips are not passes; Windows doctor `--permissions` is unsupported. +- Win32 capture of a minimized window returns `unsupported operation` without restoring it. +- A successful `activate` does not qualify a refused physical click. + +### Open work + +- [#4348](https://github.com/microsoft/fluentui-react-native/issues/4348): Windows and Win32 capture qualification. + +## Storybook story control + +### Behavior and limits + +- Catalog, select and reset are an experimental integration. +- Authenticated epochs persist per native-verified application incarnation, with versioned marker conformance, prompt disconnect failure and stale preview rejection. +- Unknown navigation effects quarantine. +- Story selection can return allowlisted, service-local timing observations and a native-verified PID and start-time pair; they are diagnostics, not marker fields, context identity or authority to clear quarantine ([#4496](https://github.com/microsoft/fluentui-react-native/pull/4496)). +- A zero advisory timing count is not visual qualification. +- Story args are unsupported on the protected driver (`story-args-unsupported`), including for the session-owning grant and the controlled orchestrator. +- HTTP and foreign WebSocket args are refused or not relayed, without logging values. +- Controlled HTTP mutations return 403, and foreign WebSocket selection, prepare and args never reach the runtime. +- Authenticated driver prepare and select use settlement and fencing. +- The non-agent channel remains legacy and unprotected; non-agent smoke is not driver evidence. +- Preview context checks preserve chrome and secondary references. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): native cancellation and fault injection, fresh preview after lost context, secondary windows. + +## Crash containment and continuity + +### Behavior and limits + +- The exact ownership ledger, surviving-resource reconciliation and lifelines belong to desktop-test. +- Windows hard-terminates non-detached children when the owner crashes, so cooperative crash-disposal contracts are skipped there. +- After a full-service restart, a private authenticated-epoch record plus exact live-incarnation proof yields renderer-specific relaunch-required with effect none. +- Initial absence and exited or replaced applications stay bridge-unavailable. +- Failed or unknown probes are contract-tested only. +- There is no transparent restoration or quarantine recovery. +- The inspected RN Storybook 10.7.1 and Storybook 10.6.5 transport does not reconnect. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): supervisor crash, descendants and interrupted publication on live endpoints. + +- [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451): repeated-run reliability window. + +## Recording, replay and visual comparison + +### Behavior and limits + +- There is no agent-session-to-WDIO export. +- The window-events route returns its full bounded ring; there is no event-tail API. + +## Platform scope + +### Behavior and limits + +- The macOS helper is arm64; Intel macOS is unsupported. +- The service is loopback-only; distributed and remote control are unsupported. +- macOS requires Accessibility and PostEvent (and ScreenCapture for capture) permissions; the stable-signed helper reports them without prompting. + +## Native qualification + +Results are per endpoint and tied to the revision that produced them; skipped checks are not passes. +Full per-run evidence lives in the linked pull request or issue. +The parameterized procedure is `@fluentui-react-native/desktop-test/references/native-qualification.md`. +Live requalification is tracked in [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441) and the repeated-run support matrix in [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451). + +### macOS Fabric + +- **Story control, restart:** bounded marker, stale-reference and restart checks passed at `6fca14788`; M3 at `41fc7473f`. + +- **Neutral ownership, inspection, exclusion:** signed checks and 224 accessible smoke checks passed. + +- **Physical input at `b66140d7b`:** `DESKTOP_INPUT=physical yarn storybook smoke --macos --mode stories-and-tests` passed 343, 3 declared unsupported, 0 failed locally, and the advisory hosted physical job passed the same selection ([#4500](https://github.com/microsoft/fluentui-react-native/pull/4500)). + +- **Earlier gap closed:** the lifecycle run at `141b88b4b` had the Checkbox physical click refused as obscured and Accordion focus-clicks failing; [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500) fixed them ([#4493](https://github.com/microsoft/fluentui-react-native/pull/4493) holds the failing run). + +- **Exact application ownership:** signed AppKit ownership, inspection, exclusion and physical-input mutex contracts passed ([#4492](https://github.com/microsoft/fluentui-react-native/pull/4492)); permission-enabled supporting evidence in [#4493](https://github.com/microsoft/fluentui-react-native/pull/4493). + +- **Handoff:** native selection at `87bc8b710` passed 27 identity and backend cases, but the real-descriptor consumer failed because Accessibility and PostEvent were false then; native handoff and worker-crash drain are unqualified ([#4491](https://github.com/microsoft/fluentui-react-native/pull/4491)). + +- **Still required:** native cancellation and fault injection, fresh preview after lost context, secondary windows, supervisor crash, descendants, interrupted publication and the reliability window ([#4441](https://github.com/microsoft/fluentui-react-native/issues/4441), [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451)). + +### macOS Paper + +- **Story control, restart:** hosted build and smoke, and bounded checks at `6fca14788`. + +- **Neutral ownership, inspection, exclusion:** separately signed checks and 224 accessible smoke checks passed. + +- **Physical input at `b66140d7b`:** physical smoke passed 343, 3 declared unsupported, 0 failed on rerun after an unrelated Captive Network Assistant window obscured six cases (the first attempt failed 7); a focused smoke passed 225/225; the Paper hosted physical lane does not exist ([#4500](https://github.com/microsoft/fluentui-react-native/pull/4500)). + +- **Keyboard clear:** fixed by [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500) after [#4493](https://github.com/microsoft/fluentui-react-native/pull/4493) found the value stayed nonempty; semantic `activate` remains Unsupported by the Paper provider. + +- **Still required:** cancellation and fault injection, fresh preview after lost context, secondary windows, and failure and reliability qualification ([#4441](https://github.com/microsoft/fluentui-react-native/issues/4441)). + +### Windows Fabric + +- **Observation and story MCP path:** M3 real-app observation recorded at `600235c35`. + +- **Connected-console physical smokes at `fec94cec4`:** Checkbox and Callout physical-input Storybook smokes passed 203/203 with zero failures, skips or retries; the real MCP descriptor path passed physical checkbox clicks with hash-verified PNGs ([#4469](https://github.com/microsoft/fluentui-react-native/pull/4469)). + +- **Secondary windows and activation at `6c626602c`** (remote unlocked RDP session, not console): a distinct secondary window's physical-pointer click was applied with `focus: confirmed` and its handle retired; modern main controls refuse semantic activation with `activation-unsupported`, effect `none` ([#4498](https://github.com/microsoft/fluentui-react-native/pull/4498)). + +- **Application ownership at `da89fe9bc`:** 11/11 actual-endpoint consumer cases and 6/6 provider cases passed locally; hosted selection ran 0 of the 11 consumer cases ([#4494](https://github.com/microsoft/fluentui-react-native/pull/4494)). + +- **Handoff at `b1bbb26df`:** the generic runner contract against the plain Win32 fixture passed 5/5 with zero skips (handoff, preemption, worker crash, lost grant response, capture), and the identity and backend selection passed 24 with 19 macOS-only cases skipped ([#4491](https://github.com/microsoft/fluentui-react-native/pull/4491)). + +- **Hosted CI:** Windows Storybook PR smoke 255/255 and native desktop-driver 1,254 tests passed at `86be8531e` ([#4498](https://github.com/microsoft/fluentui-react-native/pull/4498)). + +- **Still required:** 200% and mixed DPI, negative-origin and minimized windows, live lock, RDP and elevation transitions, popups other than Callout, capture cancellation, kernel PID recycling, arbitrary descendants, and a console rerun of the interactive wheel, keyboard and capture contracts ([#4441](https://github.com/microsoft/fluentui-react-native/issues/4441), [#4348](https://github.com/microsoft/fluentui-react-native/issues/4348)). + +### Win32 Paper + +- **Agent path and native fixture:** M3 and agent path recorded; native fixture physical input and capture at 100% and 150%. + +- **Connected-console physical smokes at `fec94cec4`:** Checkbox and Callout physical-input Storybook smokes passed 177/177; physical checkbox clicks through the MCP descriptor passed ([#4469](https://github.com/microsoft/fluentui-react-native/pull/4469)). + +- **Secondary windows and fresh-preview fencing at `6c626602c`:** the exact owned tool secondary registered with exact-root UIA content and a physical-pointer click was applied (native delivery only, not proof of the RN focus command); injected fresh-preview fencing returned `ref-region-changed` and `story-context-stale` with effect `none` (provider retirement, not crash or reconnect proof) ([#4498](https://github.com/microsoft/fluentui-react-native/pull/4498)). + +- **Real lost context:** after an authenticated-bridge disconnect the fresh preview reference is not exposed (Partial), and the exact REX process outlived the closed window after MCP close, reporting failed cleanup. + +- **Application ownership at `da89fe9bc`:** 11/11 consumer and 6/6 provider cases passed locally ([#4494](https://github.com/microsoft/fluentui-react-native/pull/4494)). + +- **Handoff at `b1bbb26df`:** the same generic plain-Win32-fixture runner contract passed ([#4491](https://github.com/microsoft/fluentui-react-native/pull/4491)). + +- **Hosted CI at `86be8531e`:** accessible 233/233; physical 288/289 with 1 unsupported; legacy E2E passed ([#4498](https://github.com/microsoft/fluentui-react-native/pull/4498)). + +- **Still required:** mixed DPI, production-bundle Metro reload (unsupported), failure paths, and a console rerun of the current native contracts ([#4441](https://github.com/microsoft/fluentui-react-native/issues/4441)). Hosted input and capture skips are not passes. + +### Hosted CI and portable evidence + +- **Windows oversized-tree fallback:** compiled and contract-tested on hosted CI only; no live Windows qualification. + +- **Generic native applications:** historical macOS Calculator and plain Win32 evidence; a probe alone is not real-app lifecycle qualification. + +## Measurements + +Measured on macOS Fabric Storybook at `41fc7473f`, 2720 x 1988-pixel window (1360 x 994 points, 2x), via `yarn storybook driver --macos --agent` and an MCP SDK stdio client over three marker-confirmed stories (button, checkbox, menu). +Token counts use `o200k_base`, include every returned text block and the result envelope (about 150 tokens for observation and delta payloads), and exclude MCP discovery (see `@fluentui-react-native/desktop-driver-mcp/references/current-status.md`). +All observations were complete. Windows and Paper need their own rows; do not extrapolate. + +| Result | Measured text tokens | Scope and notes | +| ------------------------------ | --------------------- | ------------------------------------------------------ | +| `furn/tree`, pretty JSON | 12,972-13,643 | 63-67 KB; 142-150 nodes | +| Same tree, minified | 7,102-7,469 | Same nodes | +| Page source XML | 6,244-6,562 | Same app state | +| `furn/window-events` | 5,331-5,352 | Entire bounded ring, not a recent tail | +| `observe`, `all` outline | 1,602-1,859 | 86-88% below pretty JSON | +| `observe`, `all` JSON nodes | 3,290-3,375 | Explicit JSON projection | +| `observe`, default interactive | 1,122-1,304 | 74-76 elements; not the same scope as all nodes | +| `observe`, content | 1,199-1,373 | 79-80 elements | +| `observe`, story-root scope | 252-311 | 2-6 elements | +| Unchanged delta | 149-150 | Retained-sample comparison | +| One-element delta | 192 | Checkbox toggled | +| Find story root / button role | about 200 / about 490 | Unique / multiple matches | +| Check | about 81 | Small predicate result | +| Capture without pixels | about 140 | Metadata, hash, resource link | +| Journal, default 50-entry page | about 3,570 | Bounded page | +| Stories, first 25 | 637 | Full 209-story catalog in three 100-entry pages: 5,429 | + +Latency through the MCP client at `41fc7473f` (one run unless noted; a historical measurement, not a performance guarantee): + +| Operation | Median | Qualification | +| ------------------------------ | -------- | ---------------------------------- | +| Default observe | 29-31 ms | 20 warm calls per story; max 35 ms | +| Delta | 32-35 ms | 20 warm calls per story; max 37 ms | +| Physical click + checked check | 181 ms | One run, not a distribution | +| Named Toggle + check + observe | 48 ms | One run; 386 text tokens | + +Helper-side journal timings at `c84acb741` (61 recorded macOS failure-evidence sets, different app states; do not subtract from end-to-end timings): snapshot 0-2 ms, attach 2 ms (max 983 ms), probe 5 ms, perform actions 19 ms, tree 58 ms, find 65 ms (p90 498 ms), source 79 ms, physical click 121 ms, capture 152 ms. +Recorded median full tree at `c84acb741`: 157 nodes, 69 interactive roles, 12 testIDs, depth three; pretty 14.2k tokens, minified 7.8k, XML source 6.8k, window events 5.0k (512 + 128-entry ring). +Earlier synthetic compact-outline and screenshot-token figures were estimates, not measurements, and are not retained. +The historical report kept no portable raw-artifact locator or tokenizer package version. + +When recording a measurement, state commit, OS and architecture, renderer, story, window size and scale, helper build and signing mode, cold or warm state, sample count, tokenizer version, projection and budgets, completeness, discovery overhead and the client-versus-helper timing boundary; report p50, p90 and max with retries, failures and skips. + +## Maintaining this document + +- Edit only the topic your change affects, and keep that topic's line in the status overview in step with it. +- Describe what is true now. Do not add dated entries, run logs, or "updated" stamps. +- Record per-run evidence (commands, counts, run ids, artifact locations, cleanup outcome) in the pull request description or the tracking issue, and keep only the latest result for each endpoint here with a link to that record. +- Put remaining-work detail in the tracking issue and link the issue from the topic's open-work list; remove the link when the issue closes. +- Write one sentence or fact per line, and keep a blank line between adjacent list entries that different changes are likely to edit, so concurrent changes merge independently. +- Historical results stay tied to the revision that produced them and are never reported as qualifying a later revision. +- Skipped checks are not passes; report each endpoint (macOS Fabric, macOS Paper, Windows Fabric, Win32 Paper) separately. +- The detailed per-run ledger that preceded this format is preserved in git history: `git show a7af2b4b2:packages/agentic/desktop-driver/references/current-status.md`. diff --git a/packages/agentic/desktop-driver/references/native-helpers.md b/packages/agentic/desktop-driver/references/native-helpers.md index 893566ca19..ff0c9bec7c 100644 --- a/packages/agentic/desktop-driver/references/native-helpers.md +++ b/packages/agentic/desktop-driver/references/native-helpers.md @@ -304,7 +304,7 @@ a popup stays listable using native identity and ownership rather than another UIA read. New candidates share one 1,000 ms discovery budget per listing and are retried later when that budget is exhausted. This does not broaden global title discovery. See the [Windows identity rules](../native/windows/README.md#window-events-and-identity) -and [Windows qualification evidence](current-status.md). +and the Windows entries in [native qualification](current-status.md#native-qualification). ## Diagnostics diff --git a/packages/agentic/desktop-driver/references/remaining-work.md b/packages/agentic/desktop-driver/references/remaining-work.md index a7793bf8df..837be08ddf 100644 --- a/packages/agentic/desktop-driver/references/remaining-work.md +++ b/packages/agentic/desktop-driver/references/remaining-work.md @@ -46,8 +46,9 @@ portable safety fixes may land before complete native qualification. Full submission still requires the shared architecture and separate macOS, Windows Fabric, and Win32 Paper evidence for the changed surfaces. Do not describe a macOS-only or capability-skipped result as cross-platform parity. -The [current-status ledger](current-status.md) owns capability/measurement -summaries; the sections below retain outstanding delivery work and prior evidence. +The [current status](current-status.md) owns the capability summary and the +latest result per endpoint; the sections below retain outstanding delivery work +and prior evidence. This is the canonical open-work plan under R-PKGDOC-009. Each milestone states its outcome and completion condition. As implementation lands, move durable @@ -58,8 +59,23 @@ contracts into the owning references and remove completed work here. ### macOS activation and popup polling Part of #4435. The 2026-11-12 refresh-reduction work deliberately leaves -activation semantics unchanged; the [status ledger](current-status.md) -records the source trace and bounded observational evidence. +activation semantics unchanged. Its bounded timing observations are in the +"Story selection and refresh reduction" and "Visual flash claims" sections of +`@fluentui-react-native/storybook-desktop/references/current-status.md`. + +Source trace: the W3C click path activates before its server-side hit test. +The operation click path conditionally raises an externally obscured target and +rechecks; in-window obstruction remains a no-effect refusal. The native +physical-click provider independently activates under its physical-input scope, +rereads the target, and verifies ownership of the click point. These are +separate postconditions, not interchangeable duplicate calls: removing the +server activation can move hit-testing before foreground establishment, and +removing the provider activation assumes ownership cannot change between +commands and weakens callers that use the provider independently. macOS +activation sets main/raise attributes even when already frontmost, and the menu +test helper `inWindowWith` calls activating `switchToWindow` while polling each +candidate window and restoring its owner. Both are plausible flashing +contributors, not established visual causes. Measure activating menu/popup polling before changing it. Trial explicitly non-activating, window-scoped reads rather than silently changing @@ -1114,7 +1130,8 @@ secondary physical delivery passed. The complete native selector passed the shared discovery deadline and deferred retry, truthful publication, and the deterministic close-after-enumeration self-test. The default-environment clean-cache aggregate passed all 178 tasks on its first -attempt. See the separate [Windows evidence entry](current-status.md). +attempt. See the Windows Fabric and Win32 Paper entries in +[native qualification](current-status.md#native-qualification). Keep #4441 open for the unrun 200% DPI, mixed-DPI, and negative-origin cases and other issue-owned gaps. Native delivery receipts do not prove the diff --git a/packages/agentic/desktop-test/README.md b/packages/agentic/desktop-test/README.md index 9ccf5ec432..4431738030 100644 --- a/packages/agentic/desktop-test/README.md +++ b/packages/agentic/desktop-test/README.md @@ -11,7 +11,7 @@ integration in `@fluentui-react-native/storybook-desktop`. **Work in progress:** [current status](references/current-status.md) separates implemented capabilities from incomplete ownership/recovery guarantees and -records consumer qualification evidence. +links the latest per-endpoint qualification results. ## Quick start diff --git a/packages/agentic/desktop-test/references/README.md b/packages/agentic/desktop-test/references/README.md index 5de1507f3f..af93954c88 100644 --- a/packages/agentic/desktop-test/references/README.md +++ b/packages/agentic/desktop-test/references/README.md @@ -7,14 +7,14 @@ owner_paths: # Desktop Test references -| ID | Document | Read when | -| ----------------------------------- | --------------------------------------------------------------- | -------------------------------------------------------------------------------- | -| `desktop-test.configuration` | [Configuration](configuration.md) | Adding an application target or suite | -| `desktop-test.commands` | [Commands](commands.md) | Running stages, suites, or service mode | -| `desktop-test.platforms` | [Platform setup](platforms.md) | Preparing or building a platform-specific app | -| `desktop-test.wdio-suites` | [WDIO suites](wdio-suites.md) | Configuring suites or native page adapters | -| `desktop-test.lifecycle` | [Lifecycle](lifecycle.md) | Changing ownership, supervision, or cancellation | -| `desktop-test.reporting` | [Reporting](reporting.md) | Reading run results, artifacts, or CI output | -| `desktop-test.contributing` | [Contributing](contributing.md) | Editing the harness or its tests | -| `desktop-test.current-status` | [Current status](current-status.md) | Assessing lifecycle readiness, remaining gates, and historical consumer evidence | -| `desktop-test.native-qualification` | [Native descriptor-path qualification](native-qualification.md) | Running renderer-specific native and descriptor-path qualification | +| ID | Document | Read when | +| ----------------------------------- | --------------------------------------------------------------- | ------------------------------------------------------------------------ | +| `desktop-test.configuration` | [Configuration](configuration.md) | Adding an application target or suite | +| `desktop-test.commands` | [Commands](commands.md) | Running stages, suites, or service mode | +| `desktop-test.platforms` | [Platform setup](platforms.md) | Preparing or building a platform-specific app | +| `desktop-test.wdio-suites` | [WDIO suites](wdio-suites.md) | Configuring suites or native page adapters | +| `desktop-test.lifecycle` | [Lifecycle](lifecycle.md) | Changing ownership, supervision, or cancellation | +| `desktop-test.reporting` | [Reporting](reporting.md) | Reading run results, artifacts, or CI output | +| `desktop-test.contributing` | [Contributing](contributing.md) | Editing the harness or its tests | +| `desktop-test.current-status` | [Current status](current-status.md) | Assessing lifecycle readiness, per-endpoint qualification, and open work | +| `desktop-test.native-qualification` | [Native descriptor-path qualification](native-qualification.md) | Running renderer-specific native and descriptor-path qualification | diff --git a/packages/agentic/desktop-test/references/current-status.md b/packages/agentic/desktop-test/references/current-status.md index a642784634..f79cfae035 100644 --- a/packages/agentic/desktop-test/references/current-status.md +++ b/packages/agentic/desktop-test/references/current-status.md @@ -5,3632 +5,367 @@ owner_paths: - packages/agentic/desktop-test --- -# Current lifecycle and qualification status - -**Work in progress, updated 2026-11-12.** Done means the listed implementation, -not full cross-platform qualification. This package is the intended neutral -workload host; the existing Storybook lifecycle takeover is transitional. -No new coordinator package, daemon, dynamic plugin loader, or package rename -is required now. Follow-up work is tracked in -[desktop driver improvements (#4435)](https://github.com/microsoft/fluentui-react-native/issues/4435) -and its sub-issues. - -## Capability matrix - -| Capability | State | Limit or next gate | -| -------------------------------------------------------- | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| App/suite configuration and standalone installed-app use | Done | Storybook is optional; runtime boundary and standalone-consumer contracts reject React, React Native, and Storybook imports/dependencies. Executable config is trusted operator code, not a sandbox. | -| Native service and helper composition | Done | Generic supervisor and supervised Storybook child use the same real `driverService` factory, helper policy, authority and settled close. | -| Standalone WDIO/Jasmine execution | Done | WDIO owns its session. Storybook CSF callbacks use a runner-owned session borrowed by a Node worker. | -| Required services, reporting, structured results | Done | One collector; bounded JSON/NDJSON summaries. Service-ready is not app/context-ready. | -| `serve --agent` descriptor publication | Done | Protects driver routes; no automatic build/launch from MCP. Normal WDIO cannot use the agent-served path yet. | -| Cleanup failure fence | Done, conservative | Failed or timed-out disposal retains the claim with an exact resource generation and unresolved effect; manual inspection is required. No automatic reclaim. | -| Complete lifecycle settlement | Done, conservative | Guarded work settles before successful close; late acquisitions are disposed; required-service failures stop dependants; owned resources dispose in reverse order. | -| Typed generic resource contributions | Done | Static contribution plans validate IDs/dependencies and record readiness, generation, identity, ownership, and settled disposal through public `resources` and `ownership` subpaths. | -| Durable quarantine wiring | Done | Both native service paths use owner-only storage and exact-incarnation probes; restart is not recovery proof. Live endpoint qualification remains separate. | -| Supervisor crash containment | Work in progress | Cooperative Node service lifelines and correlated service readiness implemented; portable parent/wrapper crash, late-spawn, failed-withdrawal and stalled-child contracts. Live macOS survivors remain unqualified; no automatic reclaim. Windows: the helper's Job Object ends an unpackaged app it launched when the helper dies (qualified with REX); packaged launches are `uncontained` and survive in a fenced claim. Non-detached children are hard-terminated with a crashed owner (observed A/B), so the three crash-disposal contracts are skipped there. | -| Storybook resource contributions | Work in progress | Current integration substitutes its CLI lifecycle; retain the channel/driver child until relocation is justified and qualified. | -| MCP/WDIO handoff | Work in progress | Generic attach-only phase is implemented and qualified on real macOS AppKit and Windows/Win32 fixtures: bounded busy, operator preemption, settled deletion, exclusive reservation, worker-crash/lost-response drain, fresh resources, and credential confinement. Storybook handoff remains deferred; see the dated F11 receipts below. | -| Supervisor-owned app lifetime | Work in progress | Opt-in macOS supervisor launch, neutral exact attach, causal ownership and application claim resources are implemented. Session-owned launch remains the default; external attach never transfers ownership. Signed AppKit supervisor/session lifetime, exact cleanup and external preservation passed on 2026-11-12. Renderer evidence is scoped below; descendants, interrupted native publication and Storybook lifecycle relocation remain gates. | -| Cross-service/multi-target claims | Work in progress | Configured aliases have distinct registered IDs while retaining physical identity. Exact macOS application exclusion, lease release and the physical-input mutex passed signed AppKit contracts on 2026-11-12; cross-platform and broader lifecycle qualification remains open. | -| Bounded persisted run/status/artifact queries | Done | Harness-local read service/client; atomic snapshots, terminal evidence, scoped expiring grants, bounded pages and allowlisted opaque artifact IDs. No management transport. | -| Workload management/control API | Work in progress | Future management selects only declared workload IDs, not commands/paths/environment; lifecycle/handoff gates still apply. | -| Automatic service reuse and per-user daemon | Deferred | Requires exact identity/config matching and evidence of need. Never adopt an unknown port occupant. | - -See [lifecycle](lifecycle.md), [commands](commands.md), and -`@fluentui-react-native/desktop-driver/references/current-status.md` for the -native and driver boundaries. - -## Historical consumer evidence - -Migrated from the former root capability report. These results were recorded -before this update; they are not reruns or qualification of new cleanup behavior. - -| Evidence window | Consumer/lane | Recorded result | Limit | -| ----------------------------------------- | ---------------------------------------- | ------------------------------------------------------------------------------ | ------------------------------------------------------------- | -| Baseline report, `c84acb741` (2026-11-02) | FluentTester E2E, macOS Fabric | 62/62 locally and hosted | No skips/retries recorded; not new lifecycle evidence | -| Same report | FluentTester E2E, Windows Fabric | 39/39 | Endpoint-specific historical result | -| Same report | FluentTester E2E, Win32 | 201/201 | Endpoint-specific historical result | -| Same report | macOS Fabric Storybook smoke | 280/280, including 81 callbacks | Full smoke was not rerun for the November 4 update | -| Same report, accumulated local runs | Storybook attempts | 1,118 passed, 24 failed, 40 skipped | Mixed runs; not a controlled reliability rate | -| `59dad9515`, 2026-11-04 | Windows Storybook hosted native contract | Recorded 814 passed, 34 skipped | Physical-input/capture skips do not qualify those paths | -| This update | Portable cleanup contract | Injected disposal failure retains the claim and refuses a replacement workload | Fake-host evidence only; Windows live cleanup remains not run | - -The historical passing-attempt median was 52 ms (p90 2.2 s, max 9 s). -Single-test Storybook commands were recorded at about 4-4.6 s wall time; -full bundle/prep/build/launch at about 70-78 s. These are historical mixed-run -observations, not a benchmark of the proposed shared host. The original report -did not retain a portable raw-report locator; do not imply otherwise. - -## Full-submission checklist - -| Area | Remaining evidence/work | -| --------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| macOS | Signed AppKit and separate Fabric/Paper positive lifecycle, inspection, exclusion and exact primary cleanup passed on 2026-11-12. Rebased native contracts also pass interrupted acquisition publication, owner death with native app/known-descendant survivors, retained claim fences and exact operator cleanup. Arbitrary descendants, Metro/channel survivors, automatic recovery and repeated-run reliability remain unqualified. | -| Windows / Win32 | Same-branch MCP restart/revocation/dispose/args tests; physical console Ctrl+C requalification of `serve --agent`; interactive input/capture rerun. Generic launched-application claim fields and durable survivor contracts are implemented; live uncontained packaged-survivor journaling/recovery remains not run for this change. | -| Storybook | Authenticated mutation ingress, context fencing/reconnect, consistent helper build policy. F6 strict provenance and independent nonces are qualified by same-branch Windows Fabric and both Win32 Paper lanes in run 43156299266; see the F6 evidence below for scope and limits. | -| Recovery | Durable uncertainty is wired in both paths; live restart qualification, ledger and exact-resource reconciliation still precede automatic reclaim. Claim schema 3 records the driver's OS start identity; schema 2 and unverified claims inspect as unknown. LIFE-8's clock-estimate mismatch is fixed; LIFE-9's existing stable per-user claim root has a TMPDIR regression contract. Automatic reclaim remains unsupported. | -| Concurrency | No physical-input suites in parallel on one interactive desktop. Separate machines may qualify independent endpoints. | -| Reliability | Establish the required repeated-run window rather than interpreting aggregate past successes as a pass. | - -Tracking: live macOS, Windows, and Win32 requalification -[#4441](https://github.com/microsoft/fluentui-react-native/issues/4441); -recovery [#4436](https://github.com/microsoft/fluentui-react-native/issues/4436) -and [#4442](https://github.com/microsoft/fluentui-react-native/issues/4442); -Storybook [#4437](https://github.com/microsoft/fluentui-react-native/issues/4437), -[#4443](https://github.com/microsoft/fluentui-react-native/issues/4443), and -[#4445](https://github.com/microsoft/fluentui-react-native/issues/4445); -application ownership [#4446](https://github.com/microsoft/fluentui-react-native/issues/4446) -(macOS) and [#4447](https://github.com/microsoft/fluentui-react-native/issues/4447) -(Windows); reliability window -[#4451](https://github.com/microsoft/fluentui-react-native/issues/4451). - -For new evidence, append the date, exact commit, app/OS/renderer, command, -selection, passed/failed/skipped counts, cleanup outcome, and approved artifact -locator. Distinguish tests that actually ran from build/bundle or fake-host -checks. Keep measured text/operation costs in the driver status reference and -MCP discovery overhead in the adapter's status reference. - -### Provisional F15 macOS evidence — 2026-11-06 - -**Provisional; exact source revision `388a60cc803ab78202a6f64f92e3f53b779c5c39`.** -App: Agentic Components Storybook / ReactTestApp (configured bundle ID -`com.microsoft.fluentui.agenticstorybook`). OS: macOS 27.2, arm64. - -| Renderer | Command and actual selection | Result | Cleanup | Artifact | -| -------------- | --------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | -| macOS Fabric | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook smoke --macos --mode stories-and-tests`; 211 selected/executed | Passed 211, failed 0, skipped 0, unsupported 0; retries 0 | No owned app/driver/channel/Metro process; ports 21338/30029/45382 had no listeners; `interactive.lock` absent | `apps/storybook/artifacts/desktop/0f99740f-631b-444a-a705-76a9b4a761b7/run.json` | -| macOS Paper | `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook smoke --macos --paper --mode stories-and-tests`; 211 selected/executed | Passed 211, failed 0, skipped 0, unsupported 0; retries 0 | No owned app/driver/channel/Metro process; ports 21338/30029/45382 had no listeners; `interactive.lock` absent | `apps/storybook/artifacts/desktop/e82a83c2-a01d-41c0-b308-2551630e8b1e/run.json` | -| Windows Fabric | Not run — this machine is macOS-only | Not run; 0 selected | Not applicable | None | -| Win32 Paper | Not run — this machine is macOS-only | Not run; 0 selected | Not applicable | None | - -| Descriptor-path renderer | Command / selection | Passed / failed / skipped / unsupported | Cleanup | Artifact | -| ------------------------ | --------------------------------------------------- | --------------------------------------- | -------------------------------------------- | -------- | -| macOS Fabric | Not run; no real descriptor consumed; selection N/A | 0 / 0 / 0 / 0; Not run | N/A for descriptor path; smoke cleanup above | None | -| macOS Paper | Not run; no real descriptor consumed; selection N/A | 0 / 0 / 0 / 0; Not run | N/A for descriptor path; smoke cleanup above | None | - -The Windows Fabric and Win32 Paper "Not run" rows above describe that macOS -host only. They are superseded, not erased, by the Windows-host section -"Windows Fabric and Win32 Paper qualification — 2026-11-06" at the end of this -file. - -These are real native Storybook smoke runs through the non-agent WDIO path. -They do not demonstrate descriptor publication, service restart, failed -cleanup, or exact attached-app preservation. The `STORYBOOK_SMOKE_STORY` -environment value did not constrain the selection; the artifact reports 211 -callbacks. The initial Fabric attempt failed before selection due to an -unsynchronized CocoaPods sandbox; `yarn storybook prep --macos` repaired the -generated project before the passing retry. See the runbook in -[`native-qualification.md`](native-qualification.md). The authoritative -descriptor-path and F2/F3/F4/F5 cases remain **Not run** until the combined -integration commit is available and the native lock is granted again. - -## F15 combined-candidate macOS qualification — partial, 2026-11-06 - -**Exact tested source:** `3c780cfd1bccaf510ece080e4be2b516c645989a`. -App: Agentic Components Storybook / ReactTestApp, bundle identifier -`com.microsoft.fluentui.agenticstorybook`. OS: macOS 27.2, arm64. -This is a partial ledger for this exact candidate, not final acceptance. The -F5 reference-generation fix at -`2d8e8db001b64ebea85f0d6dce141f040fb5182d` is not in this tested source and -requires a new native qualification. The original accepted-reference probe -remains inconclusive and is not attributed to that fix. - -**Commands and selection:** helper build and permission preflight were -`yarn storybook build-driver --macos` and -`yarn desktop-driver doctor --platform macos --permissions` (no `--prompt`). -The Paper descriptor pass used -`yarn storybook driver --macos --paper --agent` and -`yarn storybook run --macos --paper`; Fabric used -`yarn storybook driver --macos --agent` and -`yarn storybook run --macos`. Both used the MCP SDK stdio client with the -default private `agent-connection.json`; its exact inline client launch command -was not retained. Those app/driver `run.json` records each report zero selected -and zero executed WDIO tests; MCP operations are manual evidence, not suite -counts. The separate Paper legacy lane used -`yarn storybook smoke --macos --paper --mode stories-and-tests`: 319 selected, -308 executed/passed, 0 failed, 11 skipped, 0 unsupported/incomplete/retries/ -flaky. Its artifact is -`apps/storybook/artifacts/desktop/a569510a-ac99-4ec9-8c78-4d8607bdaadd/run.json`. -Two additional 319/308/11 reports -(`8c6d031d-32e3-49c4-9e80-049202b74626` and -`e5117371-764c-45ff-8b14-97c5b7ec5036`) do not retain renderer metadata and -are not credited to Fabric or Paper. - -The 24-case matrix is recorded without promoting partial evidence to passes. -`Not run` means no case-specific native evidence was retained; it is not a -skip or pass. - -| Case | macOS Fabric | macOS Paper | Retained evidence / limitation | -| ------------------------------ | ------------ | ----------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Same-port replacement | Not run | Partial | Paper restart and rediscovery succeeded; explicit confirmation of old-session close was not retained. | -| Grant revocation | Not run | Pass | Paper delegated-grant revocation and child drain were reported complete; private run record only. | -| Dispose during open | Not run | Not run | No acquisition hold was injected. | -| Failed delete | Not run | Not run | No DELETE dispatch-loss injection was run. | -| Ambiguous create | Not run | Not run | No session-create response-loss injection was run. | -| Resource invalidation | Inconclusive | Partial | Fabric had a separate preview-child stale-element rejection (`stale element reference`, `ref-region-changed`), but the earlier accepted handle's kind/identity was not retained. Paper's old-preview rejection lacks retained handle kind/generation. | -| Story quarantine and recovery | Not run | Not run | No unknown-effect transition/recovery injection was run. | -| Protected args refusal | Partial | Pass | Fabric final attempt returned `story-args-unsupported`, unchanged story, and no probe value in error/journal; an earlier recorded post-attempt unchanged-state assertion was false. Paper refusal was reported with no effect. | -| Failed-cleanup claim retention | Not run | Not run | No owned-resource cleanup failure was injected. | -| Launch-mode discovery | Partial | Partial | Only attach mode was discovered; no launch-capable target was exercised. | -| Live cleanup and restart | Not run | Partial | Paper service restart/rediscovery ran; old-session close confirmation was not retained. | -| Attached-app preservation | Pass | Pass | The same ReactTestApp process survived agent-service shutdown; the qualification-launched app was stopped only after that check. | -| F4 post-dispatch cancellation | Not run | Not run | No cancellation-after-native-dispatch injection was run. | -| F4 readiness/render faults | Not run | Not run | No delayed-ready, correlated error, or render-fault injection was run. | -| F4 lost-context fencing | Not run | Not run | No lost-context mutation-fencing case was run. | -| F4 lost-context recovery | Not run | Not run | No new authenticated epoch recovery case was run. | -| F4 secondary windows | Not run | Not run | No secondary-window case was run. | -| F4 physical input | Failed | Not run | Fabric returned `physicalInputApplied:false`; no retry. This is not a pass. | -| F3 startup/shutdown/survivors | Partial | Partial | Normal bounded shutdown and cleanup completed; injected startup/shutdown survivor reconciliation was not run. | -| F5 ingress bypass | Partial | Partial | Fabric anonymous/foreign HTTP select, prepare, and args returned 403 and preview stayed unchanged. The first two foreign-WS checks failed; a later attempt passed with no observer-message delta. Paper bypass checks were reported passed, but detailed status fields were not retained. | -| F5 owned selection/reset | Partial | Partial | Selection/reset markers settled. Stale preview-reference identity and generation evidence is incomplete; see the MCP ledger. | -| F5 channel lifetime | Pass | Pass | Fabric stayed open after 21,003 ms idle; Paper's 21-second check was reported connected. | -| F5 manifest smoke | Not run | Partial | Paper smoke passed 308/319 with 11 skipped; one-session/no-HTTP-fallback evidence was not retained. It is not protected-ingress evidence. | -| F5 legacy separation | Not run | Not run | No separate legacy-channel behavior case was retained. | - -Fabric outcome totals: 2 Pass, 5 Partial, 1 Inconclusive, 1 Failed, 15 Not -run. Paper totals: 4 Pass, 8 Partial, 12 Not run. These are matrix-case -statuses, not WDIO counts. Windows Fabric and Win32 Paper are Not run because -this host is macOS-only. - -The fresh doctor reported Accessibility, PostEvent, and ScreenCapture -preflight `true`, no permission prompt, console on/unlocked, and the arm64 -release helper ad-hoc signed with designated requirement -`cdhash H"0d2e7a81dfd4695b7fb84aa1b3242f5e1773e8b4"`. Host load was 3.25 / -4.41 / 3.81 (1/5/15-minute averages). Per-renderer driver metadata is -`apps/storybook/artifacts/desktop/a58d2fdd-5584-47f4-8747-5307607b4c0d/run.json` -(Paper) and -`apps/storybook/artifacts/desktop/c15a35b9-7b8d-4edb-adba-84e3cd5a3023/run.json` -(Fabric); Fabric app launch is -`apps/storybook/artifacts/desktop/8fb9a576-cde8-470c-a8bb-a76b6d8fabc0/run.json`. -All report zero selected/executed automated work, so they do not increase the -matrix totals. Sanitized operation records and private captures are under -`f15-native-qualification-3c780cfd-20261106T0200/` in the machine-local log -root; they are not committed. - -Retained retries/failures are not erased by later successes: Fabric's first -two foreign-WebSocket assertions failed before the final no-message-delta -attempt passed; an intermediate protected-args attempt recorded -`currentStoryUnchanged:false` before the final attempt recorded `true`; and the -first capture caller shape failed before the corrected call passed. No cause -is inferred from those retries. The Fabric physical-input attempt remained -failed and was not repeated. - -At handoff, ports 21338/30029/45382 had no listeners, the workload lock and -default descriptor were absent, and the Fabric cleanup record showed its -service owner exited, grant revoked, descriptor removed, and attached app -preserved through service shutdown. Paper cleanup was reported, but its -process IDs were not retained, so PID-level Paper process exit is not -independently asserted here. No further native workload was started. No -product implementation was changed in this ledger update. - -## F15 combined-candidate macOS requalification — partial, 2026-11-06 - -**Exact tested source:** `a19fd3e1e7b3f1f443a241b1571cb029162f4ef2`. -App: Agentic Components Storybook / ReactTestApp, -`com.microsoft.fluentui.agenticstorybook`; macOS 27.2 arm64. The exact helper -was rebuilt with `yarn storybook build-driver --macos`; fresh, nonprompting -`yarn desktop-driver doctor --platform macos --permissions` reported -Accessibility, PostEvent, and ScreenCapture as granted. No permission changes -were made. The arm64/release helper was ad-hoc signed with designated -requirement `cdhash H"0d2e7a81dfd4695b7fb84aa1b3242f5e1773e8b4"` and build -fingerprint -`b58e4d8692702fd1dfa16b73ac987d1738e6e17d16757cbc3c6007722b8e3869`. -Observed quiet host load was 2.49 / 5.27 / 5.47 (1/5/15-minute averages). - -The descriptor pass used the default private `agent-connection.json` through -an MCP SDK stdio client. Manual MCP calls have no suite selection; each -companion Storybook driver `run.json` recorded 0 selected and 0 executed tests. -The separate manifest-backed, non-agent smoke command selected and executed -211 callbacks in each renderer. These are legacy-lane results, not -protected-ingress or no-HTTP-fallback evidence. - -| Case | macOS Fabric | macOS Paper | Evidence and artifact | -| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ----------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Same-port replacement | Partial (recorded Failed at qualification; reclassified on contract review): the old client's first post-replacement `desktop_session.status` returned `adapter` / `service-unavailable`, the designed fail-closed rejection of the old incarnation; old session close then succeeded with no mutation replayed. Rediscovery on a later call was not demonstrated, because close is a no-op once invalid authority forgets the session. `desktop_observe.find` returned 0 matches, so no old-reference assertion was possible. | Not run | Fabric private sanitized record: `fabric-same-port-replacement-final.json`; service runs: `apps/storybook/artifacts/desktop/414cf449-bfd4-4b9f-b630-63e86ebdd5c4/run.json` and `apps/storybook/artifacts/desktop/d732ae38-df56-4b81-a56c-2791bdbac5e8/run.json`. | -| Grant revocation | Not run | Not run | No live revocation case on this SHA. | -| Dispose during open | Not run | Not run | No acquisition hold was injected. | -| Failed delete | Not run | Not run | No DELETE dispatch-loss injection was run. | -| Ambiguous create | Not run | Not run | No session-create response-loss injection was run. | -| Resource invalidation | Passed | Passed | Same-session preview-element refs were rejected after marker-settled select/reset; fresh find returned a distinct working alias and the old alias remained stale. Private records: `fabric-reference-sequence-retry.json`, `paper-reference-sequence.json`. | -| Story quarantine and recovery | Not run | Not run | No unknown-effect transition/recovery injection. | -| Protected args refusal | Not run | Not run | Not attempted on this SHA. | -| Failed-cleanup claim retention | Not run | Not run | No forced owned-resource cleanup failure. A normal runner retry was blocked by the live prior claim, retained until its owner exited and empty resources were verified. | -| Launch-mode discovery | Not run | Not run | No launch-capable versus attach-only target comparison on this SHA. | -| Live cleanup and restart | Partial | Not run | Fabric old service stopped and replacement started on the same ports; the old adapter's next status failed with `service-unavailable`. App PID 26792 remained alive through service stop/restart. | -| Attached-app preservation | Passed | Not run | Fabric app PID 26792 remained running after agent-service shutdown; it was stopped only after the check. Paper exact survival through service shutdown was not retained. | -| F4 post-dispatch cancellation | Not run | Not run | No dispatch-confirmed cancellation injection. | -| F4 readiness/render faults | Not run | Not run | No delayed-ready, correlated error, or render-fault injection. | -| F4 lost-context fencing | Not run | Not run | No lost-context mutation-fencing case. | -| F4 lost-context recovery | Not run | Not run | No new authenticated-epoch recovery case. | -| F4 secondary windows | Not run | Not run | No secondary-window case. | -| F4 physical input | Partial | Not run | One corrected Fabric `desktop_act` physical-click request was issued. MCP returned success and sequence completion, but the client did not retain the nested action mechanism/disposition; no physical-input pass is claimed and it was not repeated. Paper physical input was not attempted. | -| F3 startup/shutdown/survivors | Not run | Not run | No injected startup/shutdown survivor reconciliation. | -| F5 ingress bypass | Not run | Not run | No anonymous/foreign HTTP or WS mutation attempts on this SHA. | -| F5 owned selection/reset | Passed | Passed | Same-session authorized select and reset both applied with marker verification; preview-element old/new identity checks passed. | -| F5 channel lifetime | Passed | Not run | Fabric descriptor session remained open for 21,002 ms idle; subsequent status and close succeeded. No Paper idle interval on this SHA. | -| F5 manifest smoke | Partial | Partial | Non-agent `yarn storybook smoke --macos [--paper] --mode stories-and-tests` passed 211/211 with no skips or retries, but one-session/no-HTTP-fallback instrumentation was not retained. Separate from protected descriptor evidence. | -| F5 legacy separation | Partial | Partial | Non-agent smoke ran and was labeled separately as the legacy lane. No additional legacy-channel mutation trace was retained. | - -Fabric and Paper each passed the retained-preview element-reference sequence -and 211/211 separate legacy smoke callbacks. The same-port Fabric case was -recorded as Failed because its first post-replacement call returned -`service-unavailable`. Contract review reclassified it as Partial: that -response is the designed fail-closed rejection that -`packages/agentic/desktop-driver-mcp/src/server.test.ts` asserts before the next -call re-verifies discovery, but the native sequence ended with a close that does -not prove rediscovery. A native status call after the rejection remains -required. The -stale workload claim encountered on the first replacement attempt was not -removed until its exact owner was verified exited with no resources or -unresolved effects; the retry then started normally. On cleanup, owned app and -service PIDs were gone, ports 21338/30029/45382 had no listeners, the default -descriptor and workload lock were absent, and no external app was stopped. -The legacy run.json locators are -`apps/storybook/artifacts/desktop/a2475e8d-7ea9-465f-8461-b8f115cfcee5/run.json` -(Fabric) and -`apps/storybook/artifacts/desktop/87f0452f-5fd7-48eb-8494-a63202842521/run.json` -(Paper). Driver run records include the same-port attempt -`apps/storybook/artifacts/desktop/73563809-2d77-49cb-a297-c74fd57ecd36/run.json`, -blocked by the retained claim with 0 selected/executed tests. - -Windows Fabric and Win32 Paper remain **Not run** on this macOS-only host. -Fault-injection cases remain Not run because no existing documented native -hook was exercised. The older `3c780cfd1bccaf510ece080e4be2b516c645989a` -ledger and its inconclusive probe remain unchanged and separate. The Windows -results that follow were recorded later on a Windows host; they do not change -these macOS results. - -## Windows Fabric and Win32 Paper qualification — 2026-11-06 - -**Tested source:** the PR branch at `1f5333e25` plus two follow-up fixes from -this pass: `8458bde48` (a relaunched runtime reports `bridge-unavailable` -until its initial render) and `a6b89e924` (a reloaded runtime that -authenticates first retires the old bridge). Windows records started before -21:31Z predate both; `d558ed5d` includes only `8458bde48`; every later Windows -record and every Win32 record includes both. Host: Windows 11 Enterprise build -26300, AMD64, Node 26. Windows uses the React Native Test App host (window -title `Agentic Components Storybook`); Win32 uses the `rex` host, component -`AgenticStorybook` (window title `Agentic Components Storybook (Win32)`). -Windows App Runtime 1.8 was installed. The runbook is the Windows section of -[`native-qualification.md`](native-qualification.md). - -Preflight: `yarn storybook build-driver --windows` and `--win32` succeeded. -`yarn desktop-driver doctor --platform windows --permissions` and -`--platform win32 --permissions` report that `--permissions` is -**Unsupported** on Windows; doctor otherwise passed. No OS privacy or input -setting was changed. - -Each descriptor pass ran `yarn storybook driver -- --agent`, then -`yarn storybook run --`, then connected the MCP SDK stdio client to -`desktop-driver-mcp` through the default private `agent-connection.json`. -Only one interactive workload ran at a time. Manual MCP calls have no suite -selection, so each driver and app `run.json` reports 0 selected and 0 -executed tests; the cases are manual evidence in machine-local sanitized -records that are not committed. - -### Smoke runs (legacy lane, not descriptor evidence) - -Counts are selected / executed / passed / failed / skipped. - -| Endpoint | Command | Result | Artifact | -| -------------- | --------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------- | -| Windows Fabric | `yarn storybook smoke --windows --mode stories-and-tests` | Failed: 202 / 202 / 202 / 0 / 0; the relaunch phase reported `relaunch-required`. This led to `8458bde48`. | `apps/storybook/artifacts/desktop/2a80a82b-febc-40f8-85de-50c36f92f842/run.json` | -| Windows Fabric | Same | Failed: 206 / 206 / 205 / 1 / 0; the one failure is the physical-input test (foreground not acquired). | `apps/storybook/artifacts/desktop/9506772a-d3f9-4f53-966e-b6c75799b463/run.json` | -| Win32 Paper | `yarn storybook smoke --win32 --mode stories-and-tests` | Failed: 180 / 180 / 179 / 1 / 0; the one failure is a physical-pointer test, and 43 groups were not run. | `apps/storybook/artifacts/desktop/7a737861-c31b-4eeb-ac6d-b21b8a96739d/run.json` | -| Win32 Paper | Same, with physical input disabled as on hosted runners | Passed: 266 / 214 / 214 / 0 / 52. | `apps/storybook/artifacts/desktop/4ba5a00d-40d8-4a34-86f9-e564448623e2/run.json` | -| CI A/B | The `win32-storybook` and `windows-storybook` job steps | Win32 baseline failed at owned-service shutdown with 214 / 214 passed; the lifeline fix passed 266 / 214 / 214 / 0 / 52. Windows baseline failed 204 / 204 / 202 / 2; the readiness fix passed 312 / 234 / 234 / 0 / 78. | Baselines `47471c35-…` and `b7c67987-…`; fixes `047f5c3a-…` and `6cbfe8c5-…` | - -### Descriptor matrix - -| Case | Windows Fabric | Win32 Paper | Evidence | -| ------------------------------ | -------------- | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Same-port replacement | Pass | Pass | The old client's next status returned `service-unavailable`; old refs returned `no-session`; rediscovery, close, and a new session succeeded. No old mutation replayed. The new session's first select returned `relaunch-required` (rejected, effect none); after an app relaunch, select and reset applied with marker verification. | -| Grant revocation | Pass | Pass | After an external revoke (200), the next call and status returned `service-unavailable` ("authority is no longer valid"), and reopening on the same client was refused; a fresh connection opened and closed. Windows: the first attempt's post-revoke call was rejected by MCP input validation (the harness passed an unknown `budgets` key); the rerun passed. | -| Dispose during open | Pass | Pass | The client was disposed while `POST /session` was held; the traffic shows the late session deleted and the grant revoked, and a fresh connection opened. Incarnation unchanged; no target quarantined. | -| Failed delete | Pass | Pass | After the injected DELETE failure (`service-unavailable`), the session was retained and an observe succeeded; the retried close matched. Windows: the first attempt recorded `sessionUsableAfterFailure: false`, and its traffic has no observe request after the failure; the rerun recorded `true`. | -| Ambiguous create | Pass | Pass | With the create response dropped, open returned `service-unavailable`, status reported the session `uncertain`, and reopen and close returned `session-uncertain`; after the grant was revoked, a fresh connection opened. Incarnation unchanged. | -| Resource invalidation | Pass | Pass | See the reference evidence below. The immutable PNG resource stayed readable after reset with a matching capture digest. | -| Story quarantine and recovery | Pass | Pass | Windows: a Metro reload 15 ms into a select returned `story-not-ready` (indeterminate, quarantined), and later mutations returned `target-quarantined` while reads worked. Win32: an app kill 30 ms into a select did the same; after relaunch the same session stayed quarantined. On both, a new session ran the exit probe, lifted the quarantine, and select and reset applied. | -| Protected args refusal | Pass | Pass | `story-args-unsupported`; no value appeared in the response. | -| Failed-cleanup claim retention | Not run | Not run | No forced owned-resource cleanup failure was injected. | -| Launch-mode discovery | Partial | Partial | Attach-only targets were observed and never launched or terminated the app; no launch-capable comparison was run. | -| Live cleanup and restart | Not run | Not run | Cleanup was not forced to fail. Same-port service restart and reconnection are recorded above but are not this case. | -| Attached-app preservation | Pass | Pass | The externally started app process stayed running through service shutdown and was stopped only afterward. Win32 was observed by process presence. | -| F4 post-dispatch cancellation | Pass | Pass | Aborts at 150 ms and 40 ms landed at 160/47 ms (Windows) and 156/49 ms (Win32). The journal shows the dispatch completed; replaying the request ID returned the recorded applied result, and the next mutation applied in the same session. | -| F4 readiness/render faults | Not run | Not run | No readiness or render fault was injected through the descriptor path. | -| F4 lost-context fencing | Pass | Partial | Windows: in the in-flight run, a fresh preview root was observed while quarantined, preview mutations returned `target-quarantined`, and a chrome check passed. A quiet reload run on `a6b89e924` let the new runtime authenticate before the mutation, so it shows recovery rather than fencing. Win32 (production bundle): Metro reload has no effect (Unsupported); after app exit, select returned `bridge-unavailable` (rejected, effect none), but no fresh preview could be observed. | -| F4 lost-context recovery | Pass | Pass | After a relaunch or reload, the old preview ref was `ref-stale`, and select and reset applied with marker verification; the old context was not resumed. | -| F4 secondary windows | Not run | Not run | Not attempted. | -| F4 physical input | Failed | Pass | Windows: the descriptor `act` physical click was rejected ("The target window did not become the foreground window"; effect none), and the smoke physical test failed. Win32: the descriptor click applied with mechanism `physical-pointer`; separately, smoke 1's physical-pointer test failed and is retained. A semantic activation is not counted. | -| F3 startup/shutdown/survivors | Not run | Not run | Survivor reconciliation was not exercised. | -| F5 ingress bypass | Pass | Pass | HTTP `/select-story-sync`, `/send-event`, `/mcp`, and an unknown path returned 403; an anonymous WebSocket received only `furn:desktop:request-hello`; the leaf ref and region were unchanged. | -| F5 owned selection/reset | Pass | Pass | Select and reset applied with marker verification in one session. | -| F5 channel lifetime | Pass | Pass | The session stayed usable after 21,010 ms (Windows) and 21,006 ms (Win32) idle. | -| F5 manifest smoke | Not run | Not run | The one-session instrumentation was not run; the smokes above are legacy-lane evidence. | -| F5 legacy separation | Not run | Not run | The non-agent smokes ran in separate workloads, but no legacy-channel trace labeled unprotected was retained. | - -**Counts.** Windows Fabric: 15 Pass, 1 Partial, 1 Failed, 7 Not run. -Win32 Paper: 15 Pass, 2 Partial, 7 Not run. `--permissions` is Unsupported -on both. Win32 Metro-reload fencing is Unsupported (production bundle). - -**Reference evidence.** Each reference-sequence run used one session and -service incarnation 1 throughout. - -- Windows (target `windows-f33cce69c1a7d0ae`): preview story-root `9z4z23` - and child `9z4z66` became `ref-region-changed` after reset; window ref - `9z4z1` (chrome) survived on `window-df4a4ab9-b2`; the fresh root `9z4z102` - became stale after the second select (region 3). -- Win32 (target `win32-5e01f2ae23282f34`): preview refs `okdw21` and `okdw38` - became stale after reset; window ref `okdw1` survived on - `window-9951db02-b3`; the fresh root `okdw54` became stale after the second - select. -- The matrix scripts' ref labeled "preview element" (`wbwm18`, `1roq21`) was - actually sidebar chrome. It survived select and reset, which chrome may do; - it is not staleness evidence. - -**Run records** (`apps/storybook/artifacts/desktop//run.json`): - -- Windows driver supervisors: `d558ed5d-4117-4afa-b69b-5fe2b7777e26`, - `dc41f8c6-a7d0-4741-9a94-da88e3ab8619`, and the same-port replacement - `102c54a5-417d-4266-acdd-eb83372292c4` (all passed). The lifecycle records - `6424cc6f-6ffd-4ca1-83c0-8020f7817818` (passed), - `a5e9d797-1bd9-4ea4-a106-2f23f75dbff0` (failed, `rnx-cli` ENOENT), and - `e30e6500-105c-4055-a51d-a1f88d4f7f33` (failed) are retained attempts. -- Win32 driver supervisors: `80b2b8da-8428-4b90-a70a-d0ad9b6b5eae`, - `9044e1fe-d81e-4c14-b078-5542d1ad62be`, and - `2207ae33-ea4d-4b92-a4d8-5c3339a3faa0` (all passed). The app records - `3a6ee6cd-…`, `e805a6b8-…`, `380063db-…`, `acc6cd02-…`, and `1dd7cdfe-…` - report failed because the cases deliberately killed or externally stopped - the app. - -**Retained failures and deviations.** - -- Windows, before `a6b89e924`: after a quiet Metro reload, select returned - `story-not-ready` (indeterminate, quarantined) and later mutations returned - `target-quarantined`. This led to the bridge-retirement fix; the same case - on `a6b89e924` recovered. -- Windows: one matrix retry returned `bridge-unavailable` (rejected, effect - none) for a select, and the harness then failed on its own argument error. - The cause of that `bridge-unavailable` is not determined. -- Both matrix scripts requested a story list limit above 100, which schema - validation rejected; the reference-sequence runs used a limit of 100. - -**Cleanup.** After each pass, no owned app, driver, channel, or Metro process -remained; ports 8081, 8082, 6006, 7007, 4723, 17202, 29419, and 39225 had no -listeners; no `interactive.lock` or default `agent-connection.json` remained; -and the grant was revoked. Pre-existing unrelated processes were left -running. - -**Still Not run on Windows and Win32:** failed-cleanup claim retention, live -cleanup and restart, F4 readiness/render faults, F4 secondary windows, F3 -survivors, F5 manifest smoke, and F5 legacy separation. Hosted CI disables -physical input, so physical input is not qualified there. - -## Windows follow-up qualification - 2026-11-07 UTC - -Source: `386f0585edae2450ab22e6d359e2fe4db3399555` with repairs to the -package-boundary loader, link-check worker, and atomic reporting. Native/runtime -implementation was unchanged. Same Windows 11 Enterprise build 26300, AMD64 -host; Node 26.13.0. The earlier matrix and every failed attempt above remain -historical evidence, not erased by this follow-up. - -The seven previously Not run rows were exercised. All other rows retain their -earlier outcomes: **Windows Fabric 21 Pass, 2 Partial, 1 Fail; Win32 Paper -21 Pass, 3 Partial; neither has a Not run row.** This is not all-pass native -qualification: launch-capable discovery and separately registered secondary -windows remain Partial on both; fresh-preview lost-context fencing remains -Partial on Win32; Windows physical input remains Failed. Permission preflight -and Win32 Metro reload remain Unsupported. Hosted physical-input skips do not -change those results. - -| Follow-up case | Windows Fabric | Win32 Paper | Evidence | -| ------------------------------ | -------------- | ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| Failed-cleanup claim retention | Pass | Pass | Injected failure before the exact owned channel stop dispatched. The claim stayed fenced with channel generation 2 unresolved; replacement was refused before acquiring resources. | -| Live cleanup and restart | Pass | Pass | Original recorded stop callbacks settled, exact resources exited, descriptors disappeared, then retained owners/claims were reconciled. Fresh descriptor discovery in a new service incarnation selected and reset with marker verification. Restart alone was not used as cleanup proof. | -| F4 readiness/render faults | Pass | Pass | Held a real correlated ready for 1,500 ms after native render and delivered a stale correlation first. Selection settled only after the real message; old preview refs expired. A correlated render error returned unknown-effect quarantine and fenced later mutations. | -| F4 secondary windows | Partial | Partial | `native-callout--default` reported `Native window: Shown`; selection back to Button cleaned up in the same session. Only the registered primary window was addressed; this does not prove selection of a separately registered secondary window. | -| F3 startup/shutdown/survivors | Pass | Pass | Startup failed after a real typed service-ready receipt, and shutdown failed before the owned stop. Exact channel/helper survivors were reconciled through the recorded stop; Metro generation 1 settled, channel generation 2 retained the failed claim until cleanup was proved. | -| F5 manifest smoke | Pass | Pass | Instrumented manifest-backed smoke used one traversal session per phase and confirmed its DELETE before the next phase/authored session. All driver sessions were paired with successful deletes, without overlapping sessions or controller HTTP fallback. Overall smoke/report failures are retained below separately. | -| F5 legacy separation | Pass | Pass | Non-agent smoke traced unprotected `/select-story-sync` requests nested inside driver-owned story calls. This is the explicitly configured legacy transport, not fallback from the controller and not protected-ingress evidence. No descriptor or grant was used by this lane. | - -### Ownership failures and reconciliation - -`yarn desktop-test --target windows|win32 --reporter json serve --agent` -acquired real Metro, channel, native helper, and private descriptor resources. -Local instrumentation failed `running.stop` **before** dispatch for only the -recorded `storybook-channel` resource; the startup variant also threw after -the actual correlated `service-ready` receipt. The lock recorded -`service:storybook-channel:2`, generation 2, `disposal: failed`, and an -unresolved disposal effect. A second identical `serve --agent` command failed -with the workload-held diagnostic. It did not replace or steal the service. - -Calling the saved original stop callback then settled the exact channel and -helper. On the corrected startup runs and Win32 shutdown run, instrumentation -also recorded `authority.revoke` returning true; no grant identifier or token -was logged. The descriptor disappeared and ports 8081, 17202, and 39225 were -free. The supervisor retained its failed claim and remained alive; only after -that resource proof was its exact process stopped and its exact stale claim -removed. This operator reconciliation is not an automatic recovery claim. -External app incarnations survived throughout and were stopped separately -only after preservation was checked. - -| Endpoint / phase | Run ID | Terminal result | -| ------------------------------------------ | -------------------------------------- | ----------------------------------------------------------------------------------------------------------------- | -| Windows first secondary/fault harness | `9adb8cc6-b2de-4e12-b833-31314c03f111` | Passed supervisor; initial harness failures retained. | -| Windows Metro/native-launch attempt | `d2c2d92a-c561-4f57-8d71-9109c53a04b6` | Failed; Metro watch ENOENT during native build activity, followed by exact settled-resource/owner reconciliation. | -| Windows failed shutdown | `ed3d74cb-769f-4e9f-9ca6-7f00772e98c5` | Failed as injected; exact original stop subsequently settled. | -| Windows correlated readiness/render faults | `949dde01-8fa9-46b3-b6ee-9e2082eb0b56` | Passed supervisor and corrected MCP fault probe. | -| Windows first startup failure | `d729b429-89bc-44c4-9b06-13d3a00e0de6` | Failed; the local control-file reader crashed on a partial JSON write before reconciliation. | -| Windows corrected startup failure | `ea10eb32-ad0f-462d-8384-7d824357e5b2` | Failed as injected; original stop, grant revocation, and resource exit were observed. | -| Win32 failed shutdown and MCP faults | `509c46ef-d6ee-47d6-9fc5-4f9c475faa8a` | Failed as injected; original stop, grant revocation, and resource exit were observed. | -| Win32 startup failure | `f831f9d8-7335-4b98-8dad-5e8b6dade5e1` | Failed as injected; exact survivor reconciliation completed. | -| Windows fresh-service recovery | `5cad63bb-2e67-4033-a3ac-84ec44f5553d` | Passed; new MCP session selected and reset with verified markers. | -| Win32 fresh-service recovery | `dd134838-ca66-415c-afdb-828510739ef3` | Passed; new MCP session selected and reset with verified markers. | - -Every artifact is `apps/storybook/artifacts/desktop//run.json`. -Manual descriptor operations are not suite selections: these lifecycle runs -have zero selected/executed tests, not passing test counts. The new Windows -service-ready incarnation was `dd5b8e7c-cca3-4fb6-9bf5-b6838b270f97`, MCP -session `fbbdadc4-aa0e-45a9-9862-124052bc5e4e`; Win32 used incarnation -`28c0beaf-ffb2-4161-a230-3b9edabe1c1e`, session -`0245de0c-c004-4c76-865d-7458b62f999f`. Their normal shutdowns removed the -lock and descriptor, revoked authority, released all owned native resources -and ports, and preserved the attached app. - -The crashed first Windows startup run was **not** graceful cleanup: its -service and owner exited, leaving a stale descriptor and claim. Exact -process absence, no helper/listeners, and matching dead descriptor ownership -were established before removing those two stale files. Service-local -authority was invalidated by process exit; explicit grant revocation was -not observed on that attempt. The corrected probe retained the last valid -control during a partial write and used atomic updates. - -### Correlated render evidence and retained harness failures - -Both corrected probes used a real SDK stdio MCP client, default descriptor, -verified native service, and one uninterrupted driver session at incarnation - -1. Session regions were 1 -> 2 -> 3; runtime preview generation is a separate - identity, reported below. - -| Endpoint | MCP session | Preview / chrome refs | Window identity | Runtime epoch; delayed/error generations | Measured held-selection duration | -| -------- | -------------------------------------- | --------------------- | --------------------- | --------------------------------------------- | -------------------------------- | -| Windows | `46a48bed-5f9c-40c6-aac7-c2929461d222` | `0yoo23` / `0yoo1` | `window-4124f4f1-1be` | `7a4241eb-dc2c-4c9e-b6c5-bd7e7142ec74`; 2 / 3 | 1,637 ms | -| Win32 | `a29dc7af-c96f-4b0a-a98f-2eae04d7f8c1` | `oyno25` / `oyno1` | `window-5d019501-11b` | `a8a10ef6-7935-43c5-b198-79f4ced043bf`; 5 / 6 | 1,618 ms | - -The old preview returned `ref-region-changed` after the delayed transition. -The injected correlated error returned `story-not-ready`, indeterminate, -effect unknown, quarantined. Chrome remained present and a later reset was -rejected `target-quarantined`, effect none. Neither faulted tree exposed a -fresh preview root, so this stimulus's fresh-preview mutation subcase is -Partial, not additional fencing proof. Windows retains its earlier distinct -successful fresh-preview fence probe; Win32 remains Partial. - -Secondary-window probes used Windows session -`0ad62612-8bc5-4f9f-98f3-f987f9d0b4fb`, window `window-42ae1e0b-4`, -status ref `1kos1`, and Win32 session -`2d5302b1-e602-40b3-8536-66fad0f7d4a2`, window `window-5d019501-2`, -status ref `k0kf1`. Each changed region 2 -> 3 when selecting Button again. - -Retained harness failures are not product diagnoses: initially matching the -public request ID against the runtime's different internal ID did not inject -the fault; repeatedly using the same public request ID replayed a cached -effect-none `bridge-unavailable`; a stale-ref `check` returns a failed check -envelope rather than necessarily an MCP error; finding a fresh preview after -render failure did not succeed; opening a new session before quarantined-app -exit correctly failed. The corrected probe retried with a new ID only after -proven rejected/effect-none responses and never retried an ambiguous mutation. - -Windows app run `1885a5a3-0ba3-4206-8f8d-7894e77dc68a` failed at rnx-cli -Developer Mode enabling (2406); no OS setting was changed. Later Windows -launches used the exact already registered AUMID. The earlier app run -`cc9cebaf-9029-4a05-b234-5aaa96121886` passed. Win32 app runs -`177d3c9d-cbee-4387-a5bf-7ed135cc6811` and -`727752b3-427a-40e7-8cec-0e33cbef32d9` report failed because their externally -started apps were deliberately stopped after preservation/cleanup proof. - -### Instrumented smoke and atomic-report follow-up - -These are separate, explicitly **unprotected** non-agent workloads, with -`FURN_DESKTOP_DRIVER_DISABLED_INPUT_FEATURES=keyboard,physicalClick,wheel` -matching hosted restrictions. They do not replace the physical-input result -or descriptor-path evidence. Counts are selected / executed / passed / -failed / skipped; all rows have zero unsupported, incomplete, retries, flaky, -and groups not run. - -| Endpoint / attempt | Command | Result | Run ID | -| --------------------------------------- | --------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- | -------------------------------------- | -| Win32 instrumented | `yarn storybook smoke --win32 --mode stories-and-tests` | Pass: 266 / 214 / 214 / 0 / 52. | `670a4dea-adad-4082-b597-56160040fcf5` | -| Windows first instrumented | `yarn storybook smoke --windows --mode stories-and-tests` | Fail overall: 312 / 234 / 234 / 0 / 78; nine `EPERM` atomic-report rename failures. | `5ac696da-3ea3-47dc-ad90-1d4bf1110377` | -| Windows retry without live report reads | Same Windows command | Fail overall: 312 / 234 / 234 / 0 / 78; fifteen `EPERM` atomic-report rename failures. | `a67d9961-40c9-4118-95bc-0bc988599ab7` | -| Windows with bounded reporting recovery | Same Windows command | Pass: 312 / 234 / 234 / 0 / 78; one replacement recovered after one sharing retry, journaled as a warning. | `1825d9aa-2c25-4a2c-8524-ceadcf52ad68` | -| Win32 with bounded reporting recovery | Same Win32 command | Pass: 266 / 214 / 214 / 0 / 52; no sharing retry needed. | `8c6f58e1-3591-4f1b-b008-c52a718d3ca2` | - -Artifacts are `apps/storybook/artifacts/desktop//run.json`. -The first Windows attempt overlapped live PowerShell reads of the projection; -the retry did not, so those reads are not an established explanation. The -exact holder of the conflicting Windows file handle was not identified. -Atomic replacement failed transiently while later replacements succeeded; -the collector correctly retained those failures, making otherwise successful -test runs fail. - -A deterministic real Windows reader opened `run.json` without delete sharing -for 100 ms. The pre-fix collector failed the regression; the corrected -collector passed. Only Windows `EPERM`/`EACCES` renames are retried, at most -ten times with 25 ms spacing, retaining the old complete snapshot. Recovery -is journaled, not silently discarded; persistent conflicts, other errors, -and non-Windows failures still fail reporting. Neither native mutation nor -producer-handshake retries were added. Tests cover both recovery and -permanent-failure retention. - -Win32 paired all 91 session creates with successful deletes. Traversal -`797fabc1-6291-4f06-9a10-87486c48d2dd` made 176 story calls; readiness -traversal `ce2d8587-3357-4765-8174-75d5e7527868` made one, then was -deleted before the first authored session. Both Windows attempts paired 112 -creates/deletes; their story-phase traversals were -`3b691f62-dc41-4aef-96ab-a05e3115a47e` (233 calls) and -`e423ef0d-4c3b-4357-bbae-8eeb7dbcfd09` (232 calls), respectively. -Their post-relaunch readiness traversals were -`b64ede17-91b3-40c1-9554-88bb38857359` and -`e1b4b77a-28de-4bb4-8c2d-331d0e0cc20f` (eight calls each). -Each was deleted before authored-test session acquisition, without -overlapping sessions. Readiness retries are not test retries. - -All 266 Win32 and 313 Windows internal legacy selections per run were nested -inside exactly one driver-owned story call. There was no controller HTTP -fallback. The non-agent orchestrator intentionally uses the legacy transport; -the presence of those nested requests is not evidence of protected-ingress -bypass. After every attempt, no owned app/helper/channel/Metro or listener -remained, the workload lock and default descriptor were absent, and no -external app was terminated. - -The fixed Windows run again paired 112 creates/deletes: traversal -`beb044e6-d9c5-4ee4-bf77-57febd117123` made 233 calls, and readiness -traversal `690c6ae8-1bfd-49bb-8f7e-fa6253249a4e` made eight. The fixed -Win32 run paired 91 creates/deletes: traversal -`dccddd56-608c-4ad5-bdda-3aa4a0e2e753` made 176 calls, and readiness -traversal `c7233695-703d-4f70-ba6b-83d9d53bf1ba` made one. Both again -deleted the readiness traversal before acquiring an authored-test session. -Terminal reports have no failures. The Windows journal preserves its one -recovered write; this is reporting recovery, not a test or mutation retry. -After both fixed runs the exact owned processes were absent, all relevant -ports were free, and the per-user workload claim and default private -descriptor were absent. These non-agent lanes issued no descriptor grant. - -A full desktop-test package test attempt during the Windows smoke failed -the `agent-cli` serve fixture (96 passed, four skipped, one failed). -It also acquires a workload; overlapping it was a qualification scheduling -error. Contention is a plausible explanation, but the retained child output -does not establish the cause. The attempt remains recorded separately from -the subsequent serial validation. - -After both native workloads and their cleanup finished, `yarn test` passed -serially in desktop-test (97 passed, four skipped), desktop-driver -(813 passed, 72 skipped), desktop-driver-mcp (26 passed), and scripts -(76 passed). The desktop-test `agent-cli` fixture passed without a live -smoke workload; no product behavior was changed to suppress that failure. - -## Windows application ownership and foreground follow-up - 2026-11-08 UTC - -Source: `9d3e11516` (F16, stacked on -[#4468](https://github.com/microsoft/fluentui-react-native/pull/4468)) plus -the Windows application-ownership (F10-W, -[#4447](https://github.com/microsoft/fluentui-react-native/issues/4447)) and -foreground-diagnosis -([#4441](https://github.com/microsoft/fluentui-react-native/issues/4441)) -helper changes. Same Windows 11 Enterprise build 26300, AMD64 host; Node -26.13.0. The console session was **disconnected** (RDP, `LogonUI` in the -session) for the launch and measurement cases below. Foreground, physical input, -Windows Graphics Capture, and retention require separate connected-session -evidence. - -**Updated matrix: Windows Fabric 22 Pass, 1 Partial, 1 Fail; Win32 Paper -22 Pass, 2 Partial.** Launch-capable discovery moves from Partial to Pass on -both endpoints. Separately registered secondary windows (both), Win32 -fresh-preview fencing, and Windows physical input keep their earlier results; -they need a connected, unlocked session. - -| Case | Windows Fabric | Win32 Paper | -| ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Launch-capable discovery | Pass. `desktop_session` status listed `launchModes: ['launch']`; `open` launched the packaged app in 341 ms and `close` ended `ReactApp.exe`. Run `0876d93b-f689-47e5-8379-4c09d95d6cc9`. | Pass. REX FluentTester (`apps/win32`) listed `['launch']`; `open` launched in 625 ms and `close` ended `ReactTest.exe`. Run `101c5916-027d-40cf-a26a-163535df4c03`. | -| Attach-only refusal | Unchanged: Storybook targets are always attach-only and advertise only `attach`. | Pass: the Storybook Win32 target refused `launch` before acquiring resources (run `411c5218`). | -| Helper death after launch | Expected `uncontained`: killing the helper left the packaged `ReactApp.exe` running. The claim was fenced (`webdriver` disposal failed, one unresolved effect). The survivor was ended manually after its identity was checked. | Pass, `job-object`: killing the helper ended `ReactTest.exe` within 5 s. The claim was fenced the same way. | -| Agent MCP measurements | Recorded in the adapter's status reference (run `caba9e12-d9da-42a0-8b06-7300940e8318`). | Recorded there (run `5f75aac2-d74c-4969-aba3-bd8c925d5725`). | - -The launch-capable targets are not Storybook targets. Win32 used -`apps/win32/desktop.config.mts`. Windows used a temporary, uncommitted -`installed` host naming the registered Storybook package AUMID; the -committed Storybook targets are attach-only. Each probe used the SDK stdio -client and the default descriptor. In both crash variants the service exited -with code 1 and wrote a failed `run.json` naming the helper's unexpected -exit. Neither the claim nor the settlement journal names the surviving -uncontained application, because the journal records applications only for -mutations and quarantines. Recording a launched application as a claim -resource belongs to F11's retained-application handoff -([#4448](https://github.com/microsoft/fluentui-react-native/issues/4448)). - -**Native contracts.** With `FURN_NATIVE_DRIVER_TEST=1`, `nativeDriver` -passed. While the session was still connected, `win32Fixture` passed 17 with -2 skipped, before review fixes. While disconnected, the three new -containment cases passed, but window capture and the 500-cycle retention -case failed. The retention case grew from 316 to 1,516 retained elements, -exactly three per cycle. A control run of the same case on the unchanged -helper failed identically while disconnected. After reconnection, -`FURN_NATIVE_DRIVER_TEST=1 yarn test win32Fixture --runInBand` passed all -11 cases, including capture and retention: retained elements stayed at 20 -from cycle 100 through cycle 500. - -**`serve --agent` interrupt.** Interrupting a Storybook `serve --agent` with -a console Ctrl+C three times (Win32 runs `411c5218` and `5f75aac2`, Windows -run `caba9e12`) settled every owned -resource. Metro, the channel, the helper, and the descriptor were gone, and -the ports were free. The claim then stayed `active` with no unsettled -resource or unresolved effect, `run.json` stayed `incomplete`, and the CLI -process stayed idle without exiting. Each time, the exact owner was ended -only after that proof, and then its stale claim was removed. The physical -console Ctrl+C case is otherwise out of scope for this pass. - -The idle CLI had one cause on every platform: the CLI tracked the Storybook -integration's run in the service scope it lent to that run, and the run closed -the scope, so closing waited on itself. On macOS, a Storybook-integration -`serve` was still running 90 seconds after SIGINT, with its claim `active`. -Integrations now leave the borrowed scope to the CLI, the CLI waits for the run -outside the scope with a bound, and scope draining is bounded. With that fix the -same macOS `serve` exits promptly after SIGINT and leaves no owned process or -claim. A physical console Ctrl+C on Windows and Win32 still needs requalification. - -**Reconciled claims.** Every stale or fenced claim in this pass was removed -by hand only after its owner was proved dead and its resources were proved -settled, or after the fenced launch probe's survivor was ended. Nothing -reclaims a claim automatically. - -## Windows Fabric and Win32 Paper — #4441 connected-session qualification — 2026-11-12 UTC - -**Observed baseline:** `505a2187b0a2451aed7081c9260b343d55742f93`, -Windows 11 Enterprise build 26300 x64, Node 26.13.0. Native/GUI work ran -serially in the W3 campaign, 14:50:24–15:22:51 UTC, under the machine-wide -lock. Every before/after physical gate reported an active unlocked -**remote RDP** session. The attached display was one 1728 x 1084, -origin-zero, 96-DPI / 100% monitor; no display setting changed. - -The registered Fabric native build was `bdeb0b291`; the compared app/native -sources differed only in native-core version/changelog metadata, not Windows app/native -implementation. **Inferred:** reuse was valid. Own-checkout Metro, -generated manifests and protected bundles were used. Paper ran the prebuilt -REX 0.81.1 host. Each supervisor matched the newly launched app's exact -image, start time, title and window before probing it. - -### Per-case outcomes - -These are manual native/descriptor probes, not selected authored suite -cases. **Observed** results must not be inferred from supervisor exit alone: - -| Case | Windows Fabric packaged | Win32 Paper REX | -| ------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Matching-OS F16 native contracts | Passed: shared Windows-provider/fixture run, 3 suites / 27 tests | Same shared run includes the real Win32 fixture; 2 macOS-only build cases skipped, not passes | -| Nonempty RN physical keyboard clear | Passed, keyboard receipts 1 -> 2 | Passed, keyboard receipts 1 -> 2 | -| Initially-empty accessible clear | Passed, semantic-clear receipt | Passed, semantic-clear receipt | -| Readonly accessible clear refusal | Passed, effect none, receipts unchanged | Passed, effect none, receipts unchanged | -| Separately registered secondary observation / interaction / cleanup | Passed: one exact new opaque handle, complete 11-node observation, unique control, physical-pointer click, `focus: confirmed`, then popup handle removed | Not run / partial: popup shown, but no separately registered handle | -| Additional secondary semantic activation effect | Failed: applied/accessibility receipt but actual main operation status stayed `Idle` | Not run: no registered secondary | -| Fresh native preview after lost provider context | Not requested | Passed: old preview rejected `ref-region-changed`; fresh unique native preview rejected `story-context-stale`; both effect none, not quarantined; state unchanged and old chrome ref survived | -| Physical MCP click measurement | Passed: 6 clicks and post-click state checks | Passed: 6 clicks and post-click state checks | -| Minimized capture and explicit restore | Passed: unsupported-operation refusal without implicit restore, then valid restored capture | Passed: same safety contract | -| 200% scale | Not run: 100% RDP scale; changing it not authorized | Not run: same | -| Mixed DPI | Not run: one attached monitor | Not run: same | -| Negative-origin layout | Not run: one origin-zero monitor; no approved eligible layout | Not run: same | -| Physical console Ctrl+C | Not run: separate implementation lane owns it | Not run: same | - -The native strict-clear self-test supplies nonempty **semantic** clear -evidence; the RN semantic endpoint probe started empty. Provider -keyboard-denial instrumentation is not substituted for the separate real -RN keyboard deletion. The native run skipped only the two macOS-specific -build/signing cases. - -### Commands and retained evidence - -Run from the qualification checkout. `$CampaignTools` is the separately -managed private campaign helper directory, not committed configuration: - -```powershell -$QualificationRoot = (Get-Location).Path -$env:FURN_NATIVE_DRIVER_TEST = '1' -pwsh -NoProfile -File "$CampaignTools\with-native-lock.ps1" ` - -Lane W3 -Label 'W3 qualification native strict-clear and fixture contracts' ` - -RequirePhysical ` - -WorkingDirectory (Join-Path $QualificationRoot 'packages\agentic\desktop-driver') -- ` - yarn test --runInBand nativeDriver win32Fixture operations.contract -``` - -Endpoint lifecycle commands below each ran within the same -`with-native-lock.ps1 -Lane W3 -RequirePhysical` scope. The local supervisor -awaited driver readiness, ran the declared bundle stage before app launch, -then owned probes and cleanup: - -```powershell -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs win32 -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs win32 --fence-only -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --mcp-only --with-secondary -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs win32 --mcp-only -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only --physical-secondary -``` - -Artifacts are retained locally under the qualification checkout at -`artifacts\desktop\w3-qualification\runs\\`, including -`qualification.json`, command logs and the named probe JSON. They are not -published package files. `artifacts\desktop\w3-qualification\qualification-summary.json` -is the machine-readable matrix. - -| Case / endpoint | Qualification ID | Probe JSON / driver run | -| ----------------------------------------------- | -------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- | -| Clear and minimized, Fabric | `5c97fc26-71ef-483c-aa2a-b454d6ff4458` | `descriptor-cases-windows-clear.json`, `descriptor-cases-windows-minimized-capture.json` | -| Clear, minimized, unregistered secondary, Paper | `22332065-6fa3-4061-a2b6-93f83295ca2e` | `descriptor-cases-win32-clear.json`, `descriptor-cases-win32-minimized-capture.json`, `descriptor-cases-win32-secondary.json` | -| Fresh-preview fence, Paper | `60751352-c158-40a2-b4f9-9251e15fec95` | `descriptor-cases-win32-fence.json`; driver `64d2de39-df69-4a2b-a164-fc842f0ef64e` | -| MCP physical clicks, Fabric | `f188fae5-67c9-47e4-9800-ea584185fed9` | `mcp-click-windows.json`; driver `57d4f6cb-2dc3-4b00-bb95-4331efd8d34b` | -| MCP physical clicks, Paper | `6d474fd4-61cb-41f6-824e-7a80df68db45` | `mcp-click-win32.json`; driver `918930ba-e498-4e20-ad98-52e4187b5618` | -| Secondary semantic failure, Fabric | `675160f7-e069-413c-a6c8-486de0c37b69` | `descriptor-cases-windows-secondary---modern.json`; driver `a9f2e71e-71ab-40f0-a375-fd6d76a0b971` | -| Secondary physical pass, Fabric | `7fbe3b0b-df33-476b-80e3-95b97bf19f02` | `descriptor-cases-windows-secondary---modern---physical.json`; driver `16aac92b-bb36-4742-bd29-01df381d9594` | - -### Baseline failures, scope and cleanup - -**Observed:** Fabric semantic activation targeted the unique native popup -button and returned applied/accessibility, but six status checks over -2090 ms and a final native observation remained `Idle`. Main was read -through its cited root without changing foreground. The same test with -physical click passed in one attempt / 79 ms. **Inferred:** native -accessibility-pattern/renderer dispatch may acknowledge without executing -Pressable `onPress`; no exact pattern trace was captured. - -**Observed:** Paper's shown Callout produced a visible, untitled root -`Net UI Tool Window` with ToolWindow true, but no registered secondary. -Modern Callout was absent from its catalog. **Inferred:** the provider's -intentional untitled-toolwindow filter explains the registration gap. -Both findings are being fixed on this branch; neither is claimed fixed -by this baseline evidence. - -The Win32 fence deliberately invoked the **real** provider's bridge -retirement while preserving the socket/native preview. In one uninterrupted -session, old ref `hkbo1` was fenced, fresh unique ref `hkbo3` was fenced -in the next preview epoch, and chrome ref `hkbo2` remained usable for -presence checks. This proves injected provider-context loss, not a native -crash, real network disconnect/reconnect, or chrome mutation. - -Earlier missing-catalog/setup failures, invalid probe selectors/facts and -secondary receipt-only attempts remain in the run history; they are not -erased by independent case passes. Fabric's final MCP case passed inside -an enclosing run that failed its separate semantic-secondary assertion. -Win32 window close hid REX without ending it; the uniquely owned run CLI's -cooperative stop drained its process tree. This is not physical Ctrl+C -verification. No other lane's portable test/build workers were stopped. - -All owned apps, helpers, Metro/services, sessions and child grants drained. -Endpoint listeners were absent. Campaign end and both audits **passed**: -`FURN-AUDIT: no FURN desktop processes found`. Display-layout and console -qualification remain not run. MCP costs/budgets are recorded in -`@fluentui-react-native/desktop-driver-mcp/references/current-status.md`; -Storybook context details are in -`@fluentui-react-native/storybook-desktop/references/current-status.md`. - -## Connected Windows physical-input follow-up - 2026-11-08 UTC - -Source `fec94cec4995c7372b798877c5412ddcb156523b`; Windows 11 Enterprise -build 26300, AMD64, Node 26.13.0. With the console connected, the following -separate, non-agent smokes ran from `apps/storybook` with -`yarn storybook smoke -- --mode stories-and-tests --input physical ---reporter json` and `STORYBOOK_SMOKE_STORY` set to the listed story. -Each had zero failures, skips, unsupported cases, incomplete work, retries, -or flaky cases. The totals include navigation work, not just authored tests. - -| Endpoint | Story and authored assertion | Selected / executed / passed | Physical-click receipts | Run ID | -| -------------- | ----------------------------------------------------------------- | ---------------------------- | ----------------------- | -------------------------------------- | -| Windows Fabric | `components-checkbox--default`: unchecked -> checked -> unchecked | 203 / 203 / 203 | 2 | `ef1ed966-3be3-4cc1-b940-ac0afbb02155` | -| Win32 Paper | `components-checkbox--default`: unchecked -> checked -> unchecked | 177 / 177 / 177 | 2 | `54e6a815-21a4-418c-bfda-848e28549ffe` | -| Windows Fabric | `native-callout--default`: open, light-dismiss, recreate | 203 / 203 / 203 | 3 | `1e521138-f529-45a1-974e-74f90edd42ac` | -| Win32 Paper | `native-callout--default`: open, light-dismiss, recreate | 177 / 177 / 177 | 3 | `44d9dd17-7132-4ebd-8b1e-18fa923ee355` | - -Artifacts are `apps/storybook/artifacts/desktop//run.json`. The -Windows totals contain 202 navigation cases and one authored case; Win32 -contains 176 and one. Popup interaction does not establish F4's separately -registered Storybook secondary-window identity. - -The connected native fixture contract passed 11 tests under the per-user -claim; its retained-element counts were 20 and 23 at cycles 100 and 500. -The separate generic fixture probe passed exact secondary-handle switching -and minimized-capture refusal as described in the driver status reference. -Windows `doctor --permissions` remains Unsupported. - -The real MCP descriptor path independently passed physical checkbox -transitions and target capture on both endpoints. Standalone agent driver -runs `af97fdb3-68de-4ce9-836e-935331528c8e` (Windows) and -`3294d66d-9332-4639-95fe-8d3047f47e46` (Win32) each ended with a -terminal passed lifecycle report and zero selected tests; the manual MCP -operations are not suite selections. The first Windows probe exhausted its -30 safe `bridge-unavailable`/effect-none attempts while the cold Metro bundle -was still building; the later probe passed after the app connected. - -After MCP session close, Windows accepted an exact-app window close. On -Win32 the window closed but its exact REX process remained; it was stopped -only after identity verification. Both deliberately stopped Win32 app runs, -`665e6e85-f0ce-4612-a34d-f140d2bab433` and -`fc71d8fd-ffa4-464e-96ea-616d88d49e0e`, report failed, not graceful cleanup. -The standalone drivers were stopped by a session-local injected SIGTERM, -not physical Ctrl+C. Both owned app/service processes exited, the recorded -ports were free, and the per-user claim and default descriptor were absent. -The separate `serve --agent` interrupt limitation above is fixed but not yet -requalified on Windows. - -## F6 generated manifest provenance — 2026-11-12 - -Implementation source: `efd6dadeda87344911fb4000fe293c686c678602`, with -resolved-case compatibility in `fa0681635` and the awaited shared workload -claim in `a5de740f01ca025bf548a923ae4561b36fc9872b`. -Generated driver manifests are strict version 3 and contain no secrets. -Version-1 private credentials use independent bridge and application attachment -nonces. The owner-private runtime module contains only bridge data and a -data-only `export default` statement. The parent's independent handoff pins -manifest/credential/runtime hashes and the resolved helper; the actual channel -child validates that binding and re-verifies the executable before a helper -can spawn. Metro receives the canonical schema locator from trusted process -configuration, not from editable generated files. Channel lifecycle placement -and application lease ownership have not moved. - -Review fixes route Windows/Win32 smoke lease publication, replacement, and -cleanup through the trusted Node helper and OS private-storage adapter; -PowerShell receives only a digest receipt. Smoke control now uses the strict -shared version-3 parser. Publication resolves the existing generated directory -once before opening private storage, supporting symlinked projects and macOS -`/var` aliases while retaining later fail-closed ancestor validation and -actionable diagnostics. - -**Portable/local boundary checks:** the initial 364-test integration result -used a project-local temporary directory and did not qualify stock macOS -temporary paths. Review exposed that gap; those counts are superseded by the -default-`TMPDIR` review validation below. Runtime format, lint, build, and all -33 tests pass with the default temporary directory. The driver package passed -format, lint, build, and 1,221 portable tests with 66 skips; focused private -storage tests passed 65 with ten platform skips. -The reviewed integration package passes format, lint, build, and 379 tests -with four Windows-only skips. Its focused generated-manifest, lease, and real -WDIO-runner contract selection passed 38 tests, including the actual v3 smoke -control path. The ledger-owning harness passes format, lint, build, and -380 tests with one skip; its claim-sensitive tests ran under the shared lock. -These review runs used the default macOS `TMPDIR`, without an override. -The channel-child rejection fixture now canonicalizes its temporary parent -too, ensuring it reaches helper/hash refusal instead of an alias-path error. -Evidence: `apps/storybook/artifacts/f6/review-{driver-focused,driver-test,integration-focused,integration-final,runtime,harness}.log`. -Controlled Metro fixtures cover missing trust inputs, unknown fields, and -manifest/runtime tampering. Actual channel-child fixtures reject substituted -helper path/hash, helper bytes, and unknown fields before listening or spawning. -The real macOS private-storage adapter round-trips the records with directory -mode `0700` and file mode `0600` and rejects a permissive credential mode. - -**Pre-review native helper checks:** on macOS 27.2.1 arm64, the locked -`FURN_NATIVE_DRIVER_TEST=1 yarn workspace -@fluentui-react-native/desktop-driver test --runInBand` run passed 1,240 tests; -47 platform/capability tests were skipped (75 suites passed, six skipped). -Evidence: `apps/storybook/artifacts/f6/native-contract.log`. The explicitly -built helper uses the stable `FURN Desktop Driver Development` identity; -`doctor --platform macos --permissions` reported ready, unlocked/on-console, -and Screen Recording preflight true, but Accessibility and PostEvent preflight -false. No permission prompts, privacy changes, or TCC resets were requested. -The native consumer runs below must not be interpreted as passed from that -helper readiness result. - -**Native consumer preparation:** Fabric bundle passed -(`apps/storybook/artifacts/f6/bundle-fabric.log`). Initial Fabric prep failed -because CocoaPods called `JSON.fast_generate` after Ruby selected JSON 3.1. -Selecting the already installed JSON 2.18 through a worktree-local Ruby -preload, without installing or modifying shared tools, made prep pass -(`prep-fabric-json2.log`, run `8df5c79a-9618-423d-b65d-e9828644a75b`). -The first Fabric smoke exposed rejection of the extractor's legitimate -resolved `inputSource` metadata; `fa0681635` fixes it and adds declared/derived -descriptor regressions while preserving unknown-field refusal. - -**Fabric native consumer: blocked, not passed.** The corrected smoke built and -launched the app, started the verified channel, and reached authenticated -bridge epoch recording. Native attach then reported Accessibility -`apiDisabled` (`-25211`), and session creation failed because the target exposed -no supported click input mode. No stories or tests executed. -Evidence: `apps/storybook/artifacts/f6/smoke-fabric-resolved-options.log` and -`apps/storybook/artifacts/desktop/65fb216f-89ee-43db-86a6-333b3d91331e/run.json`. -The harness's AppleEvent cleanup also timed out (`-1712`); its leftover app -process was verified against this worktree's executable, terminated by that -exact PID only, and confirmed gone. No privacy workaround was attempted. - -**Paper native consumer: blocked, not passed.** Paper bundle and prep passed -with the same existing-JSON-2 selection (`bundle-paper.log`, run -`3c281fb3-eb3a-4125-a189-86ced929d6ad`; `prep-paper.log`, run -`9b575663-fa72-4625-9f09-b745bfa1e83e`). Paper smoke also built and launched -the app and reached bridge epoch recording, then failed on Accessibility -`apiDisabled` and no supported click input mode. Zero tests executed. -Evidence: `apps/storybook/artifacts/f6/smoke-paper.log` and -`apps/storybook/artifacts/desktop/cfab6e50-9b38-45d3-b114-05b3ea995037/run.json`. -Its AppleEvent cleanup timed out as well; the exact worktree-owned leftover -app PID was verified, terminated, and confirmed gone. Positive Fabric/Paper -traversal and authored-case qualification remain pending an authorized -Accessibility/PostEvent-capable host. - -**Reviewed repository checks:** source `ffb4944db` passed all 90 -`yarn lage test-links --verbose --grouped` tasks. After `yarn clean-lage-cache`, -the locked `TF_BUILD=1 yarn lage buildci --concurrency 4` passed all 178 tasks -with the default macOS `TMPDIR`. The existing test wrapper uses `TF_BUILD` -only to run Jest in-band; it does not skip agentic tests. Evidence: -`apps/storybook/artifacts/f6/review-links.log` and `review-buildci.log`. -The Accessibility-blocked macOS consumer smoke was not repeated for review. - -**First PR CI follow-up (2026-11-12):** run `43145388693` on -`2739349a3` exposed five Linux CLI fixture failures and both Win32 smoke -lanes refusing private storage before any stories ran. The Linux cause is -confirmed: the command factory dropped the injected `createPrivateStorage` -adapter. It now forwards that option, and platform selection explicitly -supports macOS/Windows while other production credential paths fail closed. -Portable contracts inject storage and simulate all three platform branches. - -The Win32 logs retained only the generic unsafe-storage message, not the -adapter's non-enumerable refusal reason; the exact native refusal category -cannot be recovered from that evidence. Native inspection confirms that the -workspace root was checked before creation and ancestor grants allowing -untrusted child replacement are refused. Unsafe hosted-workspace ancestry -is therefore an inferred cause, not a confirmed refusal category. -Windows publication now asks the verified helper for its OS-selected per-user -known-folder root and opens a deterministic, independently protected child -namespace for the canonical generated directory. Reading rederives that -namespace only after verifying the parent handoff and helper. No environment -path, editable manifest root, relaxed ACL, or root repair chooses authority. -Storybook now surfaces the bounded native refusal token. Metro retains its -workspace roots; this initial fix also watched the private runtime projection, -which the subsequent CI run exposed as incompatible with cross-volume Windows -workspaces. - -The separate Windows setup job failed four native build/handshake/UIA -timeout cases, but its existing Windows private-storage native contracts -passed. This branch does not change Windows helper build, resolution, -handshake, C++, UIA, or fixture code; setup runs before Storybook publication. -Runner flakiness is an inference, not a confirmed explanation. The next -same-branch CI run passed setup, supporting that inference; no timeout or -native-helper behavior was changed. - -**CI-fix portable validation (default `TMPDIR`):** driver format/lint/build/test -passed (1,222 tests; 66 skipped); Storybook integration passed (386; 4 skipped); -runtime passed (34). The new contracts cover OS dispatch, helper-selected -project isolation, trust-before-helper ordering, changed-root and tamper -refusal, protected child calls, and outside-workspace Metro resolution. -Evidence: `apps/storybook/artifacts/f6/ci-fix-{driver,integration,runtime}.log`. -These simulated-platform results are not Windows ACL qualification. - -**CI-fix repository validation:** source `17032203b` passed all 178 -`TF_BUILD=1 yarn lage buildci --concurrency 4` tasks after -`yarn clean-lage-cache`; both commands held the shared native lock and used -the default `TMPDIR`. Links passed all 90 tasks. Repository lint, instruction, -changeset, and whitespace checks passed. Evidence: -`apps/storybook/artifacts/f6/ci-fix-buildci.log` and `ci-fix-links.log`. -The Accessibility-blocked macOS consumer smoke was not repeated. - -**Second PR CI follow-up:** run `43150784347` on `ca27c1b0b` passed JS PR, -all macOS lanes, Windows/Win32 PR, and Windows native-helper setup. All three -Windows Fabric / Win32 Paper Storybook lanes opened the revised protected -per-user namespace, then failed before smoke during Metro bundling. The -workspace was on `D:` and private projection on `C:`; Metro's relative-path -file map recombined the absolute cross-volume result into `D:\C:\...`. -This confirms the new namespace works on the hosted runner, not that the -consumer flow is qualified. - -The bridge projection contains the bridge nonce, so it remains protected. -Windows/Win32 Metro now indexes only the constant workspace stub and supplies -the pinned private buffer through its supported custom-transformer hook, -delegating compilation/minification to the consumer's original worker. The -worker rechecks the digest, and its cache key binds that digest to avoid -stale nonce transforms. No cross-volume private watch folder or public -generated source copy is created. macOS private-path behavior is unchanged. -Windows namespace/path comparisons normalize drive letters, casing, and -separators without replacing the independent byte-digest trust checks. - -**Cross-volume portable evidence:** controlled fixtures cover a `C:` private -source with a `D:` workspace, case/drive aliases, a wrongly located stub, -unchanged watch roots, digest tampering, nonce cache invalidation, and -unrelated-module delegation. The actual consumer Metro worker/Babel/minifier -also compiled both Windows targets. A bounded full Metro file-map/worker-farm -fixture built minified Windows and Win32 bundles with the private source -explicitly excluded from the file map. These JavaScript checks are not native -Windows qualification. Evidence: -`apps/storybook/artifacts/f6/crossdrive-real-worker.log` and -`crossdrive-real-bundle-final.log`. - -**Cross-volume package validation (default `TMPDIR`):** format/lint/build/test -passed for the driver (1,222 tests; 66 skipped), Storybook integration -(387; 4 skipped), and runtime (37). The final transformer cache-key change -also passed the 11 controlled Metro cases. Knip now explicitly includes the -Metro config/worker entrypoints; repository lint, instruction integrity, -changesets, whitespace, and all 90 link tasks passed. Evidence: -`apps/storybook/artifacts/f6/crossdrive-{desktop-driver,storybook-desktop,storybook-desktop-runtime}.log`, -`crossdrive-final-format.log`, `crossdrive-repo-checks.log`, and -`crossdrive-repo-safeguards.log`. - -**Cross-volume repository validation:** source `e760a36de` passed all 178 -`TF_BUILD=1 yarn lage buildci --concurrency 4` tasks after -`yarn clean-lage-cache`. Both commands held the shared native lock and used -the default `TMPDIR`; no native smoke or privacy-setting change was repeated. -Evidence: `apps/storybook/artifacts/f6/crossdrive-buildci.log`. - -**Windows/Win32 qualification: sufficient for F6's changed paths.** -[Run 43156299266](https://github.com/microsoft/fluentui-react-native/actions/runs/43156299266) -passed on source `58ec7ab4fb7ce6e449d0e351ecb02352bf88f319` on -2026-11-12. Full job logs and the uploaded terminal `run.json` reports agree: - -| Lane / job | Full story index rendered | Authored tests | Aggregate | -| ---------------------------------------------------------------------------------------------------------------------------------------- | ------------------------- | ------------------------------------- | ---------------------------------------------------- | -| Windows Fabric accessible / [130976919836](https://github.com/microsoft/fluentui-react-native/actions/runs/43156299266/job/130976919836) | 215/215 | 40/40 passed | 255 selected, executed, and passed | -| Win32 Paper accessible / [130976919943](https://github.com/microsoft/fluentui-react-native/actions/runs/43156299266/job/130976919943) | 188/188 | 45/45 passed | 233 selected, executed, and passed | -| Win32 Paper physical / [130976920038](https://github.com/microsoft/fluentui-react-native/actions/runs/43156299266/job/130976920038) | 188/188 | 100 passed; 1 quarantined unsupported | 289 selected; 288 executed and passed; 1 unsupported | - -All three reports have zero failed, skipped, incomplete, retried, or flaky -tests and zero groups not run. Each navigation result has a distinct story -ID and passed; the render totals match the complete runtime index, not an -early-stop subset. The physical lane's existing #4373 quarantine is not a -pass or an F6 failure. - -The five required paths are exercised in **each** lane. The successful -guarded execution establishes the mandatory checks below; these checks do -not emit separate success messages or upload secret records: - -1. **Version-3 control:** - `packages/agentic/storybook-desktop/config/storybook-control.cjs` uses the strict shared - generated-contract parser before driver traversal and authored tests. - The rendered-story and authored-test results above therefore cross the - v3 gate in all three lanes. -2. **Trusted helper verification:** each log records native Storybook/WebDriver - readiness before traversal. The channel child's - `packages/agentic/storybook-desktop/config/server.cjs` - requires the inherited parent handoff and awaits `readDriverManifest`, - which checks its independent byte/helper binding and calls - `verifyNativeDriverArtifact` before opening Windows private storage or - starting the native service. Readiness plus completed native traversal - demonstrates this guarded spawn path, not merely parent-side resolution. -3. **Protected-DACL lease lifecycle:** both PowerShell scripts require a - successful trusted Node publication and digest receipt before control, - and propagate lease-helper cleanup failures. Successful script completion - exercises publication, native consumption, and receipt-checked cleanup in - all lanes. Windows Fabric additionally stops/relaunches the app between - its stories/tests phases and replaces the lease using the prior receipt; - the Win32 lanes use one app incarnation. The Windows job also passed the - real NTFS `privateStorage.contract.test.ts` (log line 436), including - exclusive publication/removal and refusal of permissive root/record ACLs, - junctions, hard links, and replaced roots. This is native boundary - evidence, not portable mock coverage or an independent post-run ACL dump. -4. **Private runtime resolution:** all three logs finish writing the minified - Windows/Win32 bundle before native startup. On these hosted runners the - workspace is on `D:` and the protected per-user projection is on `C:`. - The Windows runtime transformer must read and hash-check that private - buffer before delegating the workspace stub to Metro's original worker; - the subsequent authenticated render confirms the resulting bundle works. -5. **Live bridge authentication:** successful native story selections require - `StorybookChannelOrchestrator.acceptHello` to match the nonce, instance, - endpoint, target, manifest digest, and connection epoch, followed by - correlated render-ready responses. The 215/188/188 successful selections - prove live authenticated runtime traffic, not just a listening server. - -**Evidence locators:** the Windows full log records bundle completion at -line 655, native readiness at 669-670, all stories at 709, authored tests at -1230, and terminal totals at 2295. Win32 accessible uses lines 502, -508-511, 1097, and 1118; Win32 physical uses 499, 505-508, 1809, and 1845. -The three artifacts are `Storybook_windows_fabric_43156299266_1` -(13916946307), `Storybook_win32_paper_accessible_43156299266_1` -(13917010025), and `Storybook_win32_paper_physical_43156299266_1` -(13916318019). Their main smoke reports under `apps/storybook/artifacts/desktop` -are respectively `414aa6f1-a539-4d67-9aa4-a92db05b58a4/run.json`, -`00cae5a0-ef44-45cf-aaa9-68485ecf2403/run.json`, and -`e5767c29-e4b1-4a9b-9685-8cdf8e0d3ee1/run.json`. Full logs, inventory, -and extracted reports are retained locally under ignored -`apps/storybook/artifacts/f6/ci-43156299266/`. - -**Limits:** collected channel stderr includes non-terminal -`bridge-unavailable` navigation diagnostics on Windows and an authenticated -epoch `lease-invalid` diagnostic on Win32 physical. Do not infer their -original timing from the buffered output, diagnostic-free operation, or -general bridge-recovery qualification from the terminal passes. The local -macOS Accessibility block above remains unchanged; no local smoke was -repeated. `pwsh` is unavailable locally, so its parser check remains not run; -the actual scripts executed in these CI lanes. Keep the PR draft until the -documentation follow-up's CI passes and the coordinator changes its status. - -## 2026-11-12: F10-M neutral application ownership - -Implementation source: -`76a2be59abfde3a7bcb0abd8d86a2052e7378274`. -Host: macOS 27.2.1, arm64, isolated signed AppKit fixture; no React renderer. -The helper used the stable `FURN Desktop Driver Development` identity, -artifact `73d2e52b2ad9fbd51410301399f3b784a2d17ae8c2b34ebb12b36cd9b41626a1`, -source digest `b0e40122cc2226ffa65486677af7ba2495e1138cbc2c7e70af22fd36671e796f`. -Evidence records the exact source revision and the three documentation-only -working-tree changes present during the run. - -The neutral `application-ownership` public entry provides exact macOS attach -identity and opt-in supervisor-owned launch; session launch stays the default. -Attached apps remain external. Both macOS launch kinds are explicitly -uncontained: LaunchServices causal ownership is not direct-child containment. -Pending and acquired applications are claim resources, including exact -PID/start and launch/containment metadata when available. Unknown publication -or cleanup outcomes remain fenced. `claimDesktopWorkload` is now asynchronous; -schema 3 uses the driver's OS start-time identity, and schema 2/unverified -identities inspect as unknown. Existing Storybook lease migration remains -deferred until after F9. - -All commands below ran under the exclusive native-validation lock with -project-local ignored `TMPDIR`; native commands additionally used -`FURN_NATIVE_DRIVER_TEST=1`, the stable signing identity, and an isolated helper -cache. Commands are shown from the repository root. - -| Selection | Exact command | Result | Workspace-relative artifact | -| ------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------ | -------------------------------------------------- | -| Full portable harness | `yarn workspace @fluentui-react-native/desktop-test test --runInBand` | Passed 399, failed 0, skipped 9; 25 suites passed, 1 skipped | `artifacts/desktop/f10m/harness-unit-final.log` | -| Native owner and rejected-launch safety | `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.contract --testNamePattern 'kernel owner identity\|failed native'` | Passed 3, failed 0, skipped 5 | `artifacts/desktop/f10m/native-negative-final.log` | -| Native identity and application exclusion backend | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand applicationIncarnation serviceIdentity` | Passed 69, failed 0, skipped 14; 4 suites passed, 2 skipped | `artifacts/desktop/f10m/native-identity-final.log` | - -The first full harness run had one missing compiled `desktop-driver-mcp` -boundary dependency. Building that unchanged workspace corrected the -prerequisite; all eight standalone-consumer boundary cases then passed, -including the new public entry, and the full harness rerun above passed. -Portable Windows/Win32 cases prove durable launched-app survivor records and -fail-closed claims only; they are not native packaged-survivor qualification. - -Native evidence: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-0fNQ8C/evidence.json`. -Schema 3 inspected its native OS owner as live. Both direct-child and -LaunchServices launches reached actual AX queries, which returned -`apiDisabled` (-25211), then journaled and settled their exact primary-process -identities. Direct child PID 8558 started at -`2026-11-12T02:37:23.794023Z`; LaunchServices PID 8616 started at -`2026-11-12T02:37:24.304837Z`. Native probes proved those incarnations exited; -an independent process check found both PIDs absent and no fixture/helper -process under this worktree. No external process was terminated. -Post-run verification is retained at -`artifacts/desktop/f10m/native-cleanup-final.json`. - -An earlier fixture inherited Mach-O minimum macOS 28.0 on this 27.2.1 host. -LaunchServices failed before publishing a PID -(`NSCocoaErrorDomain` 256, underlying `NSOSStatusErrorDomain` -10825); -that older implementation left the pending resource uncertain. Completed -no-application failures are settled by the review follow-up below. -The corrected fixture pins the helper's macOS 11.0 minimum in both compiler -target and bundle metadata; -`vtool` confirmed 11.0 and the corrected rejected-launch run passed above. -The earlier evidence is retained at -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-UiVcfS/evidence.json`. - -**Remaining qualification:** positive supervisor/session lifecycle, exact -external-app preservation and independent-service application exclusion are -blocked: Accessibility is not granted to the signed helper. Physical-input -mutex qualification is not run; PostEvent preflight is also false. Separate -macOS Fabric and Paper signing/TCC-responsibility lanes are not run on this -branch because of that permission block. No privacy settings were changed. -Arbitrary surviving descendants, interrupted native publication and live -Windows/Win32 uncontained packaged-survivor claims remain gates. The PR's -Windows driver-only native suite and ordinary E2E/smoke jobs do not exercise -the new harness survivor contracts; Windows CI is insufficient for those -gates. This is partial F10-M delivery, not full issue completion. - -### Pre-review repository validation - -At source `48717724f0df1e8f89d76b66c48d37fac93e871b`, -`yarn clean-lage-cache` passed once, followed by the locked aggregate: - -```sh -TMPDIR="$PWD/artifacts/desktop/f10m/buildci-tmp" \ -GIT_CEILING_DIRECTORIES="$PWD/artifacts/desktop/f10m/buildci-tmp" \ -TF_BUILD=1 yarn lage buildci --concurrency 4 -``` - -Result: passed, 176 successful tasks, 2 cache skips, 0 pending/aborted/failed. -`TF_BUILD=1` selects the declared Jest task's in-band runner; the Git ceiling -keeps project-local temporary test fixtures from discovering the outer -repository. No baseline tests or dependencies were changed for this setup. -An initial aggregate stopped at two owned cross-package guidance links; -replacing them with explicit code paths fixed the policy violation, and -`yarn check-agent-instructions` then passed. The corrected aggregate above -passed, including all owning workspace test/lint tasks. Package format/build -checks passed separately. `yarn lage test-links --concurrency 4` passed 90 -tasks with no failures. Logs: `artifacts/desktop/f10m/buildci-first.log`, -`artifacts/desktop/f10m/buildci-final.log` and -`artifacts/desktop/f10m/test-links-final.log`. Aggregate success does not -qualify the blocked or unrun native gates above. - -### Review follow-up: launch rejection and lock-failure settlement - -Source `62e6552028ea8fb81e76d430379321145fab709c` fixes the three review -findings. Native pre-spawn rejection and completed LaunchServices failure with -no application report `applicationSettlement: 'not-started'`. Ordinary no-data -macOS/Windows/Win32 rejections release the claim; reported process identity, -unknown or unrecognized settlement, lost dispatched responses and completion -after cancellation remain fenced. An acquired lease is never cleared by the -no-data fallback. The async `claimDesktopWorkload` export and awaited Storybook -caller are unchanged; desktop-test now has a **minor** changeset. - -Default-TMPDIR focused tests used only project-local test files and Jest cache. -Native commands ran under the exclusive lock, with the stable signing identity, -native opt-in, an isolated helper cache and project-local temporary storage: - -| Selection | Exact command | Result | -| ------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------- | -| Default-TMPDIR ownership | `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.test --cacheDirectory "$PWD/artifacts/desktop/f10m/jest-cache"` | Passed 24, failed 0, skipped 0 | -| Default-TMPDIR transport | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand NativeHostProcess --cacheDirectory "$PWD/artifacts/desktop/f10m/jest-cache"` | Passed 59, failed 0, skipped 0 | -| Signed helper rebuild | `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver build-driver --platform macos --force` | Passed | -| Native identity and termination | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand nativeDriver.test applicationIncarnation serviceIdentity` | Passed 79, failed 0, skipped 15; 5 suites passed, 2 skipped | -| AppKit safety | `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.contract --testNamePattern 'kernel owner identity\|failed native'` | Passed 4, failed 0, skipped 5 | -| Portable CLI retry, native-only environment removed | `yarn workspace @fluentui-react-native/desktop-test test --runInBand contracts --testNamePattern 'agent-cli\|lane-cli'` | Passed 2, failed 0, skipped 28 unselected | -| Full portable harness, native-only environment removed | `yarn workspace @fluentui-react-native/desktop-test test --runInBand` | Passed 409, failed 0, skipped 10; 25 suites passed, 1 skipped | - -The native self-test distinguishes actual flock contention from other errors. -Injected EBADF cleans up a real causally owned child and proves its exact exit; -injected EWOULDBLOCK preserves its real child. The AppKit safety case proves -claim release/reacquisition for exact-identity launch refusal, a missing direct -executable, a missing installed bundle, and a completed LaunchServices rejection. -Both AX-rejected launch kinds still prove exact-primary cleanup. - -New evidence: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-JMEGuu/evidence.json`; -it records a clean source tree. Helper artifact: -`1d68fae8df504432900e30e5b2a6714029762f5dd690cd953a8ef07b1f5da4f3`; -source digest: -`2c8ac56b2b824b05b1cb10d358050dfb643b44eb6e209508d53eaa0f2744b0f8`. -Direct-child PID 70283/start `2026-11-12T03:14:29.097020Z` and -LaunchServices PID 70323/start `2026-11-12T03:14:29.588927Z` were proved exited. -Independent post-run inspection found both PIDs absent and no own-worktree -fixture/helper process. Logs: -`artifacts/desktop/f10m/review-focused-final.log`, -`artifacts/desktop/f10m/review-native-final.log` and -`artifacts/desktop/f10m/review-native-cleanup.json`. -An extra full portable run in the native batch initially passed 407, failed 2, -skipped 10 because the HOME-isolating CLI fixtures inherited native-only -signing/cache settings and could not resolve the configured signing identity. -Removing `FURN_NATIVE_DRIVER_TEST`, -`FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY` and -`FURN_DESKTOP_DRIVER_CACHE_ROOT` corrected the invocation, without a source or -fixture change. Both affected cases and then the full portable suite passed -under the exclusive lock; log: -`artifacts/desktop/f10m/review-harness-final.log`. -These safety passes do not close any positive AX, physical-input, renderer, -descendant, interrupted-publication or Windows-native gate above. - -### Final review-fix repository validation - -At clean source `e74043b0e6eb5b452e9023bc60c70ba6c9540600`, one combined -clean-cache aggregate passed under the exclusive native-validation lock: - -```sh -env -u FURN_NATIVE_DRIVER_TEST \ - -u FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY \ - -u FURN_DESKTOP_DRIVER_CACHE_ROOT \ - TMPDIR="$PWD/artifacts/desktop/f10m/review-buildci-tmp" \ - GIT_CEILING_DIRECTORIES="$PWD/artifacts/desktop/f10m/review-buildci-tmp" \ - TF_BUILD=1 \ - bash -c 'yarn clean-lage-cache && yarn lage buildci --concurrency 4' -``` - -Result: **passed, 178 successful tasks, 0 skipped, 0 pending/aborted/failed**. -All owning workspace aggregate tasks passed; native-only overrides were not -inherited. `TF_BUILD=1` retains the supported in-band Jest task, and the Git -ceiling isolates project-local temporary Git fixtures. No source, dependency or -test waiver was needed. Log: -`artifacts/desktop/f10m/review-buildci-final.log`. -The code remains the native-qualified `62e6552028ea8fb81e76d430379321145fab709c`; -subsequent changes record evidence only. Positive native acceptance gates remain -open, so delivery is draft and does not close #4446. - -### CI regression follow-up: read-only application identity - -CI run `43148275374` at `0bd0ad6f7cc1e862169a422b7614a95adf899584` -failed both macOS Storybook renderers with `application-lock-busy` while -recording the authenticated bridge epoch. The actual WebDriver session's -native helper held its legitimate exact-incarnation application lock. -Storybook's recovery callback ignored that helper and created another -`NativeDesktopHost` whose temporary `attach` requested competing control. -This was not a workload-claim leak or an ownership handoff. -The passing macOS E2E artifact reports 62 tests with session ownership: -`desktop-test/src/wdio/configuration.ts` selects `launch`, not the Storybook -attach/recovery path, and that lane has no Storybook bridge-epoch journal. - -Fix source `5e360734e9b85bd4732aabc47c7654bb7211c3d3` introduces the -macOS-only read-only `inspectApplication` operation. It shares descriptor, -legacy nonce/lease, executable/bundle and exact-window verification, capturing -the kernel identity before discovery and rechecking it afterwards. It does -not acquire control or register windows/elements. Storybook uses the existing -session helper; every real launch/attach retains exclusive exact-incarnation -control, and foreign controlling services still receive -`application-lock-busy`. Windows/Win32 keep their previous isolated identity -attachment path. Unsupported old helpers fail closed rather than falling -back to controlling macOS attachment. - -All three owning packages passed their declared format, build and lint checks. -Build/test/lint batches and native qualification used the exclusive -native-validation lock. Focused tests kept the default TMPDIR unchanged; -their fixtures and Jest caches are project-local. The full -suites instead used project-local TMPDIR and the matching Git ceiling because -existing broad fixtures use `os.tmpdir()` and this task forbids system -temporary-directory writes. This is a declared deviation from the requested -default-TMPDIR full-suite invocation, not default-TMPDIR full-suite evidence. - -| Selection | Exact command | Result | -| -------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | -| Default-TMPDIR native adapter | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand NativeDesktopHost.applicationIdentity --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-driver"` | Passed 11, failed 0, skipped 0 | -| Default-TMPDIR composed Storybook regression | `yarn workspace @fluentui-react-native/storybook-desktop test --runInBand applicationIdentity --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-storybook"` | Passed 5, failed 0, skipped 0 | -| Default-TMPDIR claim/ownership regressions | `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.test ownership.test --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-harness"` | Passed 29, failed 0, skipped 0 | -| Full portable driver | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-desktop-driver-full"` | Passed 1,236, failed 0, skipped 66 | -| Full portable harness | `yarn workspace @fluentui-react-native/desktop-test test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-desktop-test-full"` | Passed 409, failed 0, skipped 10 | -| Full portable Storybook | `yarn workspace @fluentui-react-native/storybook-desktop test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-storybook-desktop-full"` | Passed 344, failed 0, skipped 4 | -| Stable-signed helper rebuild | `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver build-driver --platform macos --force` | Passed | -| Noninteractive permissions | `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver doctor --platform macos --permissions` | Passed diagnostics; Accessibility/PostEvent preflight false, Screen Recording true | -| Native identity/termination/backend | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand nativeDriver.test applicationIncarnation serviceIdentity --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-driver-native"` | Passed 79, failed 0, skipped 15 | -| AppKit safety | `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.contract --testNamePattern 'kernel owner identity\|failed native' --cacheDirectory "$PWD/artifacts/desktop/f10m/ci-regression/jest-harness-native"` | Passed 4, failed 0, skipped 5 | - -The composed regression covers launch, application claim publication, -embedded-driver attach and authenticated epoch observation separately for -Fabric and Paper, with retained window identity and rejected foreign control. -The native self-test additionally refuses missing, malformed and stale -inspection descriptors before AX discovery. The positive AppKit contracts now -inspect an application while its controlling session and foreign helper are -present, but those AX-dependent cases remain locally blocked, not passed. - -The stable-signed native run used a clean source tree. Helper artifact: -`982f3cfa696bbde7f9868186ae9365d59014fd656687e5cbec2de74863280ed3`; -source digest: -`6bafaf5835d9cd22410992466a8d8f86086ff6b3b3efd182b3a48922f698527c`. -AppKit evidence: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-mwo1d2/evidence.json`. -Rejected direct-child PID 53320/start `2026-11-12T04:18:28.787757Z` and -LaunchServices PID 53360/start `2026-11-12T04:18:29.263568Z` were proved exited; -independent post-run inspection found both PIDs absent and no own-worktree -fixture/helper process. No privacy settings were changed. - -Logs and curated CI analysis are under -`artifacts/desktop/f10m/ci-regression/`: `focused-default-tmp.log`, -`package-validation.log`, `native-validation.log`, `native-cleanup.json` and -`ci-analysis.json`. Actual macOS Storybook Fabric/Paper CI requalification at -the fix is pending. Local positive AX/physical input, descendant, interrupted -publication and Windows-native acceptance gates above remain open; keep draft. - -#### Final CI-regression aggregate - -The first clean-cache aggregate at documentation revision -`a7c37b86abdfc3847824480300315cf88eebaf42` failed the Storybook test task -after 112 successful tasks. Its output stopped after 13 passing Jest files, -without an assertion, stack, summary or termination cause. The cause remains -unestablished. The unchanged full Storybook suite then passed 344 with 4 -skipped under the identical aggregate environment; that reproduction is not -a successful aggregate. - -A second clean-cache aggregate serialized package tasks to remove -inter-package fixture/process overlap. Under the exclusive native-validation -lock, the following invocation passed **178 tasks**, with 0 skipped, failed, -pending or aborted, in 6m 12.39s: - -```sh -env -u FURN_NATIVE_DRIVER_TEST \ - -u FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY \ - -u FURN_DESKTOP_DRIVER_CACHE_ROOT \ - TMPDIR="$PWD/artifacts/desktop/f10m/ci-regression/buildci-tmp" \ - GIT_CEILING_DIRECTORIES="$PWD/artifacts/desktop/f10m/ci-regression/buildci-tmp" \ - TF_BUILD=1 \ - bash -c 'set -euo pipefail; git rev-parse HEAD; git status --short; yarn clean-lage-cache && yarn lage buildci --concurrency 1' -``` - -No source or test waiver was applied between these runs. Serialization is -the retry's isolation measure, not a proven diagnosis of the first failure. -Logs: `artifacts/desktop/f10m/ci-regression/buildci.log`, -`artifacts/desktop/f10m/ci-regression/storybook-buildci-environment.log` -and `artifacts/desktop/f10m/ci-regression/buildci-final.log`. -Runtime source remains the native-qualified `5e360734e`; subsequent changes -record evidence only. This is portable repository validation, not positive -AX or renderer qualification. Actual macOS Storybook Fabric/Paper CI and the -other native acceptance gates remain pending; keep draft. - -### 2026-11-12 AX-enabled F10-M qualification - -**Observed.** This evidence supersedes the earlier Accessibility/PostEvent -block, not the remaining descendant/publication gates. Source is -`9395737943d7e1402a40a67a41dc605ef0afb24b` plus this delivery's live -macOS discovery fix. Host: macOS 27.2.1 arm64. Feasibility was recorded before -execution at -`artifacts/desktop/f10m/ax-qualification-20261112/feasibility.json`. -Every native, GUI, physical-input and aggregate command used -`$HOME/Library/Caches/furn-overnight/with-native-lock.sh`. -The default TMPDIR was retained for this requalification; the older -project-local-TMPDIR qualification remains historical. - -The requested diagnostic passed: - -```sh -env FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ - FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ - FURN_DESKTOP_DRIVER_BUILD_POLICY=never \ - yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver \ - --reporter json doctor --platform macos --permissions -``` - -Accessibility, PostEvent and ScreenCapture preflight were `true`; the screen -was unlocked/on-console and `promptRequested=false`. Stable signed helper: -`1429ce4663795bad31f545086ef80bffb83e9b9533287964991c5527388b2f6c`; -source digest -`c17944691081d30a320cf6e4393a5bb0ce6cece8dd71295a400dbc5e6a53b3cb`. -The certificate leaf and designated requirement match the driver's evidence -above. Process ancestry reaches `Rex Server.app`; the effective permission -result is scoped to that responsible process tree. **Inferred, not qualified:** -these passes do not establish that the helper bundle independently owns a -durable grant or that another terminal/CI process tree inherits it. - -The initial full AppKit run passed 8 and failed 1. Its newly launched app was -missing from the standalone helper's cached `NSWorkspace` application list. -The new warmed-discovery regression failed before the native fix (1 failed, -9 skipped). Live descriptor discovery now uses current-user kernel PID -enumeration with fresh `NSRunningApplication` objects; exact identity, -ambiguity refusal and launch causality remain enforced. The final focused run -passed **39, failed 0, skipped 0**: - -```sh -env FURN_NATIVE_DRIVER_TEST=1 FURN_NATIVE_PHYSICAL_INPUT_TEST=1 \ - FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ - FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ - yarn workspace @fluentui-react-native/desktop-test test --runInBand \ - applicationOwnership.contract applicationOwnership.test ownership.test \ - --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-native-harness" -``` - -Ten native tests cover positive supervisor/session lifecycle, exact OS owner -identity, external preservation, independent-service application exclusion -and lease release, read-only observation with retained windows, warmed -discovery and duplicate-bundle refusal, stale identity, direct-child and -LaunchServices causality, cancellation, rejected launch cleanup, and the -physical-input mutex. The remaining 29 are portable regressions. Physical -input was limited to the mutex's Shift down/up sequence. - -During implementation, two experimental LaunchServices attempts could not -publish causal ownership. They were not silently treated as owned or exited. -The run-specific executable paths, native start identities and launch-time -intervals were checked before exact operator cleanup; native probes then -reported both exited (PIDs 63263 and 65041). The experimental change to launch -preexistence discovery was reverted; only descriptor discovery is changed. -Evidence: -`orphan-native-launch-cleanup.json`, `orphan-physical-launch-cleanup.json`, -the red/experimental/green logs and `discovery-probe.json` under -`artifacts/desktop/f10m/ax-qualification-20261112/`. -Final passing AppKit identities and cleanup: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-1V0SrR/evidence.json`. - -#### Real Storybook renderer procedure and boundaries - -The ignored qualifier -`artifacts/desktop/f10m/ax-qualification-20261112/qualify-renderer.mjs` -uses the declared app commands below from `apps/storybook`. Fabric omits -`--paper`; Paper adds it to every renderer-specific command. Preparation is -rerun when changing renderer: - -```sh -yarn storybook --reporter json instance --macos -yarn storybook --reporter json bundle --macos -yarn storybook --reporter json prep --macos -yarn storybook --reporter json manifest --macos -yarn storybook --reporter json driver --macos --input accessible -yarn storybook --reporter json run --macos -STORYBOOK_SMOKE_STORY=components-button--default \ - yarn storybook --reporter json smoke --macos --mode stories-and-tests --input accessible -``` - -Only `run` applies the worktree instance bundle ID. Before restarting the -driver, the qualifier quits the exact app it launched and closes the old -driver/Metro/channel. Its observer is started before app launch to exercise -warmed discovery. Startup polling retries only read-only missing/incomplete -AX discovery, never launch, attach or another mutation. - -The real-app checks use public `DesktopSupervisor`, native-host, identity, -claim-inspection and driver-client APIs. Two independent services have -separate claim directories and native helpers. The first attaches; both -helpers can inspect its exact app identity without changing window handles. -The second's native controlling attach returns `application-lock-busy`, and -its WebDriver new-session request is refused. Releasing the first session -lets the second attach. Deleting sessions and closing both external owners -preserves the app. A separate supervisor-owned launch then survives two -borrowed sessions, publishes its exact owned application resource and exits -only at supervisor disposal. Each neutral claim verifies schema 3, native -owner start identity, registered target and target-instance IDs; its claim is -absent after settled close. - -**Observed integration limit:** Storybook's outer CLI claim still uses -`unregistered` target/instance placeholders and no application resource -journal. Its OS owner identity and run ID are verified separately. This -qualification does not pretend that the deferred Storybook lifecycle/lease -relocation after F9 has happened; the exact registered application-resource -assertions are on the neutral supervisor claims. - -The first Fabric prep attempt failed before any app/service launch because -the installed Ruby 4.0.6 selected JSON 3.1.0, which lacks the `fast_generate` -entry used by this CocoaPods toolchain. Selecting the already-installed -JSON 2.18.0 through a project-local Ruby preload fixed prep. The preload -contains only `gem 'json', '=2.18.0'`; no gem installation, generated-source -patch, Podfile change or global Ruby configuration change was made. -Renderer invocations use: - -```sh -env RUBYOPT="-r$PWD/artifacts/desktop/f10m/ax-qualification-20261112/ruby-json-version.rb" \ - FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ - FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ - FURN_DESKTOP_DRIVER_BUILD_POLICY=never \ - node artifacts/desktop/f10m/ax-qualification-20261112/qualify-renderer.mjs fabric -``` - -For Paper, the final argument is `paper`. The machine-wide native-lock -wrapper encloses each complete invocation, including cleanup. The qualifier -selects Button/Default by manifest title/name, not an assumed slug. -`STORYBOOK_SMOKE_STORY` filters authored callbacks, not catalog traversal. -The accessible lane executes the Button semantics callback; its declared -physical-pointer callback is excluded at selection, not counted as passed -or skipped. - -Failed exploratory Fabric attempts remain separate artifacts: -`fabric-prep-json3-failure/`, `fabric-orchestrator-scope-failure/`, -`fabric-ax-startup-failure/` and `fabric-orchestrator-refusal-failure/`. -The latter three corrected qualifier assumptions: the transitional outer -claim is not a registered neutral claim; initial AX queries can be incomplete; -and native conflict metadata is not the WebDriver error shape. They did not -change production behavior. The scope failure required scoped command-tree -shutdown; its three ports and claim were then verified absent in -`driver-exact-cleanup.json`. The startup failure occurred before the qualifier -captured a lease identity, so its original zero cleanup-failure count is not -proof that its app exited. Operator cleanup separately checked the unique -native PID, worktree executable, instance bundle ID and start within the -successful `run` interval before signaling that exact app; the native probe -reported exited in `run-candidate-exact-cleanup.json`. No unknown process or -foreign application was adopted or killed. - -#### Observed renderer results - -| Renderer | Qualification checks passed / failed / skipped | Smoke selected / executed / passed / failed / skipped / unsupported | Smoke artifact | -| ------------ | ---------------------------------------------- | ------------------------------------------------------------------- | -------------------------------------------------------------------------------- | -| macOS Fabric | 9 / 0 / 0; cleanup failures 0 | 224 / 224 / 224 / 0 / 0 / 0; retries 0 | `apps/storybook/artifacts/desktop/616bf2e8-a675-408f-bd56-2e09982ea969/run.json` | -| macOS Paper | 9 / 0 / 0; cleanup failures 0 | 224 / 224 / 224 / 0 / 0 / 0; retries 0 | `apps/storybook/artifacts/desktop/19358539-eed2-49e9-9c9a-79b5e7db4e9f/run.json` | - -Each smoke traversed 223 catalog entries plus the one manifest-selected -Button semantics callback. The accessible input policy excluded one physical -callback before selection; no physical callback is included in those passes. -Both renderer doctors observed the same stable signed helper artifact and -effective permissions recorded above. This qualifies the observed Fabric and -Paper signing/TCC responsibility path separately; it is not first-run, -clean-machine, PPPC or revoked-permission qualification. - -| Renderer | Externally attached run-owned app, PID / native start | Supervisor-owned app, PID / native start | Exact cleanup | -| -------- | ----------------------------------------------------- | ---------------------------------------- | ------------------------------------------------------------------------------------------------------------------ | -| Fabric | 82677 / `2026-11-12T10:56:35.279027Z` | 83427 / `2026-11-12T10:56:38.528966Z` | External app preserved through both external owners, then explicitly quit; supervisor app exited at owner disposal | -| Paper | 93753 / `2026-11-12T11:03:02.610130Z` | 94499 / `2026-11-12T11:03:04.921261Z` | Same independently observed preservation and exact disposal | - -Both renderer runs closed driver/channel/Metro ports 33637, 18878 and 38843 -before restart and after smoke, removed all neutral and outer CLI claims, -and found the instance application absent after smoke. Raw checks, identities, -claims, process ancestry, signing/preflight and command logs: -`artifacts/desktop/f10m/ax-qualification-20261112/fabric/evidence.json` -and -`artifacts/desktop/f10m/ax-qualification-20261112/paper/evidence.json`. - -**Still not run/unqualified:** arbitrary surviving descendants, interrupted -native publication, supervisor-death survivors and live failed-cleanup/restart -fault injection; native Windows/Win32 packaged-survivor journaling/recovery; -and the repeated-run reliability window. First-run durable helper-owned -TCC grants, clean-machine/PPPC/revocation behavior and another responsible -process tree are not inferred from these passes. Storybook lifecycle/lease -relocation remains deferred after F9. **Windows CI sufficient: no; keep draft.** -Apple/Office module and bridging-header checks are inapplicable to this -standalone helper. - -#### Final package validation and scope correction - -**Observed.** All three owning packages passed their declared -`yarn workspace @fluentui-react-native/ format --check`, `build` -and `lint` commands, for `desktop-driver`, `desktop-test` and -`storybook-desktop`. The harness build/format/lint were repeated after its -regression-fixture refinement. Command log: `package-validation.log` in the -qualification artifact root. - -The initial broad native-enabled batch passed the full driver suite, then -failed the harness suite: **416 passed, 3 failed, 1 skipped**. Two portable -CLI fixtures intentionally substitute HOME, but inherited the native signing -override, making the signing identity unavailable in that isolated home. -Those are not portable results without native environment overrides. -The third failure was a LaunchServices acquisition in the discovery test's -setup: the helper could not prove causal incarnation, returned -`applicationSettlement=unknown` with PID 9037, and did not kill a guessed -process. Its precise cause remains unestablished. It is not a successful -launch or a resolved reliability gate. - -Operator cleanup verified the unique run-owned fixture executable, exact -kernel start `2026-11-12T11:12:39.767964Z` and native-suite interval before -signaling PID 9037; the native probe then proved exited. Evidence: -`native-suite-candidate-cleanup.json` and -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-Kd4tXB/evidence.json`. -The discovery-only regression now starts both external children directly, -captures their native identities, and waits for exact attached windows. -It still warms discovery before launch, observes through another controlling -helper, refuses duplicate bundle identities and verifies disappearance. -This removes unrelated LaunchServices setup from that regression; it neither -fixes nor weakens causal ownership capture. Existing supervisor/causal-launch -tests remain selected. - -The corrected validation keeps the signing/native opt-ins only on matching-OS -native selections and clears them for portable full suites. Every command -below ran under the native lock with the default TMPDIR unchanged: - -| Selection | Exact command | Passed / failed / skipped | -| -------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------- | -| Full signed native-enabled driver | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-desktop-driver-full"` | 1,255 / 0 / 47 | -| Final signed ownership and regressions | `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.contract applicationOwnership.test ownership.test --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-native-harness"` | 39 / 0 / 0; all 10 native tests selected | -| Full portable driver | `yarn workspace @fluentui-react-native/desktop-driver test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-desktop-driver-portable"` | 1,236 / 0 / 66 | -| Full portable harness | `yarn workspace @fluentui-react-native/desktop-test test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-desktop-test-portable"` | 409 / 0 / 11 | -| Full portable Storybook | `yarn workspace @fluentui-react-native/storybook-desktop test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/ax-qualification-20261112/jest-storybook-desktop-portable"` | 344 / 0 / 4 | - -Signed/native rows use `FURN_NATIVE_DRIVER_TEST=1`, the signing identity and -helper cache above; the ownership row additionally enables -`FURN_NATIVE_PHYSICAL_INPUT_TEST=1`. Portable rows explicitly unset -`FURN_NATIVE_DRIVER_TEST`, `FURN_NATIVE_PHYSICAL_INPUT_TEST`, -`FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY`, -`FURN_DESKTOP_DRIVER_CACHE_ROOT` and `FURN_DESKTOP_DRIVER_BUILD_POLICY`. -The 11 portable harness skips include all 10 opt-in AppKit tests; their -separate signed pass is not inferred from that portable run. - -Logs: `full-native-suites.log` (includes the failed broad attempt) and -`validation-final.log` (all corrected selections passed), under -`artifacts/desktop/f10m/ax-qualification-20261112/`. -Final AppKit evidence: -`packages/agentic/desktop-test/artifacts/desktop/application-ownership-1V0SrR/evidence.json`, -with the unchanged signed helper/source digest qualified by both renderers. -**Still unqualified:** reliable LaunchServices causal-identity acquisition -when the native snapshot cannot be proven; the unknown-publication/survivor -and other gates listed above remain open. No automatic reclaim is claimed. - -#### Final aggregate and cleanup - -**Observed.** The single final clean-cache aggregate ran under the exclusive -native-validation lock, without changing the default TMPDIR: - -```sh -"$HOME/Library/Caches/furn-overnight/with-native-lock.sh" \ - env -u FURN_NATIVE_DRIVER_TEST -u FURN_NATIVE_PHYSICAL_INPUT_TEST \ - -u FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY \ - -u FURN_DESKTOP_DRIVER_CACHE_ROOT -u FURN_DESKTOP_DRIVER_BUILD_POLICY \ - TF_BUILD=1 bash -c 'set -euo pipefail; -node -e '\''console.log(JSON.stringify({tmpdir:require("node:os").tmpdir(),nativeOptIn:process.env.FURN_NATIVE_DRIVER_TEST??null}))'\''; -yarn clean-lage-cache && yarn lage buildci --concurrency 1' -``` - -**Passed 178 tasks, failed 0, skipped 0, pending 0, aborted 0**, in -6m 50.47s. The recorded Node snapshot confirms the default macOS temporary -directory and no native opt-in. The portable aggregate does not replace the -separate signed native or real-renderer qualification. Full output: -`artifacts/desktop/f10m/ax-qualification-20261112/buildci-final.log`. - -After the aggregate, the final read-only cleanup verification ran: - -```sh -"$HOME/Library/Caches/furn-overnight/with-native-lock.sh" \ - env FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ - FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ - FURN_DESKTOP_DRIVER_BUILD_POLICY=never \ - node artifacts/desktop/f10m/ax-qualification-20261112/verify-cleanup.mjs -"$HOME/Library/Caches/furn-overnight/with-native-lock.sh" true -``` - -**Passed:** signed public native probes reported all **31** recorded exact -incarnations exited; all three reported driver/channel/Metro ports refused -connections; scoped current-user command-line and working-directory -inspection found **0** worktree workload processes. Only the verifier, its -command ancestors and its `lsof` child were excluded. The shared workload -claim was absent, with **0** remaining `interactive.lock` files in eight -native/renderer/failed-attempt artifact roots. Neither unknown nor foreign -processes were signaled. The locked verifier exited 0; the following locked -`true` successfully reacquired and released the machine-wide lock. - -Evidence: `cleanup-final.json` and `cleanup-final.log` under -`artifacts/desktop/f10m/ax-qualification-20261112/`. The qualifier, installed-gem -selector and final verifier remain as ignored execution artifacts; the -experimental Swift diagnostic and completed operator-cleanup scratch scripts -were removed, preserving their JSON/log evidence. No helper, fixture, app, -driver service or Metro from this worktree remains. The causal-acquisition, -survivor, containment, interrupted-publication and TCC/reliability limitations -above are unchanged; keep the PR draft. - -Final documentation checks passed: - -| Exact command | Observed result | -| ---------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------ | -| `yarn markdown-link-check -c .github/markdown-link-check-config.json packages/agentic/desktop-driver/native/macos/README.md` | Passed 12 links | -| `yarn markdown-link-check -c .github/markdown-link-check-config.json packages/agentic/desktop-driver/references/current-status.md` | Passed 6 links | -| `yarn markdown-link-check -c .github/markdown-link-check-config.json packages/agentic/desktop-test/references/current-status.md` | Passed 15 links | -| `yarn markdown-link-check -c .github/markdown-link-check-config.json .changeset/neutral-macos-application-ownership.md` | Passed; 0 links | -| `yarn lage test-links --verbose --grouped` | Passed 90 tasks, failed 0, skipped 0 | -| `yarn oxfmt --check` | Passed repository-wide, 3,159 files | - -Logs in the qualification artifact root: -`changed-markdown-links-final.log`, `test-links-final.log` and -`root-format-final.log`. `git diff --check` also passed. Only documentation -result capture followed the final aggregate; no native/runtime source -changed after the separately qualified renderer builds. - -### Rebased F10-M publication and survivor qualification — 2026-11-12 - -The macOS layer was rebased in place from -`6489386956c0485cb91b478a8f4adc4828f4178b` onto main -`041ac8b9894aeaa7e1aa273baa2ece4e46298b94` (F6), yielding -`d58d473b7bbef26619db596f420fa837687badf3`. The awaited Storybook claim -was already present in F6; conflict resolution retained it and both independent -evidence ledgers. No Windows native source was changed. -`yarn install --immutable` passed with existing peer-dependency warnings. - -Qualification-only commit -`d5a7530a9f2d4cb78721a36c1109b5c36fbdc5a1` adds three real AppKit -scenarios; production ownership and helper behavior are unchanged: - -- Acquired-resource publication throws after native launch. The complete - pending claim remains fenced, refuses release/replacement, and settles only - after exact native cleanup and successful publication. -- An independent owner is killed with `SIGKILL` after native acquisition but - before acquired-identity publication. The durable pending record survives; - no guessed identity is adopted, and replacement remains refused. -- Another independent owner is killed after acquired-identity publication. - Its native app and known `/bin/sleep` descendant survive owner/helper exit. - Read-only inspection verifies the surviving app. The descendant also - survives exact primary-process cleanup, confirming `uncontained` semantics. - Both crash tests verify native owner/helper/app/descendant identities and - the initial child-parent relationship before exact operator cleanup. Only - the unchanged fixture claim is then removed. - -**Observed:** macOS 27.2.1 arm64, stable signer -`FURN Desktop Driver Development`. All native operations ran under the shared -native lock, with inherited TMPDIR. Nonprompting doctor passed with -Accessibility/PostEvent/ScreenCapture preflight true and unlocked/on-console. -The helper artifact and source digest remain -`1429ce4663795bad31f545086ef80bffb83e9b9533287964991c5527388b2f6c` -and `c17944691081d30a320cf6e4393a5bb0ce6cece8dd71295a400dbc5e6a53b3cb`. - -The focused command below passed **42 tests: 13 native, 29 portable; no failures -or skips**. It ran the exact fixture/test content subsequently committed above. - -```sh -FURN_NATIVE_DRIVER_TEST=1 FURN_NATIVE_PHYSICAL_INPUT_TEST=1 \ -FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development" \ -FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f10m/helper" \ -yarn workspace @fluentui-react-native/desktop-test test --runInBand \ - applicationOwnership.contract applicationOwnership.test ownership.test \ - --cacheDirectory "$PWD/artifacts/desktop/f10m/rebased-qualification-20261112/jest-ownership" -``` - -Evidence: `packages/agentic/desktop-test/artifacts/desktop/application-ownership-kffNMN/evidence.json`. -Raw command output is under -`artifacts/desktop/f10m/rebased-qualification-20261112/`. - -Full package validation also passed: - -| Exact command | Observed result | -| ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-driver format --check` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver lint` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver build` | Passed | -| `yarn workspace @fluentui-react-native/desktop-test format --check` | Passed | -| `yarn workspace @fluentui-react-native/desktop-test lint` | Passed | -| `yarn workspace @fluentui-react-native/desktop-test build` | Passed | -| `yarn workspace @fluentui-react-native/storybook-desktop format --check` | Passed | -| `yarn workspace @fluentui-react-native/storybook-desktop lint` | Passed | -| `yarn workspace @fluentui-react-native/storybook-desktop build` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/rebased-qualification-20261112/jest-driver-native"` | Passed 1,256; failed 0; skipped 47. Native opt-in, stable signing and the helper cache above were set for this command only. | -| `yarn workspace @fluentui-react-native/desktop-test test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/rebased-qualification-20261112/jest-harness"` | Passed 409; failed 0; skipped 14. Native/signing overrides absent; the 13 AppKit cases passed separately. | -| `yarn workspace @fluentui-react-native/storybook-desktop test --runInBand --cacheDirectory "$PWD/artifacts/desktop/f10m/rebased-qualification-20261112/jest-storybook"` | Passed 392; failed 0; skipped 4. Native/signing overrides absent. | - -The native driver suite includes exact-incarnation probes, refusal of -unreadable foreign process identity, stale-start/PID-reuse classification and -native self-tests. Stale timestamps supplied against live PIDs and synthetic -microsecond-reuse self-tests are not an observation of forced OS PID recycling. -The physical ownership mutex uses only bounded Shift down/up. - -**Remaining scope:** the crash fixtures prove a bounded known descendant, -not arbitrary descendant discovery or containment. They interrupt acquisition -publication, not a machine power loss during filesystem rename. Actual TCC -revocation, helper-owned durable grants outside the observed responsible -process tree, automatic survivor adoption, Metro/channel survivor topology, -transparent restart and repeated-run reliability remain unqualified. -The earlier unverifiable LaunchServices causal snapshot is not erased by these -passing launches. F9 lifecycle relocation and F11 retained-app handoff are -separate work; this change implements neither. - -#### Rebased Fabric and Paper consumer passes - -Both renderers were independently rebuilt/prepared and qualified at -`ca88c289c0fbd79b03dbd43b9bd85f0b21e7e800` (documentation after the -qualification-only commit). Each passed **all 9 ownership/lifecycle checks**, -with **0 cleanup failures**. Their declared accessible smokes each passed -**224, failed 0, skipped 0, unsupported 0, retries 0**: 223 catalog traversals -and one manifest-selected Button semantics callback. One physical callback -was excluded before selection, not passed or skipped. - -Commands ran from `apps/storybook`, separately for Fabric and Paper, under the -native lock. The helper signing/cache settings above were supplied. -The existing project-local Ruby preload selected installed JSON 2.18.0 for -CocoaPods; no dependencies, privacy settings, native project sources or -global Ruby settings were changed. - -| Exact command | Observed result | -| --------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------- | -| `yarn storybook --reporter json instance --macos` | Passed | -| `yarn storybook --reporter json bundle --macos` | Passed | -| `yarn storybook --reporter json prep --macos` | Passed | -| `yarn storybook --reporter json manifest --macos` | Passed | -| `yarn storybook --reporter json driver --macos --input accessible` | Passed readiness and settled shutdown | -| `yarn storybook --reporter json run --macos` | Passed native build and exact instance launch | -| `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook --reporter json smoke --macos --mode stories-and-tests --input accessible` | Passed 224/224 | -| `yarn storybook --reporter json instance --macos --paper` | Passed | -| `yarn storybook --reporter json bundle --macos --paper` | Passed | -| `yarn storybook --reporter json prep --macos --paper` | Passed | -| `yarn storybook --reporter json manifest --macos --paper` | Passed | -| `yarn storybook --reporter json driver --macos --paper --input accessible` | Passed readiness and settled shutdown | -| `yarn storybook --reporter json run --macos --paper` | Passed native build and exact instance launch | -| `STORYBOOK_SMOKE_STORY=components-button--default yarn storybook --reporter json smoke --macos --paper --mode stories-and-tests --input accessible` | Passed 224/224 | -| `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver --reporter json doctor --platform macos --permissions` | Passed separately for each renderer; no prompt | - -Real neutral supervisor-owned apps survived two borrowed sessions and exited -only at supervisor disposal. Externally attached apps survived both independent -owners' cleanup. Cross-helper controlling attach returned -`application-lock-busy`; independent WebDriver control was refused until the -first lease was released. Read-only inspection succeeded while control was -held and preserved window handles. Schema-3 claims carried verified native -owner/application and registered target/instance identities. F6 provenance -behavior remained intact. Storybook's outer CLI claim still has its documented -transitional placeholders; this is not F9 integration. - -Evidence: `fabric/evidence.json` and `paper/evidence.json` under the new -qualification artifact root. Smoke reports: -`apps/storybook/artifacts/desktop/15518d7a-d662-49ae-a76d-f4c0cd5e2fc3/run.json` -(Fabric) and -`apps/storybook/artifacts/desktop/92d22b26-24cc-4d87-a4b7-229d24f478b1/run.json` -(Paper). Both observed the same stable signer/designated requirement with -effective permissions under the `Rex Server.app` responsible process tree; -this is not first-run, clean-machine or permission-revocation qualification. -Apple/Office module and bridging-header checks are inapplicable to this -standalone helper. Windows/Win32 native ownership qualification belongs to -the independent upper layer and was not run here. - -#### Plain aggregate, documentation and final cleanup - -The required aggregate **is not green**. It ran under the native lock with -inherited environment: `TF_BUILD` and native/signing overrides absent, -default TMPDIR retained, and no concurrency override. The cache was cleared -once; subsequent plain attempts used only results from that fresh run. - -| Exact command | Observed result | -| ----------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| `yarn lage test-links --verbose --grouped` | Passed 90 tasks, failed 0, skipped 0 | -| `yarn clean-lage-cache && yarn lage buildci` | Failed in `check-package-contents`: desktop-test's `yarn pack --dry-run --json` exited 1; underlying cause unestablished. Log: `buildci.log`. | -| `yarn workspace @fluentui-react-native/desktop-test pack --dry-run --json` | Passed immediately without changes. Log: `desktop-test-pack.log`. | -| `yarn lage buildci` (first retry) | Failed driver tests: known `boundedCommand` parallel-startup assertion expected SIGKILL but observed SIGTERM; unmodified `privateStorage` test also exceeded 10 seconds. Log: `buildci-retry.log`. | -| `yarn workspace @fluentui-react-native/desktop-driver test --runInBand boundedCommand privateStorage` | Passed 73, failed 0, skipped 10. This focused serial diagnostic is not aggregate acceptance. Log: `aggregate-failures-focused.log`. | -| `yarn lage buildci` (final unchanged retry) | Failed only the same `boundedCommand` assertion; driver tests passed 1,236, failed 1, skipped 66. The private-storage test passed. Log: `buildci-final.log`. | - -Both failing test sources and their implementations are unchanged from pinned -main. No assertion, timeout, concurrency policy or environment was altered to -hide these failures. The passing full signed suite, focused contracts and -renderer runs remain distinct evidence; they do not turn the failed plain -aggregate into a pass. - -Changed Markdown link checks passed using -`yarn markdown-link-check -c .github/markdown-link-check-config.json `: -`packages/agentic/desktop-driver/native/macos/README.md` (12 links), -`packages/agentic/desktop-driver/references/current-status.md` (6), -`packages/agentic/desktop-test/references/current-status.md` (19), -`packages/agentic/desktop-test/references/native-qualification.md` (2), and -`.changeset/neutral-macos-application-ownership.md` (0). - -**Observed final cleanup:** native probes proved all **24** recorded exact -incarnations exited; all **3** reported driver/channel/Metro ports refused -connections. Scoped command-line and working-directory checks found **0** -worktree workload processes. The shared claim and all claims beneath the -three native/renderer evidence roots were absent. The native lock was then -reacquired and released successfully. - -An earlier cleanup scan saw four orphaned test workers/children from the -failed aggregate. All four exited naturally before the bounded cleanup -inspection; no signal was needed. Their initial failure and subsequent -absence are retained in `cleanup-first.json`, `cleanup-first.log` and -`aggregate-worker-cleanup.json`. The final successful verifier wrote -`cleanup-final.json` and `cleanup-final.log` in the new artifact root. -No helper, fixture, descendant, application, driver service or Metro from -this qualification remains. Copied execution scripts were removed after use; -their JSON results and command logs remain as ignored evidence. - -#### Review follow-up: bounded fixture retirement — 2026-11-12 - -Source commit `0fb6cd47fccc4f6210797b25085e7b56aa8ca4fe` addresses the -reviewed indefinite crash-owner wait. The changed executable source is limited -to `fixtures/application-ownership/crash-owner.mjs`, its regression in the -existing `src/applicationOwnership.test.ts`, and four lines in -`packages/agentic/desktop-driver/native/macos/fixtures/ApplicationFixture.swift`. -The ready owner now waits -at most 120 seconds, emits a fixed diagnostic and exits 1. The AppKit fixture -requests termination on its main queue after 300 seconds; its existing -`/bin/sleep 300` descendant was already bounded. - -Production ownership, helper protocol, parent cleanup, identity verification -and claim settlement are unchanged. In particular, failure to obtain an owner -identity still does not authorize signalling a guessed PID. Owner retirement -allows helper EOF; neither fixture deadline settles an abandoned workload -claim or implements automatic recovery. - -**Observed portable regression:** with the shared wait still unbounded, the -new `--retirement-probe` test failed with `spawnSync ETIMEDOUT` after five -seconds. After adding the timeout, the same probe exited 1 with the expected -diagnostic. It uses a 25 ms deadline through the same wait function, before -acquiring any claim or starting a helper/application. This is portable -retirement evidence, not a native owner-death or five-minute AppKit runtime -pass. - -Exact commands and results at the source commit: - -- `node --check packages/agentic/desktop-test/fixtures/application-ownership/crash-owner.mjs`: - passed. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand applicationOwnership.test ownership.test`: - passed 30, failed 0, skipped 0. -- `yarn workspace @fluentui-react-native/desktop-driver format`, - `yarn workspace @fluentui-react-native/desktop-driver lint` and - `yarn workspace @fluentui-react-native/desktop-driver build`: each passed. -- `yarn workspace @fluentui-react-native/desktop-test format`, - `yarn workspace @fluentui-react-native/desktop-test lint` and - `yarn workspace @fluentui-react-native/desktop-test build`: each passed. -- `yarn workspace @fluentui-react-native/desktop-driver test --runInBand`: - passed 1,237, failed 0, skipped 66; native opt-in absent. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand`: - passed 410, failed 0, skipped 14; native opt-in absent. - -Under the shared native lock, using the existing stable signing identity and -helper cache, these commands also completed: - -```sh -yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver \ - --reporter json doctor --platform macos --permissions -node packages/agentic/desktop-driver/native/macos/fixtures/build.cjs \ - artifacts/desktop/f10m/retirement-followup-20261112/fixture -``` - -Doctor exited 0 with `promptRequested=false`, Accessibility/PostEvent/ -ScreenCapture preflight true, and **`screenLocked=true`**. The fixture -compiled and passed strict signature verification; it was not launched. -**Current native runtime qualification is blocked by the locked console.** -The prior 13 native cases and Fabric/Paper passes above remain evidence of -pre-fix content, present at `16558610ff2bd14775430f86c738c12947fe81ee`; -they are not reruns of this retirement change. No unlock, privacy alteration -or native GUI retry was attempted. - -The one required follow-up `yarn clean-lage-cache && yarn lage buildci` -cleared the cache, then **failed** only the unchanged -`boundedCommand.test.ts:31` SIGKILL-versus-SIGTERM assertion: driver tests -passed 1,236, failed 1, skipped 66. Private-storage tests passed in this run. -Both implicated test files and their implementations are byte-unchanged from -pinned main `041ac8b9894aeaa7e1aa273baa2ece4e46298b94`. -The aggregate used the native lock, inherited TMPDIR, no `TF_BUILD`, no native -opt-in and no concurrency override. It was not retried. Root instruction, -package-content and publishing checks passed within that failed aggregate; -these individual passes do not make the aggregate green. - -Separate `yarn lage test-links --verbose --grouped` passed 90 tasks. -`yarn markdown-link-check -c .github/markdown-link-check-config.json ` -passed for the five changed Markdown files listed above, with unchanged -12/6/19/2/0 link counts. - -Final locked command-line/working-directory audits at 23:23:50 and 23:23:53 -UTC found no worktree workloads or unreadable candidates and issued no -signals. The initial audit conservatively counted its own short-lived -`lsof` child; the corrected audit excludes only that exact known audit PID -and its own ancestor chain. No application, Metro or native ownership -fixture was launched in this follow-up. - -Evidence: `artifacts/desktop/f10m/retirement-followup-20261112/`, including -`retirement-unbounded-red.log`, `retirement-green.log`, -`driver-portable.log`, `harness-portable.log`, `doctor.json`, -`fixture-build.json`, `plain-aggregate.log`, `test-links.log`, -`changed-markdown-links.log`, `cleanup.json` and `cleanup-final.json`. - -## Windows Fabric and Win32 Paper — F10-W application ownership — 2026-11-12 UTC - -**Source and base decision:** implementation commit -`6e881e1a05fef735c1b9bacd7d123285f47269ff`, branch -`user/jasonvmo/desktop-windows-app-ownership`, **stacked on #4492** at -`6489386956c0485cb91b478a8f4adc4828f4178b`; draft PR base remains -`user/jasonvmo/desktop-macos-app-ownership`. The Windows side implements its -neutral application host/resources, claim schema 3, native OS owner identity, -exact process pins, inspection and launch settlement messages. A separate -claim schema or duplicate service factory was not introduced. This pins the -dependency; later #4492 movement is not authorization to rebase, merge or -push to its branch. - -Application resources can be named through the inherited neutral publication -seam without touching F11's CLI, service, agent connection, WDIO, session, -grant, client or MCP files. The base's existing `driverService.ts` changes -remain inherited only; W2 adds no edits to an F11-owned file. - -**Observed, test-first portable evidence:** new Windows identity/recovery -tests failed before implementation, then passed. Later red/green regressions -covered recovery barriers after claim removal, inconsistent containment, -unrecognized/invalid owner probes, preserving recovery plus barrier failures, -and portable service/reporting scopes without a prebuilt helper. The focused -Windows suite now has **31 passing tests** and supervisor suite **11**. -Existing neutral ownership tests also pass. Storybook identity tests pass -**7**, including Windows Fabric and Win32 Paper read-only resolution alongside -both macOS renderers; the epoch journal deliberately retains its existing -PID/start schema instead of taking another controlling attachment. - -Full portable results: - -| Package | Passed suites / tests | Skipped suites / tests | Interpretation | -| ----------------- | --------------------- | ---------------------- | -------------------------------------------------- | -| desktop-driver | 74 / 1,234 | 9 / 93 | Native opt-ins unset; not native qualification | -| desktop-test | 25 / 442 | 3 / 25 | Native/unsupported-platform opt-ins remain skipped | -| storybook-desktop | 23 / 347 | 1 / 3 | Portable integration; not renderer qualification | - -The first full harness attempt failed because the new eager Windows helper -resolution also affected portable scopes; focused failures reproduced it, -and the fix permits only `no-verified-prebuilt` to retain that existing -non-native behavior. Other verification failures still abort. A stale claim -cannot be recovered without a verified helper. The existing public boundary -fixture also needed the desktop-driver-mcp public entries built; no MCP source -was changed. Both failures were resolved before the full green harness rerun. - -Commands below ran from the worktree root; `$W2` denotes that root. -`TEMP`/`TMP` were scoped to its ignored `artifacts\desktop\f10w\scratch`; -`FURN_NATIVE_DRIVER_TEST` was unset. - -| Command | Observed result | -| --------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-test format` | Passed | -| `yarn workspace @fluentui-react-native/desktop-test lint` | Passed | -| `yarn workspace @fluentui-react-native/desktop-test build` | Passed | -| `yarn workspace @fluentui-react-native/desktop-test test --runInBand --cacheDirectory "$W2\artifacts\desktop\f10w\jest-harness"` | Passed 442; skipped 25 | -| `yarn workspace @fluentui-react-native/storybook-desktop format` | Passed | -| `yarn workspace @fluentui-react-native/storybook-desktop lint` | Passed | -| `yarn workspace @fluentui-react-native/storybook-desktop build` | Passed | -| `yarn workspace @fluentui-react-native/storybook-desktop test --runInBand --cacheDirectory "$W2\artifacts\desktop\f10w\jest-storybook"` | Passed 347; skipped 3 | -| `yarn workspace @fluentui-react-native/desktop-driver-mcp build` | Passed; existing boundary fixture | -| `yarn check-agent-instructions` | Passed | -| `yarn oxfmt --check` | Passed, 3,163 files | - -Ignored evidence run: `artifacts/desktop/f10w/portable-20261112/`. -Logs include `harness-test-initial.log`, final `harness-test.log`, -`driver-test.log`, `storybook-test.log`, each owning build/lint log, and -`changed-markdown-links.log`. Per-file link checks passed for driver README, -native protocol, Windows provider README, driver contributing/native-helpers, -harness README/configuration/lifecycle/native-qualification, Storybook -contributing, and the three-package patch changeset. The root formatter -intentionally excludes `.changeset/*.md`. - -**Observed, helper build only:** the locked Debug/x64 `buildWindowsDriver` -compile with build ID `f10w-compile-02` passed, 0 warnings/errors. Its seven -changed C++ files match the compiled snapshot. Native build evidence is -`artifacts/desktop/f10w/b2/build.log`. No application, fixture, native -self-test or physical-input case has run, and session kind is **not sampled**. -The pre-campaign audit reported -`FURN-AUDIT: no FURN desktop processes found`; no W2 campaign was started. - -**Implemented, still requiring native proof:** Windows/Win32 supervisor -ownership, exact external pins and acquired resource evidence now include -canonical executable, native seven-digit creation time, session, and packaged -AUMID. The native supervisor can reconcile an acquired exact app only after -the old owner is proven exited/reused, before its replacement claim/listener. -It serializes recovery, validates file identity/content and resource/effect -consistency, durably records partial progress, and preserves external apps. -Unknown services/resources, pending/unleased launches, mixed service graphs, -interrupted recovery, unverified owners, and unrelated effects remain fenced. -Custom hosts/delegated claims acquire no automatic native termination authority. -There is no general stale-lock remover, quarantine reset, or new protected -agent recovery route. Mixed Storybook graph and arbitrary uncontained-descendant -recovery are not qualified by this bounded application-only path. - -| Native gate | Windows Fabric packaged | Win32 Paper REX | Required evidence / limit | -| ------------------------------------------------------------ | ----------------------- | --------------- | --------------------------------------------------------------------------------- | -| Exact application/native owner identity | Not run | Not run | AUMID or canonical REX/executable, PID/start/session from real native evidence | -| PID-reuse/mismatch and ambiguity rejection | Not run | Not run | Real mismatch/fixture cases; actual OS PID reuse only if inducible | -| Containment or explicit packaged fencing | Not run | Not run | Packaged stays `uncontained`; REX must report Job Object and drain on helper exit | -| Cross-service application exclusion and physical input mutex | Not run | Not run | Independent helpers, physical opt-in and connected/unlocked session gate | -| Acquired survivor journal and next-run recovery | Not run | Not run | Literal-PID, identity-verified helper/owner fault injection | -| Repeated-run reliability | Not run | Not run | Five launches/teardowns and no old exact incarnation or claim leak | - -The opt-in consumer command is -`$env:FURN_NATIVE_DRIVER_TEST='1'; $env:FURN_NATIVE_APPLICATION_OWNERSHIP_TEST='1'; $env:FURN_NATIVE_PHYSICAL_INPUT_TEST='1'; yarn workspace @fluentui-react-native/desktop-test test --runInBand windowsApplicationOwnership.contract`. -It requires explicitly resolved Windows AUMID and REX JSON descriptors and a -verified prebuilt helper, under the granted exclusive physical desktop -campaign. Each endpoint gets five launch/teardown iterations and two borrowed -sessions per launch, stale identity checks, cross-service exclusion, and -journal recovery. Three paired packaged activations also check that the -losing launcher does not clean up the winning service's app. -See `references/native-qualification.md`, “Neutral Windows application -ownership”. No skips above establish these gates. The final clean-cache -`yarn lage buildci`, independent review, push, draft PR and hosted CI remain -not run at this preparation checkpoint. Part of #4447 and #4435. - -Preparation closeout pinned to the implementation SHA above: - -| Post-commit command | Result | -| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-test test --runInBand windowsOwnership supervisor applicationOwnership.test ownership.test --cacheDirectory "$W2\artifacts\desktop\f10w\jest-harness"` | Passed 71 tests | -| `yarn workspace @fluentui-react-native/storybook-desktop test --runInBand applicationIdentity --cacheDirectory "$W2\artifacts\desktop\f10w\jest-storybook"` | Passed 7 tests | -| `yarn change:check` | Passed | -| `yarn lage test-links --verbose --grouped` | Passed 90 tasks; 0 failed/skipped | -| `yarn oxfmt --check` | Passed repository-wide, 3,163 files | - -For each file below, the exact checker invocation was -`yarn markdown-link-check -c .github\markdown-link-check-config.json `; -all file arguments are worktree-relative: - -| File | Result | -| ------------------------------------------------------------------ | --------------- | -| `packages\agentic\desktop-driver\README.md` | Passed 16 links | -| `packages\agentic\desktop-driver\native\PROTOCOL.md` | Passed 5 links | -| `packages\agentic\desktop-driver\native\windows\README.md` | Passed 8 links | -| `packages\agentic\desktop-driver\references\contributing.md` | Passed 4 links | -| `packages\agentic\desktop-driver\references\native-helpers.md` | Passed 8 links | -| `packages\agentic\desktop-driver\references\current-status.md` | Passed 6 links | -| `packages\agentic\desktop-test\README.md` | Passed 7 links | -| `packages\agentic\desktop-test\references\configuration.md` | Passed 1 link | -| `packages\agentic\desktop-test\references\lifecycle.md` | Passed 1 link | -| `packages\agentic\desktop-test\references\native-qualification.md` | Passed 2 links | -| `packages\agentic\desktop-test\references\current-status.md` | Passed 15 links | -| `packages\agentic\storybook-desktop\references\contributing.md` | Passed; 0 links | -| `.changeset\windows-exact-application-ownership.md` | Passed; 0 links | - -The same ignored evidence root retains `tested-source-revision.txt`, -`compile-source-evidence.json`, each `*-focused-committed.log`, -`test-links.log`, `root-format.log`, and `pre-campaign-audit.log`. -Only appended Windows result capture followed the checks; no macOS ledger -section or native/runtime source was changed. - -## Windows Fabric and Win32 Paper — F10-W native qualification — 2026-11-12 UTC - -**Observed source and dependency:** real application qualification ran at -`8343dc0c48a2de01475454c892e7a295523caa75`, on -`user/jasonvmo/desktop-windows-app-ownership`. The coordinator handoff -`W2-base-decision.md` records **STACK** on #4492 at exactly -`6489386956c0485cb91b478a8f4adc4828f4178b`, implementing its neutral application -host/resources, schema-3 native owner claims, process pins, inspection and -settlement contracts. Application resource naming uses the inherited seam, -without new edits to F11's CLI, service, agent connection, WDIO, session, -grant, client or MCP files. The eventual draft PR base is -`user/jasonvmo/desktop-macos-app-ownership`. Its independently owned head may -advance or never merge; this pinned dependency is not permission to rebase, -merge, rewrite history or push to that branch. - -**Observed environment:** Windows build 26300 x64, Node 26.13.0, active/unlocked -RDP, `sessionKind: remote`, native `windows-session:1`. Every native command ran -under the exclusive W2 lock with `-RequirePhysical`; the session remained -eligible before and after each command, with no exit-75 or session-change -result. The helper used Debug/x64, prebuilt-only resolution, artifact/SHA256 -`863cc2dbfc8f5792d22c8606376d226affbf8655fe4ded62ba78c23a29c516c7`, -source digest -`ea76e931dc4cdebc87e0621a7d730f81faabce5ef902063a0512cd7c2b100f2d`. -Private machine paths are retained only in ignored artifacts. - -The registered Windows Fabric product was reused from the -coordinator-specified `furn-81` build at `bdeb0b291`: there are no Windows -app/native-core Windows source changes relative to that build, and the -generated `RnwNewArch`/`UseFabric` values are true. All five native ordinary-run -identities matched its registered manifest image and AUMID -`40411fc5-8e92-4d46-b68d-b62df44b1366_syyqehzwccsyy!App`. -Windows Metro was owned by W2 on 8081 and passed `/status`; both bundles were -produced from W2. Win32 used its canonical prebuilt REX `ReactTest.exe`, Paper, -and the W2 bundle/component/working-directory descriptor. Native Win32 app -build is **Unsupported**. - -**Observed corrections before the green run:** - -- The selected native helper contract initially failed an existing C++ assertion - that dereferenced a destroyed temporary JSON release result. Commit - `4a4b87f37b574820b93e342b68721d931be07c55` retains the result, and the complete - selected native contract/self-test rerun passed. Release behavior is unchanged. -- The first real consumer run passed 3 and failed 5. The fixture's arbitrary - project settlement root was correctly refused with native reason `ancestor`; - this worktree's ancestors allow broader modification. The final fixture uses - the normal native per-user private-store default, without changing ACLs, - permissions or the provider's fail-closed checks. -- Paired packaged activation produced two different PID/start incarnations, - not two owners of one incarnation. Commit `8343dc0c4` corrects that fixture - assumption: three pairs on **each** endpoint now prove distinct controlling - incarnations, unpinned ambiguity refusal, pinned cross-service exclusion, and - that closing one winner preserves the other. Refused contenders, if present, - are disposed before checking the winners. All observed pairs had two distinct - winners; the shared-PID activation race was not induced. - -The corrected fixture passed format, lint and TypeScript build. Its focused -portable selection passed **31 tests**, with **9 native cases skipped** because -native opt-in was unset; the later real native run, not those skips, qualifies -the cases below. Earlier full portable results remain the preparation results -above. The corrected provider selection passed **6 tests**, including all five -new real-process fixture cases, with 10 unrelated filtered cases skipped. - -**Observed final application result: 9 passed, 0 failed, 0 skipped**, one suite, -at `2026-11-12T16:04:04.248Z`. The 27 evidence rows include iteration receipts -and case verdicts; they are not 27 tests. - -| Case | Windows Fabric packaged | Win32 Paper REX | Native evidence | -| ----------------------------------------------- | ----------------------- | ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| Exact identity | Passed | Passed | Canonical image, PID, seven-digit creation time, session; packaged AUMID or exact REX incarnation | -| Read-only inspection and borrowed sessions | Passed | Passed | Unchanged window ids; two sessions per supervisor launch; first deletion preserves the app | -| Start/session/executable mismatch | Passed | Passed | Both inspect and trusted recovery refuse the wrong identity without ending the real process | -| Ambiguous matching processes/windows | Passed | Passed | Three pairs of two actual incarnations; unpinned inspection refuses, exact pins remain readable | -| Independent-service exact application exclusion | Passed | Passed | Independent helper's pinned attach fails `application-lock-busy`; observations do not take control | -| Containment or explicit fencing | Passed, `uncontained` | Passed, `job-object` | Packaged survivor stays journaled and fenced; REX exits with the crashed helper's job | -| Survivor journal and next-run recovery | Passed | Passed | Acquired application resource; live owner refused; literal-PID helper/owner stops after full native identity verification; next supervisor reconciles before replacement | -| Physical-input mutex | Passed | Passed | Native Shift down, second helper cancelled with `AbortError`/effect `none`, release then transfer | -| Repeated-run reliability | Passed 5/5 | Passed 5/5 | Five distinct PID/start tuples per endpoint, two borrowed sessions each; no exact old process or claim leak | -| Concurrent cleanup isolation | Passed 3/3 pairs | Passed 3/3 pairs | Closing one exact winner leaves the other alive; final cleanup proves every incarnation exited | -| Actual kernel PID recycling | Not run | Not run | Not induced; provider's one-100-ns-tick stale pin and real endpoint start mismatch both passed | -| Shared-PID packaged activation race | Not run | Unsupported, not packaged | Every actual Windows pair created two distinct processes; no claim that this exercised a losing shared-PID launch | - -The crash receipts include packaged PID **10984**, creation time -`2026-11-12T16:02:38.6619313Z`, helper **37916**, owner **35708**; and REX -PID **35844**, creation time `2026-11-12T16:03:18.2246578Z`, helper **34864**, -owner **37256**. Native evidence verified each fault-injection identity again -immediately before a single literal-PID stop. The packaged process survived -helper death, the live owner fenced replacement, and after owner exit the next -supervisor ended that exact app and removed the old claim. REX's Job Object -drained on helper death; its next supervisor reconciled the acquired dead app -and claim. No broad name-based termination was used. - -Commands below are the inner workspace commands issued through the required W2 -lock with `-RequirePhysical`. `$W2` denotes the assigned worktree, `$E` its -`artifacts\desktop\f10w\native-20261112` evidence directory. `TEMP`/`TMP` point -to `$W2\artifacts\desktop\f10w\scratch`; the native cache is -`$W2\artifacts\desktop\f10w\n`, configuration `debug`, build policy `never`. - -| Workspace / command | Observed result | -| ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- | -| Storybook: `yarn desktop-driver build-driver --platform windows --configuration debug --cache-root "$W2\artifacts\desktop\f10w\n"` | Passed initial explicit build; provider contract explicitly rebuilt the corrected source | -| Storybook: `yarn desktop-driver doctor --platform windows --configuration debug --cache-root "$W2\artifacts\desktop\f10w\n"` | Passed selected final artifact | -| Storybook: `yarn storybook bundle --win32` | Passed; run `14c14c80-0da7-48ad-ac14-47e0c257fbe1` | -| Storybook: `yarn storybook bundle --windows` | Passed; run `f6ba6fb4-25c9-49a2-9928-8b2430abc83b` | -| Repository root: `node .yarn\releases\yarn-4.18.0.cjs workspace @fluentui-react-native/agentic-components-storybook start --port 8081` | Passed owned Metro `/status` and final shutdown | -| Repository root: `node .yarn\releases\yarn-4.18.0.cjs workspace @fluentui-react-native/desktop-test test --runInBand --cacheDirectory "$W2\artifacts\desktop\f10w\jest-harness" --runTestsByPath src\windowsApplicationOwnership.contract.test.ts --json --outputFile "$E\consumer-results.json"` | Passed 9, failed 0, skipped 0 | - -The consumer inherited `FURN_NATIVE_DRIVER_TEST=1`, -`FURN_NATIVE_APPLICATION_OWNERSHIP_TEST=1`, -`FURN_NATIVE_PHYSICAL_INPUT_TEST=1`, and the two explicit JSON descriptors in -`windows-application.json` / `win32-application.json`. REX's descriptor came -from the owning adapter, not guessed host arguments. No input features were -masked. The command runner owned Metro and awaited its normal resource cleanup; -the outer report's “lifecycle (no tests)” is not the native test verdict. - -**Ignored evidence:** `artifacts/desktop/f10w/native-20261112/` contains -`consumer-results.json`, `consumer-evidence-final.json`, -`qualification-summary.json`, `selected-helper.json`, -`application-build-provenance.json`, `metro-identity.json`, `doctor-debug.log`, -both bundle logs, corrected provider logs, and preserved initial failed -attempts. Canonical native evidence is -`packages/agentic/desktop-test/artifacts/desktop/windows-application-ownership-GiW4x1/evidence.json`; -the outer command report is -`artifacts/desktop/f10w/native-20261112/commands/4ed98f8f-8ebb-49ca-ac90-2ee6caece7dc/run.json`. -Registration, executable/bundle hashes, source revision, real PID/start/session -receipts, and every case outcome are retained there. Machine-specific paths -and crash-owner request/ready records remain ignored, not public. - -**Observed closeout:** all W2 campaigns ended. Final audit: -`FURN-AUDIT: no FURN desktop processes found`; no owned helper, fixture, Metro, -REX or packaged app survived. No push or PR was made. Independent review, the -final one-time locked clean-cache `buildci`, and hosted CI remain **Not run**. -The existing three-package patch changeset is -`windows-exact-application-ownership.md`. - -**Limits:** the tests do not establish arbitrary uncontained-descendant cleanup, -mixed Storybook-graph recovery, forced kernel PID recycling, the unobserved -shared-PID packaged activation race, or a new protected-agent recovery API. -No macOS renderer claim is made by this Windows campaign. Part of #4447 and -#4435. - -## Windows Fabric and Win32 Paper — recovery review follow-up — 2026-11-12 UTC - -**Source:** `8b2fb7e67dc5864180c7863ac94fc3c0c3c9d6e3`, following the -review of `6489386956c0..6bdcde61f`. The branch remains stacked on #4492 at -`6489386956c0485cb91b478a8f4adc4828f4178b`; the base decision recorded in -the preceding Windows sections is unchanged. No F11-owned file was changed -for these fixes; no rebase, dependency-branch write, push, or PR occurred. - -**Implemented:** parsing, target validation and native old-owner probing are -read-only and precede recovery-barrier creation. The original claim's file -identity and exact contents are rechecked under the barrier before any -application cleanup. A versioned `.recovery-v2` barrier atomically publishes -the recoverer's verified native PID/start in a unique empty owner directory. -An empty versioned root authorizes no cleanup. A later native recovery may -retire unchanged pins only after proving their recoverers exited or their -PIDs were reused, even if the old claim was already removed. Live, unknown, -malformed and legacy unowned barriers remain fenced. Replacement directories -detected during probing/cleanup are never adopted or removed. - -The paired native change settles an application's reused recorded PID as an -exited incarnation without signalling its replacement. A matching application -start still requires full identity and application-control exclusion; an -unreadable start remains fenced. Prior native results that expected stale-start -recovery rejection are historical, not qualification of this corrected behavior. - -**Observed test-first evidence:** focused tests on pre-fix `6bdcde61f` produced -**9 failed / 31 passed**: a barrier appeared during the live-owner probe, a -foreign target caused a temporary barrier, a changed claim inherited recovery -authority, and interrupted barriers could not be reconciled. Ignored -`artifacts/desktop/f10w/review-20261112/barrier-red.log` retains those failures. -After the fixes and additional publication/replacement-race coverage: - -| Case | Outcome | -| ------------------------------------------------------------------------------ | ----------------------------------------- | -| No barrier for a live owner or foreign target | **Passed**, portable | -| Claim replaced during barrier acquisition cannot inherit cleanup authority | **Passed**, portable | -| Verified dead/reused recovery owner can resume partial recovery | **Passed**, portable | -| Live/unknown owner, malformed pins and legacy barriers stay fenced | **Passed**, portable | -| Interrupted publication and dead barrier after claim deletion can be retired | **Passed**, portable | -| Replacement-barrier preservation and original-plus-cleanup error retention | **Passed**, portable | -| Corrected native application-start semantics and actual killed recoverer | **Not run** on this source; grant pending | -| Fabric/REX repeated runs, crash recovery, concurrency and physical mutex rerun | **Not run** on this source; grant pending | - -**Observed commands/results:** - -- `yarn format`, `yarn lint`, `yarn build` in the owning harness: **Passed**. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand windowsOwnership supervisor applicationOwnership`: - **Passed**, **77 passed / 21 native skipped**. Windows ownership alone has - **42 passing** portable cases. `focused-results.json` and - `focused-tests.log` retain the expanded invocation and results. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand`: - **Passed**, **453 passed / 28 native skipped**, in - `harness-results-final.json` and `harness-tests-final.log`. An earlier full - attempt caught one replacement-barrier diagnostic regression; it was fixed - without weakening foreign-barrier preservation and the full suite rerun. -- Driver package format/lint/build and **1,234** portable tests: **Passed**. - Compile-only Debug/x64 helper rebuild under the machine lock: **Passed**; - source digest `3558e8597f6cfe1ee3f4e38dca2137710b80d193a5c30bf946d54065bf8bc153`. - See the driver's matching Windows section for its binary receipt. - -**Native rerun prepared, not executed:** the actual-endpoint suite now has -**11** cases. On each endpoint, an injected stale owner/application journal -keeps the real replacements live and owned. A real child recoverer publishes -its barrier, fences another startup, and is stopped by an identity-verified -literal PID. The next supervisor must reconcile that dead barrier and retire -the absent recorded application without signalling its live replacement. -Injected start mismatch is not a claim of actual kernel PID recycling. -The existing five-run, crash-survivor, concurrent-incarnation, and gated -physical-mutex cases remain. Run under the granted W2 campaign with both -explicit descriptors and native/physical opt-ins; no native skip is a pass. -No W2 campaign was started in this phase; no renderer/session qualification is -claimed. The existing patch changeset remains -`windows-exact-application-ownership.md`. Final clean-cache `buildci`, -publication and hosted CI remain **Not run**. Part of #4447 and #4435. - -**Observed closeout:** edited-reference/ledger Markdown links, root -`yarn oxfmt --check` (**3,163 files**), `yarn check-agent-instructions`, -`yarn change:check`, and `git diff --check` **Passed**. Both retained/rebuilt -helper hashes were verified. The concurrent audit initially observed only -the portable formatter's two Node processes; after they exited, the audit -reported `FURN-AUDIT: no FURN desktop processes found`. No W2 campaign was -started and no owned desktop process survives. `validation-summary.json` -and `closeout-audit.log` retain the ignored phase receipts. These portable, -compile and hygiene results do not replace the pending native rerun. - -## Windows Fabric and Win32 Paper — default-environment buildci — 2026-11-12 UTC - -**Observed:** root `yarn clean-lage-cache` followed by `yarn lage buildci` -passed at `68d63ccfbf008b963e59772f57aca8ac8f9ff05f`, under the W2 machine -lock with no physical-session requirement. Default TEMP, default concurrency -and repository guidance were unchanged. - -| Attempt | Outcome | -| ------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------- | -| Initial W2 gate, same revision | **Failed**: 173 successful tasks, one failed Storybook task; server-integration child exit 1 and inline-worker aggregate timeout | -| Full pinned-base comparison at `6489386956c0485cb91b478a8f4adc4828f4178b` | **Passed**: 178 successful tasks, 0 failed/skipped | -| Unchanged W2 full gate rerun | **Passed**: 178 successful tasks, 0 failed/skipped, 2m 32.26s | - -No source fix, dependency installation, timeout change, concurrency override, -or guidance workaround was made. The initial failure did not reproduce in the -base comparison or the unchanged W2 rerun. Native opt-ins remained unset: -this aggregate is not Windows Fabric or Win32 Paper runtime qualification. - -All three logs and `buildci-default-summary.json` are retained in ignored -`artifacts/desktop/f10w/review-20261112/`; the final process audit is -`buildci-default-final-audit.log` and **Passed**: -`FURN-AUDIT: no FURN desktop processes found`. -No application/native contract was launched, no desktop campaign started, -and no push occurred. Corrected native recovery re-verification remains -**Not run**, awaiting an active/unlocked session and its separate grant. -Part of #4447 and #4435. - -## Windows Fabric and Win32 Paper — recovery native re-verification — 2026-11-12 UTC - -**Observed source:** native helper and harness implementation -`dbaf558939db31b4a3935ad68b1eb20d25d90d3e`, still stacked on #4492 at -`6489386956c0485cb91b478a8f4adc4828f4178b`. The consumer test alone had -the one-line assertion correction committed with this section: the public -`NativeDesktopHost.recoverApplication` returns `Promise`, after checking -the native `{ state: 'exited' }` verdict. Its consumer must expect an undefined -resolution, not the wire-level result. No production, protocol or F11 -file changed in this phase. The exact test diff and test-file digest are retained -in ignored `artifacts\desktop\f10w\review-native-20261112`. - -**Observed session and application provenance:** both granted W2 campaigns ran -in active/unlocked **remote** session 1, input desktop `Default`. Every native -wrapper used the machine lock and `-RequirePhysical`; before/after gates stayed -eligible, with no exit 75 or session change. The final consumer wrapper ran -`21:54:54Z–21:57:02Z`. Registration and image hash were checked before each -phase: packaged Fabric remained the `furn-81` Debug/x64 build from `bdeb0b291`, -with no relevant native-app/source difference. Its executable SHA-256 remained -`df9cb3559843fcfb0158ee4ab5edbae78890c4e231ebd895cbb636c24b18cd45`. -The AUMID was -`40411fc5-8e92-4d46-b68d-b62df44b1366_syyqehzwccsyy!App`. -Win32 Paper used the worktree's canonical prebuilt REX -`0.81.1-6828073195` image, its exact arguments and a fresh Win32 bundle; -a Win32 native application build remains **Unsupported**. - -Fresh Win32 and Windows bundles **Passed**, with run IDs -`3830f2d5-30b1-405b-a205-71f45dcc9490` and -`e2fb42e1-c5ec-41b3-bae3-64cedad0aeb6`. Each consumer invocation owned a -fresh Metro, verified its listener's native PID/start/session and worktree -command line, and stopped it in `finally`. No preexisting Metro was adopted. -Native scratch/cache paths stayed inside the worktree. - -**Observed test-first sequence:** - -1. Retained old helper `863cc2dbfc8f5792d22c8606376d226affbf8655fe4ded62ba78c23a29c516c7`, - source `ea76e931dc4cdebc87e0621a7d730f81faabce5ef902063a0512cd7c2b100f2d`: - **4 expected failures / 0 passed / 7 filtered skips**. Both endpoints rejected - stale-start recovery, including interrupted-recoverer journal settlement, - with `identity-start-mismatch`. This was the required native RED, not an - environment failure. -2. Review-fixed helper - `97f0072ce22449bc5c87c05de259d3e4c46eca666e6135977432b30fa16053b2`, - source `3558e8597f6cfe1ee3f4e38dca2137710b80d193a5c30bf946d54065bf8bc153`: - the first complete run had **9 passed / 2 failed / 0 skipped**. Both failures - were the consumer's wrong return-value assertion; recovery resolved - successfully, but the test expected the native wire object. The existing - host identity unit contract already expects `undefined`. -3. Correct only that assertion to `.resolves.toBeUndefined()`, retaining the - independent assertion that the real replacement is alive and all - matching-start mismatch refusals. Rerun the whole native consumer suite: - **11 passed / 0 failed / 0 skipped**. - -| Consumer case | Windows Fabric packaged | Win32 Paper REX | -| -------------------------------------------------------------------------------------- | ---------------------------------------- | ---------------------------------------- | -| Exact identity, independent-service exclusion, five launches/teardowns without leaks | **Passed**, 5/5 distinct PID/start pairs | **Passed**, 5/5 distinct PID/start pairs | -| Stale-start settlement; matching-start session/executable/AUMID mismatches fail closed | **Passed** | **Passed** | -| Crashed recoverer, stale journal and dead recovery barrier; live replacement preserved | **Passed** | **Passed** | -| Crashed owner/helper and exact application journal recovery | **Passed**, packaged survivor recovered | **Passed**, job-object drain reconciled | -| Concurrent pairs, ambiguity refusal, pinned exclusion and independent teardown | **Passed**, 3/3 pairs | **Passed**, 3/3 pairs | -| Physical input mutex, cancellation without effect, release and transfer | **Passed**, one cross-endpoint case | Same cross-endpoint case | - -Every repeated identity reported `windows-session:1`. Fabric consistently -reported `uncontained`; REX consistently reported `job-object`. All three -concurrent pairs on each endpoint acquired two distinct incarnations: the -shared-PID packaged activation race was **Not induced**. Interrupted recoverers -were identity-verified Node processes stopped by literal PIDs **54440** and -**55808**, with recorded native starts -`2026-11-12T21:55:28.3844388Z` and -`2026-11-12T21:56:07.0500082Z`. The subsequent supervisor retired the dead -recovery barrier and injected stale application record without signalling the -real replacement. These injected recorded-start mismatches are **not** actual -kernel PID recycling. - -**Exact consumer commands:** `$W2` denotes the assigned worktree and `$E` -its `artifacts\desktop\f10w\review-native-20261112` directory. From `$W2`, -inside the granted campaign and physical-gated machine lock: - -```powershell -node -- artifacts\desktop\f10w\review-native-20261112\run-consumer.mjs red artifacts\desktop\f10w\native-20261112\selected-helper.json -node -- artifacts\desktop\f10w\review-native-20261112\run-consumer.mjs green artifacts\desktop\f10w\review-20261112\selected-helper.json -``` - -The retained bounded orchestration uses `NodeDesktopCommandRunner`, the two -explicit application descriptors, all three native/application/physical -opt-ins, the selected immutable helper and `buildPolicy=never`. Its GREEN -Jest child command is: - -```powershell -yarn workspace @fluentui-react-native/desktop-test test --runInBand ` - --cacheDirectory "$W2\artifacts\desktop\f10w\jest-harness" ` - --runTestsByPath src\windowsApplicationOwnership.contract.test.ts ` - --json --outputFile "$E\green-consumer-results.json" -``` - -RED adds `--testNamePattern 'native start|crashed recoverer'` and uses -`red-consumer-results.json`. The final real consumer evidence directory is -`windows-application-ownership-3BLzLV`; outer command report -`32f3d6be-c4b0-4792-aeb6-a5d7df165fc0` is lifecycle-only, not the test-count -source. `green-consumer-results.json`, `green-consumer-evidence.json`, -`green-invocation.json`, registration receipts and `consumer-assertion-fix.diff` -retain the actual counts, identities and dirty-test provenance. Expected RED -and first-GREEN failures are preserved separately, not overwritten. - -**Observed additional validation:** - -- `yarn workspace @fluentui-react-native/desktop-test format`, `lint`, and - `build`: **Passed**. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand ---cacheDirectory "$W2\artifacts\desktop\f10w\jest-harness" ---json --outputFile "$E\harness-portable-results.json"`: - **Passed**, **453 passed / 28 native skipped**. -- `yarn workspace @fluentui-react-native/desktop-driver test --runInBand -NativeDesktopHost.applicationIdentity ---cacheDirectory "$W2\artifacts\desktop\f10w\jest-driver" ---json --outputFile "$E\host-identity-portable-results.json"`: - **Passed**, **31 passed**, including the existing void-return and unsettled - native-verdict contracts. -- The driver's matching native selection: **Passed**, **6 passed / 10 - unrelated filtered skips**, including native build/cache/handshake/self-test - and the physical mutex. Its ledger records the exact invocation. -- Locked default-environment `yarn clean-lage-cache` followed by - `yarn lage buildci`: initial attempt **Failed**, 176 successful tasks and one - Storybook task failure, `A stopping command cannot publish readiness` in - `serverIntegration`. The reported intermittent readiness failure was observed - again; its cause remains unproven. After remaining fail-fast test children - settled and the audit was clean, the **one** unchanged full rerun **Passed**: - **178 successful / 0 failed / 0 skipped**, 2m 38.65s. Neither attempt changed - default TEMP, concurrency, timeouts, dependencies or repository guidance. - Logs are `buildci-default.log` and `buildci-default-rerun.log`. - -**Closeout and limits:** both W2 campaigns ended. Native and final aggregate -audits **Passed**: `FURN-AUDIT: no FURN desktop processes found`. No W2 app, -helper, fixture or Metro survived. The existing three-package patch changeset -`windows-exact-application-ownership.md` remains applicable. The test-only -correction requires coordinator review before push; this phase has not pushed, -edited the PR body, or started new hosted CI. Initial PR #4494 CI at `dbaf558939` -was observed green; it does not qualify this new local test correction. -F11 claim naming remains deferred and untouched. Real kernel PID recycling, -shared-PID activation, arbitrary uncontained descendants and mixed-graph -recovery remain unqualified. Part of #4447 and #4435. - -**Observed reference checks:** Markdown link checks for both edited ledgers -**Passed** (15 harness links, 6 driver links). Root `yarn oxfmt --check` -(3,163 files), `yarn check-agent-instructions`, `yarn change:check`, and -`git diff --check` **Passed**. Receipts are retained beside the native results. - -## F15 macOS feasibility and resolved permission block — 2026-11-12 - -The table below preserves the initial, non-AX pass. The permission prerequisite -was resolved later the same day; see the permission-enabled follow-up below. - -Source `505a2187b0a2451aed7081c9260b343d55742f93`; macOS 27.2.1 -(26B131), arm64 (`sw_vers`). The app exercised was -`FurnDesktopDriverHost.app`, not a React Native consumer; renderer N/A. -The pre-run feasibility artifact is -`artifacts/desktop/f15-20261112/feasibility.md`. - -| Item | Needs Accessibility/PostEvent? | Plan / actual result | -| -------------------------------------------------------------------------------------- | ------------------------------------------------ | -------------------------------------------------------------------------------------------------- | -| Stable helper build, signature verification, handshake and synthetic self-test | No | Passed under the machine-wide native lock; signed self-test 448 passed, 0 failed, 0 skipped checks | -| Strict-clear refusal and negotiation assertions | No | Passed 8 supporting assertions within the self-test, not successful real-element clears | -| Native service/application identity contracts | No | Targeted native driver selection passed 27 tests, failed 0, skipped 15; default TMPDIR | -| Noninteractive `doctor --permissions` | No | Completed without prompting; Accessibility=false, PostEvent=false, ScreenCapture=true | -| Real-element clears, Fabric/Paper Storybook descriptor lifecycle/smoke, physical input | Yes; keyboard/pointer input also needs PostEvent | Initial pass: not run, 0 selected/executed; permission prerequisite subsequently resolved below | -| Paper MCP interaction measurements | Yes | Initial pass: not run; permission prerequisite subsequently resolved below | -| Windows / Win32 matching-OS contracts | Windows host required | Not run on this macOS host | - -The driver evidence ledger at -`packages/agentic/desktop-driver/references/current-status.md`, section -"F15 non-AX macOS supporting evidence — 2026-11-12", -records the exact commands, signing requirement, source/artifact digests, -counts, and cleanup. Machine-local artifacts are under -`artifacts/desktop/f15-20261112/`, including `helper-build.json`, -`doctor.json`, `signed-self-test.log`, and `native-contracts.log`. -All commands exited and subsequent process inspection found no surviving -test/helper children. No Storybook, Metro, driver service, or MCP session was -started for these mainline checks, and no privacy settings changed. -These are supporting helper contracts, **not Fabric/Paper requalification**; -the remaining acceptance matrix and historical evidence are unchanged. - -### F15 permission-enabled consumer follow-up — 2026-11-12 - -**Observed, partial; not macOS requalification.** Source: -`141b88b4b81a1bd9a0472de533681f9274d8cb00`, macOS 27.2.1 (26B131), -arm64. Consumer: Agentic Components Storybook / ReactTestApp, instance -`d915e78595`, bundle `com.microsoft.fluentui.agenticstorybook.id915e78595`; -targets `macos-a6dfeaccb13e643a` (Fabric) and `macos-de43c71d45415315` -(Paper). The stable-signed nonprompting doctor -now reports Accessibility/PostEvent/ScreenCapture true, an unlocked console -and `promptRequested=false`. This is evidence from this PR's tree only. - -Native/GUI commands were serialized by -`$HOME/Library/Caches/furn-overnight/with-native-lock.sh`, with -`FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development"` -and `FURN_DESKTOP_DRIVER_CACHE_ROOT="$PWD/artifacts/desktop/f15-20261112/native-cache"` -set from the repository root. The feasibility note, qualification harness, -logs and redacted JSON are retained in the ignored machine-local directory -`artifacts/desktop/f15-20261112T115411Z/`. No permission prompts or privacy -changes were made. - -| Renderer / command from `apps/storybook` | Result / passed, failed, skipped | Report under `apps/storybook/artifacts/desktop//run.json` | -| -------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------- | -| Fabric: `yarn storybook bundle --macos --reporter json` | Passed stage; 0 tests | `22ed99ad-99d3-464b-91e0-b0e41f426674` | -| Fabric: `yarn storybook prep --macos --reporter json` | Initial stage failed before selection: installed Ruby JSON 3.1 removed `fast_generate`; 0 tests | `ebf07ee0-38df-4c79-85cb-3d3b3d4dd42c` | -| Fabric: same `prep`, with `RUBYOPT="-r$PWD/artifacts/desktop/f15-20261112T115411Z/pin-json.rb"` set at the repository root | Passed using already installed JSON 2.18.0; no package/product changes; 0 tests | `e5e4aeea-dd87-4088-ab19-de09a740e338` | -| Fabric: `yarn storybook smoke --macos --mode stories-and-tests --reporter json` | **Failed**: 225 selected/executed, 223 passed, 2 failed, 0 skipped; 0 retries/flaky; later callbacks not run | `8e1bfc4a-3037-42af-8d52-47739db9065d` | -| Paper: `yarn storybook bundle --macos --paper --reporter json` | Passed stage; 0 tests | `f0d64be2-6b22-4692-9ddb-97dbe827a2a5` | -| Paper: `yarn storybook prep --macos --paper --reporter json` with the same Ruby selection | Passed stage; 0 tests | `2bdd101c-6269-4c3d-8e1c-edb2bd3ba34a` | -| Paper: `yarn storybook smoke --macos --paper --mode stories-and-tests --reporter json` | **Failed**: 225 selected/executed, 223 passed, 2 failed, 0 skipped; 0 retries/flaky; later callbacks not run | `9f8bb135-907f-419c-8c23-d543fc201230` | - -Each renderer's 223 passing entries were story traversal, not authored input -passes. Each report recorded zero unsupported, declared-unsupported, -quarantined and incomplete work; those zero counters do not mean the later -authored callbacks ran. Selection excluded ten platform entries for each -renderer and one additional scope entry for Paper. Both failed callbacks belong to -`components-accordion--keyboard-disclosure`: `focusByTab`'s initial -physical click was rejected as obscured by the app's own native window; -Fabric's second callback instead reported foreground activation failure. -Later authored work was not -silently counted as skipped or passed. **Open finding:** see the driver's F15 -follow-up for reproduction and suggested ownership. - -The separate real-descriptor command -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs fabric lifecycle` -(repository root) recorded **10 passed, 1 failed, 0 skipped manual checks**. -It used `yarn storybook driver --macos --agent --reporter json` and -`yarn storybook run --macos --reporter json`, never `build` as a substitute -for the instance-aware `run`. Passing cases included grant revocation, -failed DELETE before/after dispatch, ambiguous CREATE, disposal during open, -read cancellation, target capture/resource expiry, attached-app preservation -and same-port service replacement. The failure was a refused MCP checkbox -physical click, not a successful interaction. Fault-injection probes used the MCP SDK -in-memory transport with the real verified descriptor/native service; normal -operations used the shipped stdio adapter. - -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs fabric preservation` -added **1 passed, 0 failed, 0 skipped manual check**: select/reset and stale -preview rejection through the protected descriptor, then helper-described -PID/start-time/executable equality after session and service close. -Artifacts: `fabric-lifecycle-qualification.json` and -`fabric-preservation-qualification.json` in the follow-up directory. - -Paper's separate command, -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper`, used -`yarn storybook driver --macos --paper --agent --reporter json` and -`yarn storybook run --macos --paper --reporter json`. It recorded -**14 passed, 4 failed, 0 skipped manual assertions**: -revocation, delayed-open disposal, ambiguous CREATE, both DELETE-loss -boundaries, owned select/reset, stale-preview rejection, read cancellation, -capture/resource expiry, attached-app preservation and same-port replacement -passed. The real channel survived 22 seconds idle. This is a bounded lifetime -check, not a reliability window. The physical Checkbox click was refused -before dispatch (`obscured`); semantic activation was **Unsupported** by the -provider. The keyboard-clear assertion failed with a nonempty value, and the -next semantic-clear assertion failed its dependent precondition without -attempting that mutation. Do not count these as four independent defects. - -The independent follow-up, -`node artifacts/desktop/f15-20261112T115411Z/qualify.mjs paper input`, -recorded **4 passed, 1 failed, 0 skipped manual checks**. Value polling -confirmed the keyboard-clear failure was not merely the immediate read: -100 reads with 50-ms waits still found two characters. Semantic clear, -independent physical pointer/keyboard/text/wheel delivery, protected-args -refusal and a successful MCP type/value-check/observe interaction passed. -No mutation was retried. The driver ledger records receipts, reproduction -and suggested ownership; the MCP ledger records the Paper measurements. -Original results remain in `paper-qualification.json`, with the exact full -harness archived as `paper-harness.mjs` (SHA-256 -`d43c92e080896d4c29ea64b1c7cf5226009e807fe131ab68d6c2059e7fbe77e5`); -the follow-up is `paper-input-qualification.json` / `paper-input-harness.mjs`. -Paper driver runs were `53a421b0-d2f2-4037-97f0-d32b4c557dbd` -and `600456f7-575b-4f2b-877e-07387a0004c4`; their companion -`run.json` files report zero selected suite tests, not these manual counts. - -Routine shutdown signalled only each exact owning CLI, which exited 0; -driver/channel/Metro exited and the descriptor was withdrawn. The app's -exact incarnation was preserved until this operator terminated its own app. -Apps were quit before another `run` and before renderer changes. -An earlier harness attempt broadcast SIGTERM to its whole process group, -leaving a stale descriptor and failing restart. It is retained as **9 passed, -4 failed, 0 skipped manual checks** in `fabric-qualification.json`; initial -context-readiness and dependent clear checks also failed there. -After proving service PID 51504 exited and its three ports had no listeners, -the exact stale record was removed with validated storage's compare-and-remove -API (`descriptor-reconciliation.json`). This is operator reconciliation, not -automatic recovery or a clean shutdown pass for that attempt. - -**Observed final cleanup, passed:** after all app/service runs, -`node artifacts/desktop/f15-20261112T115411Z/audit-cleanup.mjs` ran -under the native lock at 2026-11-12T12:20:26Z. It found no processes from -this worktree, no listeners on 39860/23935/28029, no default descriptor -and no per-user workload claim (`cleanup-audit.json`). The first audit -mistook a concurrently running evidence-summary shell for a survivor; -that diagnostic is retained in `cleanup-audit-initial.json`. The isolated -audit passed without killing or ignoring any process. The wrapper exited 0, -and subsequent `lsof` found the machine-wide `native.lock` unheld. - -**Remaining qualification:** the failed authored Accordion callbacks and -unexecuted later callbacks, refused physical Checkbox activation and Paper -keyboard clear must be resolved. Native cleanup-failure retention, injected -startup/shutdown survivors, story quarantine/recovery, mutation cancellation, -correlated readiness/render faults, lost-context fencing/recovery, registered -secondary windows, protected HTTP/WS ingress, launch-mode discovery and -explicit legacy separation were **Not run** on this source. Fabric idle -lifetime, the reliability window and durable TCC continuity were **Not run**. -Windows/Win32 were **Not run** on this macOS host. Typed-client response-loss -tests do not establish native failure containment. #4441's macOS acceptance -remains open; no sibling-branch evidence or successful smoke is inferred. - -## F11 generic attach-only handoff — 2026-11-12 - -Reviewed implementation source: `5c9d0191e` (including `4aafeeabb` and `87bc8b710`). -Native selection ran at `87bc8b710`. This is the generic attached-application -phase of [#4448](https://github.com/microsoft/fluentui-react-native/issues/4448), -not the blocked Storybook integration or retained-launch phases. - -**Confirmed, portable:** driver, harness, and MCP package format, lint, and -build checks passed. Driver tests passed 1,233 with 66 skipped; MCP tests -passed 32; the locked complete harness suite passed 382 with two skipped. -The real WDIO deterministic normal-handoff and worker-crash tests -passed, including a lost runner-grant response, target exclusion, final -subtree drain, fresh-session resources, and credential confinement. -An earlier unlocked harness run passed 380, failed two default-claim CLI -contracts, and skipped two; the lane failure explicitly reported another -workload holding `interactive.lock`. One focused retry of the two failed -CLI contracts under the shared native lock passed both (28 unselected -cases skipped). This collision is not native handoff evidence. - -**Review fixes:** live/draining grant capacity is reusable beyond 4096 -connections; bounded settled history never evicts pending/failed cleanup -proof. Busy/failed handoffs issue no runner grant. Descriptor roots and -workload/recovery holders refuse preempt with `target-busy` and an explicit -relinquish requirement. A preempted MCP child is discarded after authoritative -rejection so the same connection can reopen with fresh resources after runner -drain. - -Broker credentials are captured and scrubbed before user hook import. WDIO -receives each worker token through its per-worker `runnerEnv`, not launcher -`process.env`; workers scrub it before hooks/specs and their children execute. -Real hook-import, launcher-hook, worker-hook and spec-child probes pass. -The worker's Node bootstrap/preload window remains trusted because its child -token initially arrives in the fork environment; this is not same-account -hostile-code isolation. - -The focused commands below passed using the untouched default `TMPDIR`: - -| Exact command | Result | -| -------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-driver test 'authority/(grants\|delegation)\|server/(runnerHandoff\|delegation)' --runInBand` | Passed 51; includes 4,101 drained delegation cycles and root fencing. | -| `yarn workspace @fluentui-react-native/desktop-driver-mcp test server --runInBand` | Passed 31; includes preempt/reopen and stale-resource refusal. | -| `yarn workspace @fluentui-react-native/desktop-test test runner --runInBand` | Passed two; skipped the native opt-in case. Shared native lock used. | - -Review-validation logs are under `artifacts/desktop/f11/review/`. The complete -driver and harness suites used the shared native lock. Full-suite scratch was -project-local with matching `GIT_CEILING_DIRECTORIES`; no native opt-ins were set. -The Accessibility-blocked native contract was deliberately not repeated. - -After review fixes, `yarn lage test-links --verbose --grouped` passed all -90 tasks. Then `yarn clean-lage-cache` passed and the locked -`TF_BUILD=1 yarn lage buildci --concurrency 4` passed all 178 tasks in -2m 2.60s, with no pending or aborted tasks. Source was `5c9d0191e`; the -matching project-local `TMPDIR`/`GIT_CEILING_DIRECTORIES` isolation and native -opt-in skips were unchanged. Log: `artifacts/desktop/f11/review/buildci.log`. -The remaining ledger update records results only, not additional source changes. - -The earlier locked `TF_BUILD=1 yarn lage buildci --concurrency 4` at -`17ee49fd1` passed all 178 tasks, including the complete harness test task. -Project-local ignored `TMPDIR` also required `GIT_CEILING_DIRECTORIES` set -to that directory so temporary instruction-check fixtures did not discover -the enclosing worktree; the existing 24-case instruction checker suite -passed with that isolation. No baseline source was changed for this setup. -The aggregate log is `artifacts/desktop/f11/qualification/buildci.log`. -Task success does not turn the native opt-in skips into qualification. - -**Observed, native:** macOS 27.2.1, arm64, plain AppKit fixture (no React Native -renderer). The explicitly stable-signed helper was ready; its noninteractive -doctor reported Accessibility and PostEvent `false`, ScreenCapture `true`, -console connected and unlocked. No privacy or certificate-trust changes -were made. Native service/application identity and backend contracts passed -27 with 15 skipped Windows cases. AX-dependent handoff is **blocked: -Accessibility not granted to the signed helper**. These identity/backend -checks do not qualify GUI handoff. - -| Selection and exact command | Result | Cleanup and retained evidence | -| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| `FURN_NATIVE_DRIVER_TEST=1 yarn workspace @fluentui-react-native/desktop-driver test serviceIdentity.contract applicationIncarnation.contract nativeDriver --runInBand` | Passed 27, failed 0, skipped 15 | `artifacts/desktop/f11/qualification/native-identity.log` | -| `FURN_NATIVE_DRIVER_TEST=1 FURN_NATIVE_HANDOFF_APP_PATH= yarn workspace @fluentui-react-native/desktop-test test nativeRunner.contract --runInBand` | Failed 1; environment-blocked before a session opened: `The target does not support any click input mode.` Worker-crash scenario **not run**. | `artifacts/desktop/f11/native-handoff-5L2MVX/evidence.json`; fixture survived service shutdown, both descriptors removed, claim released; only its spawning test then ended the exact fixture process. | - -Both native commands and helper/fixture builds used the exclusive native -lock. The first handoff attempt used an older fixture binary without -`--ready-file` and failed readiness; it is retained at -`artifacts/desktop/f11/native-handoff-En89el/evidence.json`. The one corrected -fixture retry built and signed the readiness-capable source from -`27b109072`; it reached native-verified private discovery and the permission -gate above. Helper artifact identity: -`e4a23c0ec3cc9ddcf07fde2153bdf260b7b85f65bd0a6851275dd55a6ccf0fd3`. -Redacted logs and helper diagnostics are under -`artifacts/desktop/f11/qualification/`; private descriptors are not retained. - -**Not run:** successful native MCP → WDIO → MCP handoff, native worker-crash -drain, and Windows/Win32 native transport/lifecycle handoff. Existing PR -Windows/Win32 E2E and Storybook smoke jobs do not select this generic -descriptor-path contract; Windows CI is **not sufficient** for F11. -Storybook integration still requires F9, retained launches require explicit -F10 ownership, and cross-MCP-connection resume/replay remain unsupported. - -## Windows Fabric and Win32 Paper — F11 handoff qualification — 2026-11-12 UTC - -**Observed:** tested `c9753208d0f5602d28510d84ededf3f2fc3c3855` on -Windows x64, active/unlocked RDP, `sessionKind: remote`. The W1 campaign and -machine-wide native lock serialized all native/live work; the before/after -`-RequirePhysical` gates stayed eligible. The probes perform native -window discovery, observation and capture, not physical input. - -**As-is native consumer contract:** from this package, -`FURN_NATIVE_DRIVER_TEST=1 yarn test nativeRunner.contract --runInBand` -**skipped its one case** on Windows. The guard requires `darwin` and a signed -AppKit fixture; its CJS implementation hardcodes a macOS executable path, -bundle identity, endpoint and controls. Merely changing the Jest platform -guard cannot qualify Windows. `artifacts\desktop\f11\w1\native-runner-gate.json` -records the skip. A locked repeat with **both** opt-ins set, including -`FURN_NATIVE_HANDOFF_APP_PATH` naming the installed REX executable, still -skipped that case (`native-runner-gate-both-optins.json`). -A platform fixture/CI-selection change is proposed separately, -not implemented in this evidence-only lane. - -**Observed, real consumer:** the retained -`artifacts\desktop\f11\w1\qualify.cjs` composes the real supervisor service, -Windows-verified owner-only private descriptors, MCP SDK client/server and -WDIO/Jasmine. Commands, from the worktree root: - -```powershell -node artifacts\desktop\f11\w1\qualify.cjs windows -node artifacts\desktop\f11\w1\qualify.cjs win32 -``` - -Each command ran through `with-native-lock.ps1 -Lane W1 -RequirePhysical` -inside the granted W1 campaign. This is generic externally owned, -attach-only application handoff, **not** Storybook integration handoff. -Windows used the registered Fabric app's exact AUMID, with Metro from this -worktree via the app's declared `yarn start --no-interactive --port 8081`. -Win32 used the resolved prebuilt REX `ReactTest.exe` and its configured -component/title/production bundle, after `yarn storybook bundle --win32` -**passed** (bundle run `f13082d6-4b6c-428f-92e0-00a42035890d`). -The public native host owned each external launch separately from the service. -No native app rebuild was needed: the diff against the registered -`bdeb0b291` build changes no Windows app/native-core native source, only -native-core changelog/version metadata. Win32 native build is **unsupported**. - -| Case | Windows Fabric | Win32 Paper | Observed evidence | -| --------------------------- | ------------------------------------ | ------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Busy by default | Passed | Passed | Runner grant acquisition refused `target-busy` while MCP held the native session. | -| Explicit relinquish | Passed | Passed | MCP returned `relinquished: true`; WDIO attached; MCP was refused during runner ownership and opened a fresh session afterward. | -| Operator preemption | Passed | Passed | Service opted into preemption; old MCP binding failed closed (`service-unavailable`), then MCP was busy during WDIO and regained a fresh session after drain. | -| Normal runner exit | Passed | Passed | Native WDIO window query completed; each one-test run passed; fresh MCP attach succeeded after awaited runner cleanup. | -| Runner worker crash | Passed recovery; injected run failed | Passed recovery; injected run failed | Identity-checked literal `Stop-Process -Id 34752` / `Stop-Process -Id 29856` ran under the held lock. Parent detected the missing reporting handshake, drain completed and fresh MCP attach succeeded. | -| Lost acquisition response | Passed | Passed | The real service returned HTTP 200 from `/furn/grants/runner`; the transport discarded its body and threw before the typed client received it. Broker cleanup revoked/drained its parent; fresh MCP attach succeeded without knowing the lost runner token. | -| Old resources after handoff | Passed | Passed | Old observation reference/cursor and capture resource link were refused after a fresh session. | -| Credential confinement | Passed observed probes | Passed observed probes | Real worker spec environment contained no broker/worker credential variables; model output and 45 retained text/config/report files contained no bearer credential. Portable hook/spec-child probes also passed. | - -Native evidence roots, each containing `evidence.json`, terminal `run.json` -files, generated WDIO settings and command logs: - -| Endpoint / phase | Run ID | Terminal result | -| ---------------- | -------------------------------------- | ---------------------------------------------------------------- | -| Windows service | `78e18335-571f-41c7-a724-3bab30b41fd1` | Passed lifecycle; no test cases claimed | -| Windows normal | `a3bf02e2-df22-4a63-92f8-686057ef2a80` | Passed: one selected/executed/passed | -| Windows preempt | `24d38403-b4e5-49d6-9e0f-d1d123e4fb17` | Passed: one selected/executed/passed | -| Windows crash | `8f702133-e18a-41f0-b05d-ce5fcb5cff1a` | Failed as injected: one selected, zero completed, one incomplete | -| Win32 service | `f7ef9875-059f-432a-9daa-2bcb1e384917` | Passed lifecycle; no test cases claimed | -| Win32 normal | `d79fd672-df2d-4943-8b40-c017a8978380` | Passed: one selected/executed/passed | -| Win32 preempt | `f1f17781-08aa-4cac-a6d8-34cb13bd749b` | Passed: one selected/executed/passed | -| Win32 crash | `a09a8c20-1380-4933-a7a3-9c2d786352c3` | Failed as injected: one selected, zero completed, one incomplete | - -Windows root: `artifacts\desktop\f11\w1\windows-3tYoml`. -Win32 root: `artifacts\desktop\f11\w1\win32-BefbqE`. -The crash directories retain process identity and the literal-PID kill script. -Both runs used helper -`d8ee05e4796fb3aa22f7b5dbf481bd8cdfb0c20b6d5d24676d689dc7900d5b3c`. -The original external application process/start time remained running through -service shutdown; only its own recorded launcher then closed it. Both private -descriptors and each service workload claim were removed, with no cleanup -failure. The default per-user workload claim was also absent. **Inferred:** -claim release confirms the service's owned-resource settlement checks succeeded; -absence is not an additional persisted claim state. - -One initial Windows reproduction stopped at an overly strict nested -`assert.rejects` comparison in the local evidence harness, before relinquish. -Its `windows-dqAIsk/evidence.json` records complete cleanup; correcting that -assertion required no product change. Two storage-diagnostic invocations -failed because the external lock wrapper split a single command argument; -adding an unused second argument allowed the actual diagnostic to execute. -These are harness failures, not passing product cases. - -**Additional native contracts:** driver identity/backend selection passed -23, skipped 19 macOS cases, including **all 15 Windows cases** skipped in the -earlier macOS run. `protectedService.contract` passed three, with the -recovering-quarantine case **failed/environment-blocked before its assertions**: -project-local scratch ancestry was refused (`ancestor`). The default private -store opened and supported both live passes; no ACL was changed. -Per-case results and reasons: `artifacts\desktop\f11\w1\case-mapping.json`, -`protected-service.json`, `storage-probe.json`. Windows/Win32 -`yarn desktop-driver doctor --platform windows` and `--platform win32` -each **passed**, reporting `ready: true`; `doctor --permissions` returned -**unsupported**, not passed. - -**Portable checks, native opt-ins unset, uncached:** - -| Exact command | Result | -| --------------------------------------------------------------------------- | -------------------------------------- | -| `yarn build` | Passed | -| `yarn workspace @fluentui-react-native/desktop-driver test --runInBand` | Passed 1,211; skipped 88; failed zero | -| `yarn workspace @fluentui-react-native/desktop-test test --runInBand` | Passed 376; skipped eight; failed zero | -| `yarn workspace @fluentui-react-native/desktop-driver-mcp test --runInBand` | Passed 32; skipped zero; failed zero | - -Logs/JSON results are under `artifacts\desktop\f11\w1`. Test scratch was -project-local, with matching `TEMP`, `TMP` and `GIT_CEILING_DIRECTORIES`. -No install was needed. Existing `.changeset/scoped-runner-handoff.md` covers -the three published packages. No product source, test or CI configuration was -changed. Native Storybook handoff, retained launches, cross-connection resume -and replay remain outside F11's generic phase. - -The Windows/Win32 generic consumer remains absent from hosted selection; -Windows driver-native contracts and Storybook smokes do not replace it. -This coverage change is handed back for implementation. W1's campaign is -ended; final audit: `FURN-AUDIT: no FURN desktop processes found`. - -**Evidence validation:** `yarn markdown-link-check -c -.github\markdown-link-check-config.json -packages\agentic\desktop-test\references\current-status.md` passed -(15 links). Root `yarn oxfmt --check`, `yarn change:check` and -`git diff --check` passed. No clean-cache `buildci` rerun was required for -these evidence-only markdown edits. - -## Windows Ctrl+C regression and native consumer selection — 2026-11-12 UTC - -Source: review follow-up to `0ae47d15193770f7455db306b1e4297de76b4eaf`, -based on `04d8d8b13177727492ef204cf190840fa8a3f2a1` for #4491, with -the test/CI/documentation changes described here. This is -portable implementation evidence, not a new native qualification. No native -fixture/helper/app was launched and no physical-input campaign was acquired. - -**Observed:** `contracts` mode `agent-interrupt-cli` already passes against -that unchanged baseline. It runs the actual `serve --agent` CLI in a subprocess -with a protected fake-host service and actual descriptor/grant drains. It -checks bounded process exit, terminal `run.json`, exactly one `run-finished`, -descriptor removal and exact workload-claim disposition: - -| Semantic stimulus | Regression result | Terminal report / exit / claim | -| ---------------------------------------- | ----------------- | ---------------------------------------------------------------- | -| First SIGINT after ready | Passed | `passed`, exit 0, released | -| Second SIGINT during session drain | Passed | `passed`, exit 0, released | -| Required child failure before SIGINT | Passed | Expected `failed`, exit 1, released after settled cleanup | -| Session disposal failure during shutdown | Passed | Expected `failed`, exit 1, fenced with unresolved cleanup effect | - -These are in-process `process.emit('SIGINT', 'SIGINT')` dispatches in the CLI -process, not `GenerateConsoleCtrlEvent`, physical SendInput, or a Windows -console broadcast. These four cases do not themselves reproduce the historical -borrowed-scope self-wait and are not proof of its repair. The borrowed-scope -regression in -`packages\agentic\storybook-desktop\src\cli\DesktopStorybookCli.test.ts:836–861` -checks that current integrations leave the scope to its owner. - -**Observed historical red/green proof:** the added -`agent-interrupt-borrowed-cli` contract deliberately awaits the borrowed -scope's `close()` inside the integration after a single semantic SIGINT, -with a command-style protected fake-host service that settles first. -An isolated copy of the current built package, overriding only `cli.ts` -and `services.ts` with the exact source at -`f14e2542d7980c9fc6edfc419ddff6067c573ed6` (`586e4c385^`, before #4478), -**Fails as expected** after the 15-second subprocess deadline: - -- the exact child/claim owner PID matches; -- its one owned resource is settled, but the claim remains `active`; -- `run.json` is `incomplete`, `terminal: false`; -- borrowed close starts and never settles. - -The same contract on this branch **Passes**: exit 0, terminal `passed`, -one `run-finished`, both descriptors withdrawn, borrowed close settled and -claim released. Retained `review-fixes/historical-source.json` identifies both -historical blobs and source hashes; `historical-red.json`/`.log` and -`historical-green.json`/`.log` record the exact comparison. Historical TypeScript -was erased with the installed Babel TypeScript plugin in that scratch copy, -not checked into the working tree. This confirms the owner-side repair already -present from #4478 (`586e4c38545ca3eedccd1c9c817f2f89ce16219d`); no new runtime -patch is needed for this composition. Physical console Ctrl+C on the Windows -Fabric and Win32 Paper `serve --agent` paths remains **Not run** pending review -and the exclusive desktop grant. - -**Observed selection regression:** the extracted original native consumer -gate failed the Windows opt-in test (`Expected true; Received false`; four -other selection cases passed). After parameterization, the portable selection, -platform-fixture, hosted-capture-gate and exact native-readiness tests pass. -The macOS bundle executable/ready-file path is preserved. Windows uses the -prebuilt plain Win32 fixture with native exact executable/title/PID and -`1004: Ready` polling, not an invented ready-file option. No native source -or protocol change was needed. - -The native consumer now selects handoff, operator preemption, worker crash -and lost real HTTP-200 runner-grant response as separate cases. Its crash -fixture records the real worker session before exiting, and each mode checks -deletion before reuse, exact external-process incarnation, stale resources, -terminal reports, descriptor withdrawal and claim release. A cleanup-confirmed -failure permits the next case; uncertain cleanup prevents further launches. -Windows records use the validated default private store with unique names, -not unsafe project ancestry or repaired ACLs. Review corrections add a -120-second internal abort deadline, 30-second bounded shutdown and a -135-second Windows fixture watchdog. Passing processes must exit naturally; -an unref'd five-second post-shutdown watchdog fails lingering handles. -The parent knows each case's root and removes only its exited child's exact -descriptor records through private storage with PID/content checks; it never -releases claims or treats descriptor removal as proof of native cleanup. - -Both accessible Windows and Win32 Storybook PR jobs now explicitly provision -the fixture and select the generic consumer through -`run-native-runner-contract: 'true'`. The action requires all four ownership -cases to execute and pass; an all-skipped result is rejected. Both inner WDIO -specs must also execute/pass outside the deliberate worker crash, and the -action verifies `evidence.wdio` counts and accessibility success for the three -non-crash modes. A silently skipped inner input spec is not qualification. -Capture is a separate capability-gated case, skipped under hosted disabled-input policy; -core transport/ownership still executes without physical input. The plain -fixture is not React Native renderer or Storybook-handoff qualification. -The disabled-input capture gate now applies equally to macOS and Windows. -Endpoint refusal records `capture: "unsupported"` plus `captureSkipped: true`; -ownership/cleanup can still pass without a capture pass. - -Implementation checks and retained review artifacts are under -`artifacts\desktop\f11\w1-fix`: - -- `yarn workspace @fluentui-react-native/desktop-test test contracts --runInBand ---testNamePattern agent-interrupt-cli`: **Passed** before the new coverage - changes; `interrupt-first.json` records one Jest contract/four scenarios. -- `yarn workspace @fluentui-react-native/desktop-test test nativeRunnerSupport ---runInBand`: expected **Failed** before the gate fix (`native-selection-red.json`); - the focused combined selection/interrupt run then **Passed** 11, - **Skipped** 35 (`focused-green.json`, including name-filtered cases). -- Package `yarn format`, `yarn lint`, `yarn build`: **Passed**. Action/workflow - YAML and both new embedded PowerShell blocks: **Passed** syntax validation - without executing native code. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand`: - **Passed** 388, **Skipped** 12, failed zero (`portable-test.json` and - `portable-test.log`). Native opt-ins were unset; temporary data and - `HOME`/`LOCALAPPDATA` claims were isolated under project artifacts. -- New native consumer, optional native capture and physical Ctrl+C: - **Not run**; review and W1 desktop authorization are still required. - Hosted CI: **Not run** for the unpushed candidate. Clean-cache `buildci`: - **Not run**, reserved for final locked validation after review. - -Patch changeset: `.changeset/windows-native-runner-coverage.md`. - -## Windows review corrections — 2026-11-12 UTC - -Source: `0ae47d15193770f7455db306b1e4297de76b4eaf` plus the review-fix -working-tree changes, not a native-qualified commit. Endpoint/session kind: -not applicable to portable checks; no native run IDs, desktop campaign, -native helper/fixture launch or physical input. The original issue's physical -Windows Fabric / Win32 Paper console cases remain **Not run** for this -candidate pending review and the W1 grant. - -The historical comparison and review-regression red/green receipts are under -`artifacts\desktop\f11\w1-fix\review-fixes`. The new portable support regressions -first **Failed** 8 / **Passed** 10 (`support-red.json`), including the macOS -policy regression and missing Windows fixture lifetime. After correction, -`yarn workspace @fluentui-react-native/desktop-test test nativeRunnerSupport ---runInBand` **Passed** 18 (`support-green.json`). These do not launch native -processes or qualify the new Windows consumer. - -| Review regression | Observed portable result | Receipt | -| --------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- | -| Borrowed-scope self-close after command drain | Expected **Failed** on exact pre-#4478 CLI/services: 15-second timeout, active claim, settled resource, nonterminal `incomplete`, borrowed close still pending. Same byte-identical fixture **Passed** on the branch. Historical run: `8b6be099-3f0b-4551-9905-1a40b5104197`. | `historical-source.json`, `historical-red.json`/`.log`, `historical-green.json`/`.log` | -| Optional capture | **Passed**: both host policies skip disabled capture; endpoint refusal is explicitly unsupported/skipped without failing ownership. macOS bundle/ready-file arguments are unchanged. | `support-red.json`, `focused-green.json` | -| Deadline and failed-case records | **Passed**: ignored work still reaches bounded shutdown; hung shutdown is bounded; Windows lifetime is 135 seconds; cleanup preserves foreign/replaced records and uses exact contents for removal. Real private-store fallback remains **Not run**. | `focused-green.json` | -| Required inner input | **Passed**: one executed ownership spec cannot qualify the consumer. The actual old action guard incorrectly accepts that fixture; the new guard rejects it and accepts two executed/passed specs. No native step was executed for this comparison. | `ci-input-red-green.json`, `focused-green.json` | - -Final portable validation from `D:\dev\furn-w1`: - -- `yarn workspace @fluentui-react-native/desktop-test test contracts nativeRunner ---runInBand --json --outputFile -D:\dev\furn-w1\artifacts\desktop\f11\w1-fix\review-fixes\focused-green.json`: - **Passed** 50, **Skipped** 5, failed zero. -- `yarn workspace @fluentui-react-native/desktop-test test --runInBand --json ---outputFile D:\dev\furn-w1\artifacts\desktop\f11\w1-fix\review-fixes\portable-test.json`: - **Passed** 396, **Skipped** 12, failed zero. Native opt-ins were unset; - test scratch and default claims stayed under isolated project artifacts. -- Package `yarn format`, `yarn lint`, `yarn build`: **Passed**. Root - `yarn format:check`, `yarn check-agent-instructions`, and `git diff --check`: - **Passed**. -- `yarn markdown-link-check -c .github\markdown-link-check-config.json ` - for the four edited references and patch changeset: **Passed** (20 links). - Both action/workflow YAML files and all eight actual embedded PowerShell - blocks: **Passed** syntax-only validation (`yaml-syntax/summary.json`). -- Native consumer modes, optional native capture, physical console Ctrl+C, - hosted CI and locked clean-cache `buildci`: **Not run**, reserved for the - coordinator's post-review W1 grant and final validation. Post-commit - changeset checking is recorded in the review handoff receipt. - -## Windows portable clean-cache/buildci — review fixes — 2026-11-12 UTC - -Tested source: `d94c8787dd76a1be08cbf2827cc57edb6a3a18f2`, including the -review fixes at `78236d3e8502b743055bf5cf3c48eed58a8926a3` and two -validation-guidance commits. Scope: portable repository validation only. -The coordinator authorized the W1 lock without `-RequirePhysical` while -the desktop was disconnected; no campaign, app, native fixture/helper -launch, native/GUI contract or physical console Ctrl+C was run. Endpoint, -session kind and native run IDs are not applicable. - -Receipts are under `artifacts\desktop\f11\w1-fix\portable-buildci`. -All aggregate commands ran through `with-native-lock.ps1 -Lane W1` -from the worktree root, with `FURN_NATIVE_DRIVER_TEST` unset: - -- `cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"`: - cache clear **Passed** once; the initial aggregate **Failed**. With - project-local temporary files, Git discovery escaped synthetic test - repositories into the enclosing ignored worktree. The scripts suite - reported 58 passed / 18 failed (`buildci.log`). -- Setting `TEMP`, `TMP`, `TMPDIR` and `GIT_CEILING_DIRECTORIES` to the same - project-local scratch directory restores fixture isolation without a - product-code change. `yarn workspace @fluentui-react-native/scripts test`: - **Passed** 76 / failed zero (`scripts-git-isolation-green.log`). - The operating guidance is committed in `8cbead3f6`. -- The default-parallel `yarn lage buildci` rerun **Failed** the existing - 15-second cold-worker `StorybookUI` test (`buildci-rerun.log`). - `yarn workspace @fluentui-react-native/storybook-desktop-runtime test -StorybookUI --runInBand`: **Passed** both unchanged tests - (`storybook-ui-focused.log`); no assertions or deadlines were relaxed. -- `yarn lage buildci` with `LAGE_POOL_CONCURRENCY=2` and the isolated Git - boundary: **Passed**, exit zero, 5m 55.78s (`buildci-bounded.log`). - Lage reported 81 successful tasks, 97 reused fresh-cache tasks, and zero - failed/pending/aborted tasks. Cache reuse follows the one initial clear; - it is not native capability coverage. Scheduling guidance is committed - in `d94c8787d`. - -**Observed hygiene:** the pre-run audit was clean. After the first failed -aggregate, its portable Node test children drained naturally and a later -audit was clean; an aborted fake-host runner fixture was preserved outside -the formatted source tree. No process was terminated. The final aggregate -released the W1 lock by `2026-11-12T17:45:39.6650962Z`. The required -post-buildci audit returned exit 2, `FURN-AUDIT: 6 candidate process(es):`, -all foreign W3 Node test/build processes started after that release; no W1 -candidate survived (`post-buildci-audit.log`). Those processes were not -touched; that shared-host snapshot was not globally clean. The later -closeout audit **Passed**, exit zero: -`FURN-AUDIT: no FURN desktop processes found` (`closeout-audit.log`). - -Root `yarn format:check`, `yarn check-agent-instructions`, Markdown link -checks for the edited development guide, and committed `yarn change:check`: -**Passed**. Native consumer modes, the repeated native driver contracts, -optional native capture, physical console Ctrl+C, and hosted CI remain -**Not run**. No push was made. - -## Windows default-environment acceptance and guidance reverts — 2026-11-12 UTC - -This supersedes the preceding controlled-environment acceptance receipt. -Tested source: `015998677c0912625e0f7e30b107c674a90fec4c`. Receipts: -`artifacts\desktop\f11\w1-fix\default-buildci`. - -**Observed original failures:** the first aggregate failed only -`@fluentui-react-native/scripts#test`: 18 negative assertions in -`scripts/src/tasks/agentInstructions.test.ts` expected diagnostics but received an -empty issues list. `exact-failures.json` records every exact test name, -location and `AssertionError [ERR_ASSERTION]` message. The earlier -project-local `TEMP` was an execution choice to keep transient files local, -not an F11 or nativeRunner requirement. It made the synthetic repositories -discover their ignored enclosing worktree. - -The checker/test blobs match branch base `04d8d8b13` and local `origin/main` -`505a2187b` exactly. An extracted, byte-identical base-source copy reproduced -all 18 failures with that override; its source files were then removed. -`baseline-red.json` and `source-comparison.json` retain the proof. The -separate later `StorybookUI` failure was `Exceeded timeout of 15000 ms for a -test` in `resets the whole Win32 preview for a new test generation without -resetting chrome preferences`. The entire runtime package is identical at -base/main/head; both unchanged focused tests pass with inherited defaults. -No full baseline aggregate is claimed, and no new F11 test required a fix. - -The first default-environment aggregate passed the instruction tests but -failed the existing driver test `abandons a walk when the caller -disconnects, journals it as cancelled, and keeps the session usable` at -`packages/agentic/desktop-driver/src/server/operations.test.ts:405`: expected `unique`, received -`inconclusive`. That test and find implementation match base/main; all -39 unchanged focused tests pass, and the unchanged default aggregate -retry passed. **Inferred:** these aggregate-only failures are timing-sensitive, -not evidence that the branch needs altered repository-wide test settings. - -Both out-of-scope guidance commits were reversed with new `git revert` -commits: `b5b82795a` reverts `d94c8787d`; `015998677` reverts `8cbead3f6`. -The development guide is restored exactly to its pre-guidance contents. -No assertions, test budgets, product code or machine-wide settings changed. - -**Final fresh acceptance:** from the worktree root, through -`with-native-lock.ps1 -Lane W1` without `-RequirePhysical`, -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` **Passed**, -exit zero, 2m 45.87s (`final-default-buildci.log`). `TEMP`/`TMP`/`TMPDIR` -were inherited unchanged; concurrency was default; Git-ceiling, Lage-pool -and native-driver opt-ins were absent. All 178 Lage tasks succeeded, with -zero cache hits, failures, pending or aborted tasks. Native capability -skips are not passes. - -The post-run process audit **Passed**, exit zero: -`FURN-AUDIT: no FURN desktop processes found` (`final-default-audit.log`). -No app, native fixture/helper, native/GUI contract or physical Ctrl+C was -launched; no campaign or push occurred. Native qualification remains -**Not run**, awaiting its separate grant. - -## Windows hosted nativeRunner provisioning correction — 2026-11-12 UTC - -**Observed:** PR workflow `43279226602`, branch source -`23f974a727ca2627690304cbecee0fa96cb6de83`, selected the new Windows consumer -but did not reach it. Windows Storybook job `131395742955` and Win32 -accessible Storybook job `131395742289` both failed while importing the -driver's built fixture provisioner: `ERR_MODULE_NOT_FOUND` for an extensionless -dependency. All four nativeRunner ownership modes and their accessibility-input -specs are therefore **Not run**, not skipped or passed. No flake rerun was used. - -The owning driver repair is tested at -`0e2dcb26ca156133cb577fc81869b23cebf1df3a`: explicit Node ESM imports and the -existing package-root anchor replace Jest-only resolution and `__dirname`. -The portable regression reproduces the hosted failure before the correction -and passes afterward; it never compiles or starts the native fixture. The -fixture, native protocol, consumer opt-ins, cleanup rules and macOS path are -unchanged. The driver's provisioning receipt is in -`packages/agentic/desktop-driver/references/current-status.md`. - -**Observed, portable:** from the repository root, -`yarn workspace @fluentui-react-native/desktop-test test nativeRunnerSupport ---runInBand` **Passed**, all 18 cases. The driver suite passed 1,212 and skipped 88. The locked default-environment -`yarn clean-lage-cache && yarn lage buildci` **Passed**, all 178 tasks with -zero cache hits or failures, 2m 52.86s. Package format/lint/build, root format -and instruction checks, changed-document links and committed changesets passed. -Receipts are in `artifacts\desktop\f11\w1-fix\push-ci`. - -No local native/GUI workload or physical input was launched; no desktop -campaign was started. Hosted requalification is pending the correction's push. -Local nativeRunner and physical console Ctrl+C remain **Not run** while the -desktop session is disconnected. This receipt does not upgrade the earlier -connected endpoint evidence to a rerun of the new consumer. - -## Windows native runner and physical console interruption — 2026-11-12 UTC - -**Observed, source `b1bbb26df880252a184ea5a3d2e1848bd811e83f`:** the W1 -campaign ran with an active, unlocked RDP session (`sessionKind: remote`). -Every live command used the exclusive native lock and `-RequirePhysical`. -No input-feature mask, privacy change or implicit native download was used. - -The explicitly provisioned plain Win32 fixture now exercises the generic -MCP/WDIO consumer on Windows. From the repository root, with -`FURN_NATIVE_DRIVER_TEST=1` and `FURN_NATIVE_HANDOFF_APP_PATH` naming that -fixture, `yarn workspace @fluentui-react-native/desktop-test test -nativeRunner.contract --runInBand --json --outputFile -\artifacts\desktop\f11\w1-native\native-runner-jest.json` **Passed** all five -outer cases, with zero skips. This fixture is not a React Native renderer. - -| Mode | Result | Inner WDIO run ID | Executed / passed | -| -------------------------- | --------------------------------------------- | -------------------------------------- | ---------------------------------------- | -| Handoff | Passed | `a9849169-7f2b-46e6-87de-3e0b8b8e7a3f` | 2 / 2 | -| Operator preemption | Passed | `2a1f63bd-3ac0-4951-beb5-860a2a92b69c` | 2 / 2 | -| Worker crash | Passed recovery; expected terminal failed run | `9c1a8447-2a6c-4ca6-b027-8bc302df3fc7` | 0 / 0; interrupted by the injected crash | -| Lost runner-grant response | Passed | `4e7d50fa-b055-407c-bfe2-443fe2bc504f` | 2 / 2 | -| Capture | Passed, including a real target capture | `047e8166-c3d2-4fb9-b785-288519b3342d` | 2 / 2 | - -Every non-crash mode executed both ownership and accessibility-input specs, -with zero inner skips. Every mode preserved its exact external fixture -incarnation through service cleanup, withdrew both descriptors, released -its workload claim, wrote terminal reports, and then ended its own fixture. -No cleanup deadline fired. Full service run IDs and per-mode evidence are in -`artifacts\desktop\f11\w1-native\contracts.json`; the corresponding roots are -`native-handoff-5ZFlJt`, `native-operator-preempt-arz9P9`, -`native-worker-crash-FyizDx`, `native-lost-response-7745sB` and -`native-capture-6v7nle` under `artifacts\desktop\f11`. - -**Observed, physical console Ctrl+C:** the real app consumer ran -`yarn desktop-test --config --target - --reporter json serve --agent --descriptor `. -The ignored qualification config imports the app's actual config and adds -only the verified registered Fabric AUMID/title; it does not replace the -Storybook integration. A separate owned observer ran the repository's Yarn -entry in a new visible console. Console process inspection proved isolation -from the controller, and the emitter checked that exact window was foreground -immediately before inserting Ctrl-down, C-down, C-up and Ctrl-up with -`SendInput`. The observer recorded real console `SIGINT`; no -`GenerateConsoleCtrlEvent` or injected Node signal was used. - -| Endpoint | Result | Lifecycle run ID | Exact CLI / Yarn / observer exits | -| --------------------------- | ------ | -------------------------------------- | --------------------------------- | -| Windows Fabric packaged app | Passed | `75c40800-571c-4d8f-9fa1-debf876f0664` | 0 / 0 / 0 | -| Win32 Paper REX | Passed | `9a5c618f-d54d-4599-b36f-e3fd5b5436b9` | 0 / 0 / 0 | - -Both apps first completed authenticated `components-button--appearance` -selection with native marker verification and a native MCP observation. -The MCP session/adapter was closed before interrupting the still-running -service. Each interrupt inserted exactly four key events, wrote a terminal -`passed` lifecycle report, released the per-user claim and withdrew the -descriptor. Each borrowed app remained running through service shutdown, -then its exact recorded launcher closed it. No owned service/app survived. -This confirms the inherited #4478 repair on the physical Windows console -path; the historical borrowed-close fixture's red-before/green-after proof -remains separate from this physical evidence. - -Reproduction owners are retained as -`node artifacts\desktop\f11\w1-native\physical-case.cjs windows` and the -same command with `win32`, each inside the granted W1 campaign and -`with-native-lock.ps1 -Lane W1 -RequirePhysical`. Physical receipts are in -`physical-windows-UmfSj6` and `physical-win32-sssEyg` beneath that root. -The unchanged registered Fabric native build `bdeb0b291` was reused after -package identity and native-source comparison. Two earlier Fabric setup -attempts were **Not run** for Ctrl+C: missing generated renderer properties, -then no local built manifest. `yarn storybook prep --windows` and the -supported explicit registered identity resolved those prerequisites; both -failed startup attempts exited terminally with no retained claim. -Win32 native build is **Unsupported**; the prebuilt REX host used a fresh -`yarn storybook bundle --win32` with the serving instance's port, ID and -driver manifest, as the platform recipe requires. - -**Hosted correction:** workflow `43284867711` at prior branch head -`644388cc32e3b65854572ebe18abef6ecdce93b5` provisioned the fixture and -executed four ownership modes in Windows job `131415841258` and Win32 -accessible job `131415842551`. Each passed handoff, preemption and crash -recovery; handoff/preemption each executed both inner specs. Lost-response -failed before WDIO because its assertion expected an unwrapped error message; -capture was policy-skipped. This was not a flake. The current correction -shares that real HTTP-200 response-loss fault with the portable HTTP runner -and checks the typed transport error's exact original cause instead. It was -red with the hosted message mismatch, green after repair, and now passes the -native lost-response case above. The full portable harness suite **Passed** -396 with 12 native/platform skips; build and lint also passed. - -The campaign is ended. The post-campaign broad audit reported four -**foreign portable** W5 Jest/Yarn processes, no listeners and zero owned -native/GUI survivors; none was terminated. Exact audit, session and lock -receipts are under `artifacts\desktop\f11\w1-native`. Hosted confirmation of -the corrected head is pending publication at this receipt; those earlier -partial hosted counts are not a four-mode CI pass. - -**Portable closeout, not aggregate acceptance:** the locked default-environment -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` **Failed** in -three fresh attempts at this source. Each failed only -`@fluentui-react-native/storybook-desktop#test` → -`desktop Storybook server integration` → -`hosts the Storybook channel and Stage 1 smoke-test target in one process`, -with `A stopping command cannot publish readiness.` at -`packages/agentic/desktop-test/src/processControl.ts:112`. -Both that lifecycle source and -`packages/agentic/storybook-desktop/src/driver/serverIntegration.test.ts` -are byte-identical at main, review base `04d8d8b13` and the tested head. -Timing sensitivity is **Inferred**, not a reproduced base-runtime result: -`yarn workspace @fluentui-react-native/storybook-desktop test -serverIntegration --runInBand` **Passed** 9 with one platform skip, and -the whole unchanged package's default-concurrency -`yarn workspace @fluentui-react-native/storybook-desktop test` **Passed** -340 with three skips. No unrelated source or test deadline was changed. -TEMP/TMP/TMPDIR and concurrency stayed at inherited defaults; native opt-in, -Git-ceiling and Lage pool overrides were absent. This failed aggregate does -not replace the earlier source's passing receipt or qualify this head. - -The final attempt began with a clean audit; its same failure therefore is -not attributed to the foreign portable work. Its fail-fast return briefly -left owned portable Jest workers running; they subsequently exited -naturally, without a process kill. The foreign process owner separately -removed its abandoned W5 chain. The final audit **Passed**: -`FURN-AUDIT: no FURN desktop processes found`. Receipts -`default-buildci{,-rerun}.json`, `clean-default-buildci.json`, -`aggregate-readiness-source-comparison.json`, -`aggregate-readiness-{focused,package}.json` and -`aborted-aggregate-cleanup-audit.log` are under the same evidence root. -Formatting, instruction validation and all three changed-ledger link checks -also **Passed**. - -## Windows default-aggregate follow-up — 2026-11-12 UTC - -**Observed, source `2bf750f9e87919124eee9e0c567eb377603cc9cd`:** the final -locked default-environment -`cmd.exe /d /c "yarn clean-lage-cache && yarn lage buildci"` **Passed** -all 178 tasks, zero failed, from **20:52:21 to 20:55:29 UTC**. -TEMP/TMP/TMPDIR and concurrency remained inherited defaults; native opt-in, -Git-ceiling and Lage pool overrides were absent. No source, test deadline, -dependency manifest or environment guidance changed during this follow-up. -This supersedes the earlier lack of aggregate acceptance for this source, -without erasing the three failed attempts. - -The exact earlier failure was -`@fluentui-react-native/storybook-desktop#test`, -`packages/agentic/storybook-desktop/src/driver/serverIntegration.test.ts` → -`desktop Storybook server integration` → -`hosts the Storybook channel and Stage 1 smoke-test target in one process`: -`A stopping command cannot publish readiness.` at -`packages/agentic/desktop-test/src/processControl.ts:112`. - -**Controlled baseline:** a new detached worktree at -`origin/main` (`505a2187b0a2451aed7081c9260b343d55742f93`) used its own -immutable dependency restore, not linked branch build output. The same -locked default command **Failed** at 20:42:15–20:44:20 UTC, but differently: -`@fluentui-react-native/desktop-driver#test`, -`packages/agentic/desktop-driver/src/server/operations.test.ts` → -`Desktop Driver agent operations` → `the caller leaving` → -`abandons a walk when the caller disconnects, journals it as cancelled, and keeps the session usable`. -The post-cancellation `session.find()` expected `outcome: "unique"` and -received `"inconclusive"` at line 405. Storybook's task was incomplete, -so that run does **not** reproduce the readiness failure on main. -One unchanged default repeat at 20:47:34–20:50:39 UTC **Passed** all 178 -tasks, including Storybook. Both baseline audits were clean; the scratch -worktree and its remaining empty directories were removed before the head run. - -**Persistent-state audit, 20:41–20:46 UTC:** - -| Location / state | Observation and timestamp | -| -------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| `%LOCALAPPDATA%\Microsoft\FluentUIReactNative\desktop-test\workload` | Empty: no interactive lock, claim, report or recovery file. Last modified at the initial inventory: 20:04:25 UTC. | -| `%LOCALAPPDATA%\com.microsoft.fluentui-react-native.desktop-driver` | No agent, runner or active service descriptor. Thirteen durable settlement records; directory last modified 19:40:25 UTC before cleanup. | -| `%LOCALAPPDATA%\Microsoft\FluentUIReactNative\desktop-driver\native` | Shared immutable native cache, not a descriptor/claim store; existing cache root dates from August 31. Left untouched. | -| `apps\storybook\storybook-desktop.generated` | No lease, claim, lock, ready or connection files. Windows/Win32 manifests modified 19:36:48 / 19:40:06 UTC; runtime instance 19:40:08 UTC; retained generated WDIO spec 18:00:37 UTC. | -| `apps\storybook\artifacts\desktop` | All eight existing `run.json` reports terminal. Physical Fabric/Paper reports last modified 19:37:28 / 19:40:25 UTC. No per-user `run.json` found. | -| Registered Fabric app | Existing verified `furn-81` build; package directory / manifest modified November 10 at 15:08:37 / 15:08:26 UTC. No registration change made by this campaign. | -| Firewall / loopback | Existing app rule and exemption only, queried read-only at 20:41:51 UTC. Those APIs expose no modification timestamps. Neither setting was changed. | - -Two bridge-epoch records belonged exactly to the physical campaign's Fabric -and Paper application incarnations, created at 19:37:24 and 19:40:22 UTC. -At **20:46:53 UTC**, the verified Windows storage/probe APIs confirmed -each recorded PID/start-time pair against its own qualification receipt -and proved that exact application had exited. Compare-and-remove deleted -only those two unchanged records. The other eleven records, including -unrelated quarantines, were preserved. These were settled recovery records, -not active claims or live descriptors. No process was killed. - -**Diagnosis boundary:** identical source was not runtime proof of -pre-existence. Main's first error differed; its repeat and the final head -both passed. The readiness test waits 20 seconds, and the reported guard -rejects a producer frame received after stop begins. Aggregate CPU load was -observed at 96% across 24 logical processors; load-sensitive startup/stop -timing remains **Inferred**, not an established root cause. The fake-host -server path does not open the native service/recovery store. Neither a -same-error main reproduction nor a causal repair by deleting those epoch -records is claimed. No unrelated runtime or timeout workaround was added. - -Final post-head process audit **Passed**: -`FURN-AUDIT: no FURN desktop processes found`. Baseline/head command -receipts, state inventories, exact-identity cleanup, scratch removal and -audits are retained in `artifacts\desktop\f11\w1-buildci-followup`. - -**Hosted confirmation:** workflow `43297198219` at branch source -`2bf750f9e` completed **successfully, zero failed jobs**. Windows Storybook -job `131458546719` and Win32 accessible job `131458546805` each recorded -four ownership cases passed, one capture-policy skip and zero failures. -Handoff, preemption and lost-response each executed/passed both inner -specs with zero skips and `accessibilityClick: passed`; worker-crash -recovery passed around its deliberately failed inner run. Downloaded -artifacts, not outer Jest totals alone, establish these counts. No flake -rerun was used. The existing physical/native qualification remains -separate; this follow-up launched no app and made no new native claim. - -## F11 rebased macOS positive qualification — 2026-11-12 UTC - -**Observed source:** `5dbadb53b10cd81ccf1e15739e213f0046d93323`, rebased onto -F6 squash/main `041ac8b9894aeaa7e1aa273baa2ece4e46298b94`. -The original branch/lease was `11820d6330f4ea591bed400193b13e793970b5bd`; -the historical base was `505a2187b0a2451aed7081c9260b343d55742f93`. -The owning worktree first fast-forwarded its older `c9753208d` tip, retaining -all fourteen Windows additions, then replayed all 22 branch commits. The -only conflict appended F6 and F11 evidence at the same location in this file: -both complete sections were retained. Range-diff reports 20 identical patches -and two documentation-context-only changes. Windows fixture/provisioning, -nativeRunner, input and response-loss code, hosted actions, and Windows ledgers -were preserved. No new runtime or native-provider behavior was changed. - -### Real native consumer - -macOS **27.2.1, arm64**, plain AppKit ownership fixture; no React Native -renderer. The explicitly provisioned, signature-verified -`FurnOwnershipFixture.app` uses the readiness-capable source -`27b109072002e5bd83a35c97ff9633311a74e5c7`, bundle identifier -`com.microsoft.furn.application-ownership-fixture`, unique per-mode titles, -and its exact ready-file/PID. It is not downloaded or compiled by the consumer. -The stable-signed helper artifact is -`e4a23c0ec3cc9ddcf07fde2153bdf260b7b85f65bd0a6851275dd55a6ccf0fd3`, -source digest -`b1d5233de326faf886c8d6310c2844c6857e4936686dfd493ff898f9e5ec6b9a`. -Noninteractive doctor reported **Accessibility, PostEvent and ScreenCapture -all true**, console connected/unlocked, and `promptRequested: false`. -The earlier permission-blocked macOS attempt remains historical, not the -current F11 qualification boundary. No permission, TCC or certificate change -was made. - -All commands below ran from the repository root. Native/helper commands, -portable suites and the final aggregate used the shared exclusive native lock. -`SIGNING` below denotes the existing `FURN Desktop Driver Development` -identity; the actual native commands set -`FURN_DESKTOP_DRIVER_MACOS_SIGNING_IDENTITY="FURN Desktop Driver Development"`. -`FIXTURE` denotes the absolute path to the owning worktree's -`artifacts/desktop/f11/application-fixture/FurnOwnershipFixture.app`. - -| Exact command (with the stated native environment) | Result | -| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | -| `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver build-driver --platform macos --json` | **Passed**, verified existing stable-signed artifact | -| `yarn workspace @fluentui-react-native/desktop-driver exec desktop-driver doctor --platform macos --permissions --json` | **Passed**, all three permission preflights true; no prompt | -| `codesign --verify --strict artifacts/desktop/f11/application-fixture/FurnOwnershipFixture.app` | **Passed** | -| `FURN_NATIVE_DRIVER_TEST=1 FURN_NATIVE_HANDOFF_APP_PATH="$FIXTURE" yarn workspace @fluentui-react-native/desktop-test test nativeRunner.contract --runInBand` | **Passed 5/5, zero skips**, 31.267 seconds | -| `FURN_NATIVE_DRIVER_TEST=1 yarn workspace @fluentui-react-native/desktop-driver test serviceIdentity.contract applicationIncarnation.contract nativeDriver --runInBand` | **Passed 28; skipped 15 Windows cases**. Selection includes portable built-ESM provisioning regression, not 28 purely native tests. | - -| Mode / receipt directory under `artifacts/desktop/f11/` | Inner WDIO run ID | Observed outcome | -| ------------------------------------------------------- | -------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -| `native-handoff-cZN2uI` | `737589dd-cd90-493f-8bb1-251a4e057b72` | Explicit relinquish; 2 selected/executed/passed, zero skips | -| `native-operator-preempt-gaSxgd` | `76a5b401-0022-417c-a559-87b5d24c3753` | Operator-authorized preemption; 2 selected/executed/passed, zero skips | -| `native-worker-crash-D3uh4Z` | `861bc143-5388-4802-8194-e323ef317ee2` | Outer native drain/reuse contract passed; deliberate exit 17 leaves terminal **failed** inner run with 1 selected/incomplete, 0 completed; not an inner input pass | -| `native-lost-response-kklnxj` | `96b708e7-63a9-44c1-a0a3-eb3f4e87ff6b` | Real successful grant-response loss and confirmed drain; 2 selected/executed/passed, zero skips | -| `native-capture-xGUWSZ` | `791c6749-826e-4929-b35a-cf1cb8b9cadf` | Target-scoped capture; 2 selected/executed/passed, zero skips | - -Every non-crash mode passed the real accessibility-click spec, not merely the -outer Jest assertion. Every mode opened native sessions through verified -private discovery and the MCP SDK, exercised actual WDIO/Jasmine, rejected -stale resources, confirmed worker drain and credential confinement, withdrew -both descriptors, released its claim, and preserved the exact external -fixture incarnation through service shutdown. Only its owning test then -terminated that fixture and verified exit. All five receipts report -`cleanupConfirmed: true`; no case deadline or cleanup timeout fired. Captures -were confined to the registered fixture. No Metro or Storybook app was started. - -### Portable checks and scope - -`yarn install --immutable` **Passed** after the F6 manifest/lockfile update -(existing peer warnings). For each of `desktop-driver`, `desktop-test` and -`desktop-driver-mcp`, the exact commands -`yarn workspace @fluentui-react-native/ format`, -`yarn workspace @fluentui-react-native/ lint` and -`yarn workspace @fluentui-react-native/ build` each **Passed**. - -| Exact portable command | Result | -| ---------------------------------------------------------------------------------- | ------------------------------------------------------------------ | -| `yarn workspace @fluentui-react-native/desktop-driver test 'authority/(grants | delegation) | server/(runnerHandoff | delegation)' --runInBand` | **Passed 51** | -| `yarn workspace @fluentui-react-native/desktop-test test runner --runInBand` | **Passed 20; skipped 5 native opt-ins**, separately executed above | -| `yarn workspace @fluentui-react-native/desktop-driver-mcp test server --runInBand` | **Passed 31** | -| `yarn workspace @fluentui-react-native/desktop-driver test --runInBand` | **Passed 1,235; skipped 66** | -| `yarn workspace @fluentui-react-native/desktop-test test --runInBand` | **Passed 402; skipped 6** | -| `yarn workspace @fluentui-react-native/desktop-driver-mcp test --runInBand` | **Passed 32** | - -`yarn lage test-links --verbose --grouped` **Passed all 90 tasks**. -The first locked plain `yarn clean-lage-cache && yarn lage buildci` -cleared the cache, then **failed** `check-agent-instructions`: two newly added -ledger links crossed package guidance boundaries (43 successful tasks, -85 pending). They were corrected to package-qualified text references; this -was a documentation error, not a native or runtime failure. - -The corrected `yarn check-agent-instructions` **Passed**. The next locked -plain `yarn lage buildci` **Failed** the driver task: `boundedCommand.test.ts` -expected `SIGKILL` but received `SIGTERM`, and the first macOS private-storage -publication test exceeded its 10-second budget (1,233 driver tests passed, -66 skipped, two failed). Both test files and their implementations are -byte-identical to pinned main. This establishes unchanged source, not a newly -executed baseline reproduction or the private-storage timeout's cause. -The locked diagnostic -`yarn workspace @fluentui-react-native/desktop-driver test boundedCommand privateStorage --runInBand` -**Passed 73, skipped 10** without source or deadline changes. - -One unchanged, locked plain `yarn lage buildci` repeat **Failed again** on -the same bounded-command `SIGKILL`/`SIGTERM` assertion: driver **1,234 passed, -66 skipped, one failed**; private storage passed. Lage reported **61 successful, -109 cached/skipped tasks, one failed**, with other tasks interrupted, so no -complete aggregate acceptance is claimed. The known parallel child-startup -race was not hidden, patched, or worked around; the diagnostic pass is not -aggregate acceptance. `TF_BUILD` and native opt-ins were unset, inherited -`TMPDIR` and default concurrency were unchanged, and the cache was cleared -once before the first aggregate. The exact failures remain in `buildci.log`, -`buildci-corrected.log`, and `buildci-default-repeat.log` beside the focused -diagnostic receipt. **The macOS positive native gate is closed; final portable -aggregate acceptance remains blocked.** - -Receipts and doctor output are under -`artifacts/desktop/f11/requalification-20261112/`; per-mode `evidence.json` -and terminal `run.json` files remain in the directories above. No private -descriptor or grant is copied into shared evidence. - -**Boundaries:** generic attach-only macOS handoff is positively qualified. -Existing Windows local/hosted evidence above is preserved, not re-executed on -this Mac or asserted for the rewritten SHA. Storybook/F9 handoff, F10 retained -launches, React Native renderer qualification, physical pointer/keyboard input, -cross-connection resume and automatic replay are **not qualified by this run** -and remain outside this phase. Publication, post-rebase hosted CI, and the PR -draft/readiness decision remain coordinator-owned. +# Current status + +This document states what the neutral desktop workload host does today, what it deliberately refuses, and the latest native result per endpoint. +Done means the stated implementation exists and is covered by deterministic tests; it does not mean every native endpoint is qualified. +The package is the intended neutral workload host; the existing Storybook lifecycle takeover is transitional. +No new coordinator package, daemon, dynamic plugin loader, or package rename is required. +Parent issue: [#4435](https://github.com/microsoft/fluentui-react-native/issues/4435). +See [lifecycle](lifecycle.md), [commands](commands.md), [native qualification](native-qualification.md), and `@fluentui-react-native/desktop-driver/references/current-status.md` for the native and driver boundaries. + +## Status overview + +Configuration, service composition, reporting, cleanup fencing and generic handoff are implemented. +macOS application ownership and Windows/Win32 application ownership are implemented and qualified on the endpoints named below. +Storybook lifecycle relocation, management transport, repeated-run reliability and required native CI remain open. +Native results are per endpoint and tied to the revision that produced them. + +- **App/suite configuration and standalone use:** Done. Storybook is optional and the runtime boundary rejects React, React Native and Storybook dependencies. See [Configuration and standalone use](#configuration-and-standalone-use). + +- **Service and helper composition:** Done. The generic supervisor and the supervised Storybook child share one `driverService` factory. See [Service and helper composition](#service-and-helper-composition). + +- **WDIO/Jasmine execution, required services and reporting:** Done. See [WDIO execution and reporting](#wdio-execution-and-reporting). + +- **`serve --agent` descriptor publication:** Done. See [Agent descriptor publication](#agent-descriptor-publication). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Cleanup fence and lifecycle settlement:** Done. Failed disposal keeps the claim and never reclaims automatically. See [Cleanup fence and lifecycle settlement](#cleanup-fence-and-lifecycle-settlement). Tracking: [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451), [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Typed resource contributions and durable quarantine:** Done. Live restart qualification is separate. See [Resource contributions and durable recovery](#resource-contributions-and-durable-recovery). + +- **Supervisor crash containment:** In progress. Portable contracts pass; live macOS survivors beyond the known descendant are unqualified. See [Supervisor crash containment](#supervisor-crash-containment). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Application ownership:** Qualified on macOS Fabric, macOS Paper, Windows Fabric (packaged) and Win32 Paper. See [Application ownership](#application-ownership). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441); landed in [#4492](https://github.com/microsoft/fluentui-react-native/pull/4492), [#4494](https://github.com/microsoft/fluentui-react-native/pull/4494). + +- **Generated manifest provenance:** Qualified (Windows Fabric and Win32 Paper hosted smokes); the macOS consumer smokes were blocked by Accessibility and are not qualified. See [Generated manifest provenance](#generated-manifest-provenance). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441); landed in [#4490](https://github.com/microsoft/fluentui-react-native/pull/4490). + +- **MCP/WDIO handoff:** Qualified for the generic attach-only phase on macOS AppKit and Windows/Win32 fixtures. Storybook handoff is Deferred. See [MCP/WDIO handoff](#mcpwdio-handoff). Tracking: [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449); landed in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Storybook resource contributions:** In progress. The integration keeps its CLI lifecycle. See [Storybook resource contributions](#storybook-resource-contributions). Tracking: [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Cross-service and multi-target claims:** In progress. Exact macOS and Windows exclusion is qualified; broader lifecycle qualification is open. See [Cross-service and multi-target claims](#cross-service-and-multi-target-claims). + +- **Persisted run, status and artifact queries:** Done. No management transport. See [Run queries and management](#run-queries-and-management). Tracking: [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450). + +- **Workload management API:** Deferred. See [Run queries and management](#run-queries-and-management). Tracking: [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450). + +- **Automatic service reuse and per-user daemon:** Deferred. See [Service reuse](#service-reuse). + +- **Reliability and concurrency:** In progress. No repeated-run window is established. See [Reliability and concurrency](#reliability-and-concurrency). Tracking: [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451), [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349). + +## Configuration and standalone use + +### Behavior and limits + +- Storybook is optional; installed apps run without it. +- The runtime boundary and standalone-consumer contracts reject React, React Native and Storybook imports and dependencies. +- Executable configuration is trusted operator code, not a sandbox. +- Unknown configuration fields and invalid app or suite references are rejected. + +### Open work + +- None beyond the native qualification below. + +## Service and helper composition + +### Behavior and limits + +- The generic supervisor and the supervised Storybook child use the same real `driverService` factory. +- They share helper policy, authority and settled close. +- Generic and Storybook composition stays in `src/driverService.ts`. + +### Open work + +- None beyond the native qualification below. + +## WDIO execution and reporting + +### Behavior and limits + +- WDIO owns its session; hooks never create competing sessions. +- Storybook CSF callbacks use a runner-owned session borrowed by a Node worker. +- One collector produces bounded JSON and NDJSON summaries. +- Service-ready is not app-ready or context-ready. +- Skipped capabilities, empty work and incomplete work never count as passing coverage. +- Non-agent smoke is the legacy, unprotected path and stays separate from descriptor-path results. +- Normal WDIO cannot use the agent-served path except through the attach-only handoff. +- On Windows, report replacement retries only transient `EPERM`/`EACCES` rename conflicts (see [reporting](reporting.md)). The original conflicting handle owner was never identified, so it is not attributed to a reader or antivirus. + +### Open work + +- [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349): required native CI. + +## Agent descriptor publication + +### Behavior and limits + +- `serve --agent` publishes a private descriptor that protects the driver routes. +- MCP never builds or launches the application automatically. +- A real physical console Ctrl+C on the Windows packaged Fabric app and Win32 Paper REX drained the service, wrote a terminal `passed` lifecycle report, released the claim, withdrew the descriptor and preserved the borrowed app until its recorded launcher closed it ([#4491](https://github.com/microsoft/fluentui-react-native/pull/4491)). +- That qualification ran in an unlocked remote RDP session; a non-RDP console session has not been run. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): live requalification of descriptor-path lifecycle on all four endpoints. + +## Cleanup fence and lifecycle settlement + +### Behavior and limits + +- Failed or timed-out disposal retains the claim with the exact resource generation and unresolved effect. +- A fenced claim needs manual inspection; there is no automatic reclaim. +- A fenced claim refuses a replacement workload. +- Guarded work settles before a successful close. +- Late acquisitions are disposed. +- A required-service failure stops its dependants. +- Owned resources dispose in reverse order. + +### Open work + +- [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451): failure-path matrix, kernel PID recycling and power-loss rename; automatic reclaim stays unsupported. +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): live failed-cleanup retention and restart on each endpoint. + +## Resource contributions and durable recovery + +### Behavior and limits + +- Static contribution plans validate IDs and dependencies. +- They record readiness, generation, identity, ownership and settled disposal through the public `resources` and `ownership` subpaths. +- Both native service paths use owner-only storage and exact-incarnation probes. +- Claim schema 3 records the driver's OS start identity; schema 2 and unverified claims inspect as unknown. +- Restart is not recovery proof. +- Automatic reclaim is unsupported. + +### Open work + +- [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451): exact-resource reconciliation beyond the implemented ledger. +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): live endpoint qualification of restart recovery. + +## Supervisor crash containment + +### Behavior and limits + +- Cooperative Node service lifelines and correlated service readiness are implemented. +- Portable contracts cover parent and wrapper crash, late spawn, failed withdrawal and stalled children. +- macOS applications survive an owner crash as `uncontained`; the claim stays fenced and an operator performs exact cleanup. +- Windows: the helper's Job Object ends an unpackaged app it launched when the helper dies. +- Windows packaged launches are `uncontained`; they survive in a fenced claim. +- Windows hard-terminates non-detached children when the owner crashes, so the three cooperative crash-disposal contracts are skipped there. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): arbitrary descendants, Metro and channel survivors on macOS, and live uncontained packaged-survivor journaling. + +## Application ownership + +Opt-in supervisor launch with neutral exact attach, causal ownership and application claim resources. + +### Behavior and limits + +- Session-owned launch remains the default. +- External attach never transfers ownership; an attached application is preserved. +- Only applications whose exact incarnation was recorded are cleaned up. +- Failed cleanup keeps the pending claim fenced; a pending claim with no acquired identity adopts no guessed identity and refuses replacement. +- Native launch survivors and known descendants are `uncontained`; exact primary-process cleanup does not remove descendants. + +### macOS + +- macOS ownership requires Accessibility, PostEvent and ScreenCapture permissions for the signed helper. +- Intel macOS and remote control are unsupported. +- First-run TCC grants, clean-machine, PPPC and revoked-permission behavior are not inferred from passes. + +### Windows and Win32 + +- Windows application claims use schema 3 native owner identity, exact process pins, inspection and settlement. +- Packaged launches are `uncontained` and fenced. +- Win32 native application build is Unsupported; Win32 uses the prebuilt REX host. +- Stale-start recovery and matching-start identity mismatches fail closed. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): descendants, interrupted publication on Windows, supervisor-death survivors and live failed-cleanup fault injection. +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): Storybook lifecycle and lease relocation. + +## Generated manifest provenance + +### Behavior and limits + +- Generated driver manifests are strict version 3 and carry no secrets. +- Credentials use independent bridge and application attachment nonces. +- The owner-private runtime module holds only bridge data. +- The parent's handoff pins manifest, credential and runtime hashes and the resolved helper, and the channel child re-verifies the executable before spawning a helper. +- Metro receives the canonical schema locator from trusted process configuration, not from editable generated files. +- Windows and Win32 smoke lease publication goes through the trusted Node helper and OS private storage. +- Generated manifests never authorize native helpers or expose credentials. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): Storybook lease relocation. +- [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349): consistent helper build policy in required CI. +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): macOS manifest-path consumer smokes were blocked by Accessibility at the time; macOS requalification is tracked here. + +## MCP/WDIO handoff + +Generic attach-only phase. + +### Behavior and limits + +- A runner handoff fences admission before cancellation and drain, and confirms deletion before granting an attach-only reservation. +- Busy or failed handoffs issue no runner grant. +- The target reservation is held until every worker grant is drain-confirmed; a crash or revocation callback alone never releases it. +- Descriptor roots and workload or recovery holders refuse preempt with `target-busy` and require explicit relinquish. +- A preempted MCP child is discarded after authoritative rejection, and the connection reopens with fresh resources after runner drain. +- Live and draining grant capacity is reusable beyond 4096 connections; settled history never evicts pending or failed cleanup proof. +- Broker credentials are scrubbed before user hook import; WDIO receives each worker token through its per-worker `runnerEnv`, not launcher `process.env`. +- The worker's Node bootstrap and preload window stays trusted. +- Storybook handoff, retained launches, cross-connection resume and automatic replay are not part of this phase. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): Storybook handoff and preview lifecycle. +- [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349): the unchanged `boundedCommand` SIGKILL/SIGTERM assertion that blocks macOS portable aggregate acceptance. + +## Storybook resource contributions + +### Behavior and limits + +- The current integration substitutes its own CLI lifecycle. +- The channel and driver child stay in place until relocation is justified and qualified. +- Channel lifecycle placement and application lease ownership have not moved. +- Storybook targets are always attach-only and advertise only `attach`; the supervisor never launches or terminates a Storybook app. +- Protected story args are refused with `story-args-unsupported`, leaving the story unchanged and no value in an error or the journal. +- The non-agent smoke orchestrator intentionally uses the legacy transport. +- An ambiguous mutation is never retried. +- Windows `doctor --permissions` is Unsupported. +- Windows popup interaction does not establish a separately registered Storybook secondary-window identity. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): lifecycle takeover and lease relocation after the handoff and preview lifecycle work. +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): launch-capable discovery comparison, quarantine recovery, and live failed-cleanup retention per endpoint. +- [#4348](https://github.com/microsoft/fluentui-react-native/issues/4348): Windows and Win32 capture qualification. + +## Cross-service and multi-target claims + +### Behavior and limits + +- Configured aliases have distinct registered IDs while retaining physical identity. +- Exact application exclusion, lease release and the physical-input mutex are qualified on signed AppKit and on Windows Fabric and Win32 Paper. +- No physical-input suites run in parallel on one interactive desktop. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): broader lifecycle qualification. + +## Run queries and management + +### Behavior and limits + +- A harness-local read service and client expose atomic snapshots, terminal evidence, scoped expiring grants, bounded pages and allowlisted opaque artifact IDs. +- There is no management transport. +- Future management selects only declared workload IDs, never commands, paths or environment. +- Lifecycle and handoff gates still apply to any management API. + +### Open work + +- [#4450](https://github.com/microsoft/fluentui-react-native/issues/4450): management facet and transport. + +## Service reuse + +### Behavior and limits + +- Automatic service reuse and a per-user daemon are Deferred until exact identity and configuration matching and evidence of need exist. +- An unknown port occupant is never adopted. + +## Reliability and concurrency + +### Behavior and limits + +- Aggregate past successes are not a reliability pass. +- The required repeated-run window is not established. +- Interactive runs hold the per-user workload lock. + +### Open work + +- [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451): repeated-run reliability and support matrix. +- [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349): required native CI, including managed macOS physical-input lanes. +- [#4348](https://github.com/microsoft/fluentui-react-native/issues/4348): Windows and Win32 capture qualification. + +## Native qualification + +Results are per endpoint and tied to the revision that produced them; skipped checks are not passes. +Full per-run evidence (commands, counts, run ids, artifact locators, cleanup outcome) lives in the linked pull request or issue. +Hosted physical-input and capture skips do not qualify those paths. + +### macOS Fabric + +- **Application ownership:** 9 ownership/lifecycle checks passed, 0 failed, 0 cleanup failures; smoke 224 selected/executed/passed, 0 failed, 0 skipped, 0 retries (accessible input policy excluded one physical callback); rebased snapshot `16558610f`; evidence in [#4492](https://github.com/microsoft/fluentui-react-native/pull/4492). + +- **Generic handoff (plain AppKit fixture, no renderer):** `nativeRunner.contract` passed 5/5 with zero skips at the rebased revision `5dbadb53b`; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Physical-input smoke (latest):** `yarn storybook smoke --macos --mode stories-and-tests --input physical` passed 343, 0 failed, 3 declared unsupported at `b66140d7b`; the accessible lane passed 260 with 3 declared unsupported; the advisory hosted macOS physical job passed 343 with 3 declared unsupported; evidence in [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500). + +- **Earlier failure:** the smoke at `141b88b4b` failed 2 of 225 in `components-accordion--keyboard-disclosure` (obscured physical click, foreground activation failure); [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500) fixed it (scrolling previews, verified clear, verified foreground); original record in [#4493](https://github.com/microsoft/fluentui-react-native/pull/4493). + +- **Not run:** live failed-cleanup retention and restart, quarantine recovery and launch-mode discovery on macOS; tracked in [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +### macOS Paper + +- **Application ownership:** same result as Fabric (9 checks passed, smoke 224/224 passed, 0 skipped) at `16558610f`; evidence in [#4492](https://github.com/microsoft/fluentui-react-native/pull/4492). + +- **Physical-input smoke (latest):** passed 343, 0 failed, 3 declared unsupported at `b66140d7b` after a first attempt failed 7 cases (an unrelated overlay window obscured six; one hit-test unsupported); accessible 259 passed with 3 declared unsupported; no hosted Paper physical lane exists; evidence in [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500). + +- **Earlier failures:** at `141b88b4b` smoke failed 2 of 225 and the descriptor run 4 of 18 (keyboard clear, obscured Checkbox click), fixed by [#4500](https://github.com/microsoft/fluentui-react-native/pull/4500); Paper semantic activation remains Unsupported by the provider, and Paper Menu observations hit the 100-node cap; original record in [#4493](https://github.com/microsoft/fluentui-react-native/pull/4493). + +### Windows Fabric + +- **Application ownership (packaged, `furn-81` build `bdeb0b291`):** the complete native consumer suite passed 11/11 and the provider cases 6/6 at `da89fe9bc` (native source `dbaf55893`), including recovery of a packaged survivor and the physical-input mutex; the hosted selection ran 0 of the 11 consumer cases because it has no endpoint opt-in; evidence in [#4494](https://github.com/microsoft/fluentui-react-native/pull/4494). + +- **Generic handoff:** `nativeRunner.contract` passed all 5 outer cases with zero skips at `b1bbb26df` against the plain Win32 fixture (not a React Native renderer); handoff, operator preemption, worker crash, lost grant response and capture modes all passed; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Console interrupt of `serve --agent`:** a real physical console Ctrl+C (four `SendInput` key events, no injected signal) in an unlocked remote RDP session ended with exit codes 0/0/0, a terminal `passed` report, a released claim and a withdrawn descriptor at `b1bbb26df`; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Secondary windows and activation (unlocked remote RDP session, not console) at `6c626602c`:** a distinct secondary window's physical-pointer click was applied with `focus: confirmed` and its handle retired; modern main controls and the secondary refuse semantic activation before dispatch (`activation-unsupported`, effect `none`, no physical fallback); evidence in [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Connected-session probes at `505a2187b`:** physical keyboard clear, semantic clear, readonly refusal and minimized-capture refusal without restore passed; default observe is incomplete (`maxNodes`, 13 nodes omitted); 200% scale, mixed DPI and negative-origin layouts were not run; evidence in [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Storybook physical smokes:** Checkbox and Callout selections passed 203/203 with zero failures or skips at `fec94cec4`; evidence in [#4469](https://github.com/microsoft/fluentui-react-native/pull/4469). + +- **Manifest provenance (hosted run 43156299266):** accessible 255 selected, executed and passed; evidence in [#4490](https://github.com/microsoft/fluentui-react-native/pull/4490). + +### Win32 Paper + +- **Application ownership (prebuilt REX 0.81.1):** the same 11-case consumer suite and 6 provider cases passed at `da89fe9bc` with job-object drain reconciled; the hosted selection ran 0 of the 11 consumer cases; evidence in [#4494](https://github.com/microsoft/fluentui-react-native/pull/4494). + +- **Generic handoff:** same five modes passed at `b1bbb26df`; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Console interrupt of `serve --agent`:** same result as Windows Fabric (real physical console Ctrl+C in an unlocked remote RDP session; exit codes 0/0/0, terminal `passed` report, claim released) at `b1bbb26df`; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **Connected-session probes and secondary windows:** at `505a2187b` injected provider-context loss was fenced with `ref-region-changed` and `story-context-stale` (provider retirement, not crash or reconnect proof); at `6c626602c` the owned untitled tool secondary registered with exact-root UIA content and a physical-pointer click was applied (native delivery only, not proof of the React Native focus command); evidence in [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Storybook physical smokes:** Checkbox and Callout selections passed 177/177 with zero failures or skips at `fec94cec4`; evidence in [#4469](https://github.com/microsoft/fluentui-react-native/pull/4469). + +- **Manifest provenance (hosted run 43156299266):** accessible 233 selected, executed and passed; physical 289 selected, 288 executed and passed, 1 quarantined unsupported ([#4373](https://github.com/microsoft/fluentui-react-native/issues/4373), not a pass); evidence in [#4490](https://github.com/microsoft/fluentui-react-native/pull/4490). + +- **Open observation:** the exact REX process can outlive its closed window after MCP close and report failed cleanup; no later run resolves it; tracked in [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +### Hosted CI and portable evidence + +- **Windows aggregate:** a locked default-environment `yarn clean-lage-cache && yarn lage buildci` passed all 178 tasks at `2bf750f9e`; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491). + +- **macOS aggregate:** the final aggregate at the rebased F11 revision failed twice on the same bounded-command `SIGKILL`/`SIGTERM` assertion; no complete aggregate acceptance is claimed for that revision; evidence in [#4491](https://github.com/microsoft/fluentui-react-native/pull/4491); tracked in [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349). + +- **Hosted skips:** hosted physical-input and capture skips are not passes and do not qualify those paths. + +- **Historical consumer baselines** (`c84acb741`, endpoint-specific, not lifecycle evidence): FluentTester E2E passed 62/62 on macOS Fabric, 39/39 on Windows Fabric and 201/201 on Win32. + +## Maintaining this document + +- Edit only the topic your change affects, and keep that topic's line in the status overview in step with it. +- Describe what is true now. Do not add dated entries, run logs, or "updated" stamps. +- Record per-run evidence (commands, counts, run ids, artifact locations, cleanup outcome) in the pull request description or the tracking issue, and keep only the latest result for each endpoint here with a link to that record. +- Put remaining-work detail in the tracking issue and link the issue from the topic's open-work list; remove the link when the issue closes. +- Write one sentence or fact per line, and keep a blank line between adjacent list entries that different changes are likely to edit, so concurrent changes merge independently. +- Historical results stay tied to the revision that produced them and are never reported as qualifying a later revision. +- Skipped checks are not passes; report each endpoint (macOS Fabric, macOS Paper, Windows Fabric, Win32 Paper) separately. +- The detailed per-run ledger that preceded this format is preserved in git history: `git show a7af2b4b2:packages/agentic/desktop-test/references/current-status.md`. diff --git a/packages/agentic/desktop-test/references/native-qualification.md b/packages/agentic/desktop-test/references/native-qualification.md index c604f7ecdb..5caf5d8d8f 100644 --- a/packages/agentic/desktop-test/references/native-qualification.md +++ b/packages/agentic/desktop-test/references/native-qualification.md @@ -172,8 +172,10 @@ Evidence is written to the desktop-test package's `artifacts/desktop/windows-application-ownership-*/evidence.json`, including source/helper identities, endpoint descriptors, process evidence and case outcomes. Also record the registered app build, renderer, session kind, exact -commands, and before/after process audits in the Windows ledger. Keep private -machine paths in ignored artifacts, not public evidence. +commands, and before/after process audits in the pull request description or +tracking issue; the Windows endpoint sections of the package status documents +keep only the latest result with a link. Keep private machine paths in ignored +artifacts, not public evidence. The crash-owner helper is a contract fixture, not a second production service factory. Unknown resources/effects, unleased launches, legacy/unverified claims, @@ -512,5 +514,5 @@ that no `%LOCALAPPDATA%\Microsoft\FluentUIReactNative\desktop-test\workload\inte remain, that the channel, Metro, and WebDriver ports reported by the driver have no listeners, and that no app process from the run survives. Hosted physical-input or capture skips are skips, not passes. Keep descriptor-path results separate -from smoke/WDIO and deterministic fake-host results in all three package -status ledgers. +from smoke/WDIO and deterministic fake-host results in the run record and in +every package status document's per-endpoint results. diff --git a/packages/agentic/storybook-desktop/references/README.md b/packages/agentic/storybook-desktop/references/README.md index 4f33dab5f4..7eb0621361 100644 --- a/packages/agentic/storybook-desktop/references/README.md +++ b/packages/agentic/storybook-desktop/references/README.md @@ -18,4 +18,4 @@ owner_paths: | `storybook-desktop.win32` | [Win32](platforms/win32.md) | Bundling or running the REX host | | `storybook-desktop.troubleshooting` | [Troubleshooting](troubleshooting.md) | Diagnosing build, extraction, readiness, or smoke failures | | `storybook-desktop.contributing` | [Contributing](contributing.md) | Editing package code or tests | -| `storybook-desktop.current-status` | [Current evidence](current-status.md) | Checking endpoint evidence, bounded measurements, and qualification gaps | +| `storybook-desktop.current-status` | [Current status](current-status.md) | Checking per-endpoint qualification, bounded measurements, and open work | diff --git a/packages/agentic/storybook-desktop/references/current-status.md b/packages/agentic/storybook-desktop/references/current-status.md index 6e2f09e2c9..b0ecc897a0 100644 --- a/packages/agentic/storybook-desktop/references/current-status.md +++ b/packages/agentic/storybook-desktop/references/current-status.md @@ -5,527 +5,203 @@ owner_paths: - packages/agentic/storybook-desktop --- -# Native qualification status - -## Windows Fabric and Win32 Paper — #4441 connected-session qualification — 2026-11-12 UTC - -**Observed baseline:** `505a2187b0a2451aed7081c9260b343d55742f93`, -Windows 11 Enterprise build 26300 x64, Node 26.13.0. Packaged Windows -Fabric reused native build `bdeb0b291` after comparison found only -native-core metadata/version differences; **Inferred:** reuse was valid. -Win32 Paper used REX 0.81.1. Own-checkout Metro, manifests and protected -bundles served both endpoints. Native/GUI runs were serialized in the W3 -campaign with `-RequirePhysical`; before/after gates stayed eligible, -unlocked **remote RDP**, not console. One 1728 x 1084 origin-zero monitor -at 96 DPI / 100% was attached; settings were not changed. - -### Win32 fresh-preview fencing - -**Passed, Observed:** after a verified Checkbox selection, an injected -stimulus called the real orchestrator's bridge-retirement path while -deliberately preserving the socket and native preview. Old preview ref -`hkbo1` at epoch `(1,1)` rejected mutation with `ref-region-changed`. -A newly found, unique **native** checkbox ref `hkbo3` at epoch `(1,2)` -then rejected mutation with `story-context-stale`. Both were effect-none, -not quarantined; the checkbox remained unchecked. Pre-established chrome -ref `hkbo2` survived, and old/fresh chrome presence checks passed. - -This closes the earlier bounded fresh-preview-fence gap for **injected -provider-context loss**. It does not claim a native crash, actual socket -disconnect/reconnect, transparent upstream reconnection, chrome mutation, -or quarantine recovery. - -Exact local qualification invocation, inside the machine-wide -`with-native-lock.ps1 -Lane W3 -RequirePhysical` scope: - -```powershell -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs win32 --fence-only -``` - -Qualification ID `60751352-c158-40a2-b4f9-9251e15fec95`, driver run -`64d2de39-df69-4a2b-a164-fc842f0ef64e`, session -`08419564-7f43-48d2-9e08-b0f971590a95`. Artifacts relative to the tested -qualification checkout: -`artifacts\desktop\w3-qualification\runs\60751352-c158-40a2-b4f9-9251e15fec95\descriptor-cases-win32-fence.json` -and `qualification.json` in the same directory. The harness explicitly read -current-story state before injecting loss. These retained local scripts -are not a new published Storybook CLI or supported fault-injection API. - -### Secondary-window observations and open fixes - -| Case | Observed outcome | -| ------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -| Fabric separately registered modern Callout | Passed: exact distinct opaque handle, switch, complete 11-node observation, unique `modern-callout-focus`, actual physical-pointer click, `focus: confirmed`, then popup removal | -| Fabric secondary semantic activation | Failed: exact same control returned applied/accessibility, but main status stayed `Idle` over six checks / 2090 ms and final native observation | -| Paper separately registered Callout | Not run / partial: legacy popup shown, but only main registered; modern Callout absent from actual catalog | -| 200% | Not run: current scale 100%; changing it unauthorized | -| Mixed DPI | Not run: one attached monitor | -| Negative-origin layout | Not run: origin-zero monitor only; no approved eligible layout | -| Physical console Ctrl+C | Not run: separate implementation lane owns it | - -The Fabric status assertion read main through its explicit native root -without switching foreground; physical click passed once / 79 ms. -**Inferred:** accessibility-pattern/renderer dispatch may acknowledge an -activation without reaching Pressable `onPress`; the selected pattern was -not instrumented. Paper native inventory **observed** a visible untitled -root `Net UI Tool Window`, ToolWindow true. **Inferred:** the driver's -untitled `WS_EX_TOOLWINDOW` filter prevents its registration. These two -gaps are being fixed on this branch; this entry does not qualify the fixes. - -Exact bounded follow-ups, under the same physical-gated campaign wrapper: - -```powershell -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only -node .\artifacts\desktop\w3-qualification\run-endpoint.mjs windows --secondary-only --modern-only --physical-secondary -``` - -Semantic failure: qualification `675160f7-e069-413c-a6c8-486de0c37b69`, -driver `a9f2e71e-71ab-40f0-a375-fd6d76a0b971`; report -`artifacts\desktop\w3-qualification\runs\675160f7-e069-413c-a6c8-486de0c37b69\descriptor-cases-windows-secondary---modern.json`. -Physical pass: qualification `7fbe3b0b-df33-476b-80e3-95b97bf19f02`, -driver `16aac92b-bb36-4742-bd29-01df381d9594`; same artifact-root layout, -`descriptor-cases-windows-secondary---modern---physical.json`. -Paper inventory: qualification `22332065-6fa3-4061-a2b6-93f83295ca2e`, -`descriptor-cases-win32-secondary.json`. - -The supervisor ran supported driver/bundle/run stages in order, bundling -before launch and using actual readiness identity. Earlier setup, -unsupported-story, invalid-selector/fact and receipt-only failures remain -retained, not silently counted as passes. Session/child grants drained, -the exact owned apps/services stopped, and the W3 campaign ended. -Pre/post-end audits passed: -`FURN-AUDIT: no FURN desktop processes found`. -The complete native clear/capture/MCP matrix is in -`@fluentui-react-native/desktop-test/references/current-status.md`; -MCP response costs and default-observe completeness are in -`@fluentui-react-native/desktop-driver-mcp/references/current-status.md`. - -## 2026-11-12: refresh-reduction timing baseline - -Part of #4435. This section records bounded measurements, not a reliability -claim or a full-catalog test pass. The product baseline is the instrumented -parent of the navigation optimization, stacked on PR #4490 -(`aadc2f3591d15f5d43496d7f069fa8d0d32d5384`). - -### Method and coverage - -The public `DesktopStorybookCli` API ran eight independently launched -lifecycles per macOS renderer in the accessible lane. An injected fetch -narrowed only the CLI's traversal index to the real -`components-listboxitem--native-selection` entry; the server's authenticated -manifest, runtime, native helper, and authored callbacks were unchanged. -Every lifecycle traversed that story, ran its three supported callbacks in -declaration order, and reported its one declared macOS-unsupported callback. -App, channel, helper, and Metro cleanup completed before the next lifecycle. -This deliberately bounded index is **not** complete smoke traversal. - -The measurements below are from the executable-story `run.json`, excluding -application/service startup and traversal. Millisecond p50 is the median; p95 -uses inclusive linear interpolation. Phase samples are 24 executed callbacks; -runner-total samples are eight lifecycles. Navigation subphases overlap and -must not be added together. The first lifecycle can include cold caches. - -| Baseline phase | Fabric p50 ms | Fabric p95 ms | Paper p50 ms | Paper p95 ms | -| ---------------------------- | ------------: | ------------: | -----------: | -----------: | -| Session attach | 1192.22 | 1612.56 | 1094.93 | 1609.26 | -| Manifest | 2.62 | 3.37 | 2.60 | 3.14 | -| Input policy | 0.01 | 0.04 | 0.01 | 0.04 | -| Preparation | 0.13 | 0.17 | 0.13 | 0.24 | -| Upstream HTTP | 503.23 | 504.56 | 503.08 | 503.95 | -| Correlated runtime readiness | 28.77 | 562.62 | 25.86 | 458.94 | -| Native marker verification | 75.07 | 168.00 | 97.40 | 170.07 | -| Complete navigation | 753.13 | 812.97 | 756.77 | 792.62 | -| Worker startup | 203.00 | 215.85 | 196.50 | 207.40 | -| Worker body | 505.18 | 533.74 | 523.32 | 556.19 | -| Settled worker | 723.34 | 750.74 | 736.00 | 789.32 | -| Receipts | 2.62 | 2.93 | 2.46 | 2.76 | -| Session cleanup | 115.24 | 126.77 | 108.93 | 115.69 | -| Diagnostic native reads | 13.33 | 18.08 | 6.79 | 9.11 | - -Runner-total p50/p95 were **7939.07 / 8845.37 ms**; aggregate duration was -**63950.71 ms**. Sample variance was **306828.83 ms²** (range -7478.77–9245.48 ms). Complete lifecycle p50/p95, including startup and -traversal, were 19887.17 / 28433.88 ms, with 171994.83 ms aggregate and -19950017.41 ms² sample variance. These different totals are not interchangeable. - -Paper runner-total p50/p95 were **7915.99 / 8755.09 ms**, aggregate -**60909.13 ms**, sample variance **2199477.39 ms²**, and range -4123.59–8794.34 ms. Complete lifecycle p50/p95 were 19928.62 / 62342.92 ms, -aggregate 224024.71 ms, and variance 529506165.46 ms². The first Paper -lifecycle totaled 84945.44 ms including cold startup, but had an unusually -short runner -(4123.59 ms); both remain included. The first Fabric baseline also had -recording/cold-cache overhead. No startup or reliability improvement is -attributed to those differences. - -### Lifecycle and observation limits - -Each lifecycle had one verified PID/start-time pair and one authenticated bridge -connection. Fabric accepted one hello per baseline lifecycle; Paper accepted -two on the same connection. Each callback selected a fresh preview/context -generation (2, 3, 4 after the traversal's first generation). The historical -`mountGeneration` field merely copied that value; keyed remounting was -inferred from the rendering policy, not independently observed. That derived -field has been removed. No within-lifecycle application replacement or new bridge -connection was observed. JS initialization/HMR is explicitly **not observed**: -neither a hello nor the absence of a new socket proves a JS initialization count. -The keyed generation is not an instrumented React effect-mount counter. - -Native activity reports retained fixed counters only. No activation command, -provider activation, application activation, or main/focused-window change -was observed in these baseline cohorts; Paper recorded one application -deactivation. The original truncation projection combined a process-lifetime -flag with invalid sequence-gap inference; its per-case truncation counts are -withdrawn and cannot be reconstructed from these bounded reports. `AXUIElementDestroyed` -notifications can describe preview elements, not necessarily closed windows. -The bounded native event ring can be truncated, and its observations are -advisory. These counters do not establish whether a user saw a flash. - -Target-window-only ScreenCaptureKit recordings remain local in ignored -consumer artifacts. They are separate evidence, not committed media. -Paired candidate/renderer evidence and visual analysis must be recorded before -claiming a speedup or flash reduction. - -### Gates - -Session attachment is a material, dominant cost in this cohort; the conditional -session/worker investigation gate is eligible. No session reuse, worker -sharing, worker-isolation flag, preview reuse, or runtime protocol expansion is -implemented by the timing work. Such a trial still needs independently paired -benefit and ownership/settlement evidence. Windows Fabric and Win32 Paper: -**not run locally**. - -## 2026-11-12: direct-selection candidate - -Part of #4435. Candidate implementation: `06446c276`; instrumentation: -`f0feebc90`. The driver-owned channel factory now uses authenticated direct -selection without an upstream HTTP mutation. Controlled policy remains -independent and still refuses protected args and foreign mutation relaying. -Full extracted export maps are cached by source content/location within one -owning run, including across smoke's per-story calls. Review corrections also -bind cache hits to digests of the helper files read by input-policy validation. -Manifest digest checks, -remounts, fresh sessions, fresh workers, and native marker verification remain. - -### Fabric before/after - -Eight baseline and eight candidate lifecycles selected identical callbacks in -identical order: 24 passed and eight declared unsupported on each side. -These are sequential before/after blocks, not randomized alternating runs. -Phase samples are 24 callbacks; extraction samples and totals are eight runs. -Values below are p50 / p95 milliseconds. +# Current status + +This document records the status of Storybook desktop qualification: Windows Fabric and Win32 Paper connected-session results, direct story selection, refresh reduction, and lifecycle parity. +Done means the stated implementation exists and is covered by deterministic tests; it does not mean every native endpoint is qualified. +Parent issue: [#4435](https://github.com/microsoft/fluentui-react-native/issues/4435). +See the [integration guide](integration.md) and the [reference index](README.md). +Harness, driver and MCP evidence lives in `@fluentui-react-native/desktop-test/references/current-status.md`, `@fluentui-react-native/desktop-driver/references/current-status.md` and `@fluentui-react-native/desktop-driver-mcp/references/current-status.md`. + +## Status overview + +Direct story selection and run-owned extraction caching are implemented and measured on macOS only. +Fresh-preview fencing for injected provider-context loss is qualified on Win32 Paper at baseline `505a2187b`. +Windows Fabric modern Callout physical-pointer interaction passed; semantic-activation refusal and Win32 Paper tool-window registration landed in PR [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498) and passed local native checks at its final source `6c626602c`. +Console-session (non-RDP), 200% and mixed DPI, and negative-origin cases are not qualified. + +- **Preview fencing and context loss:** Qualified (Win32 Paper, injected provider-context loss only). Native crash, real socket reconnect and chrome mutation are not covered. See [Preview fencing and context loss](#preview-fencing-and-context-loss). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Secondary windows:** Qualified (Windows Fabric and Win32 Paper, remote RDP session, limited cases). Display-scale, negative-origin and console cases are not run. See [Secondary windows](#secondary-windows). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441). + +- **Story selection and refresh reduction:** Qualified (macOS Fabric and macOS Paper, bounded cohort). Windows Fabric and Win32 Paper are not run. See [Story selection and refresh reduction](#story-selection-and-refresh-reduction). Tracking: [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Lifecycle parity:** Qualified (macOS Fabric and macOS Paper, observed). JS initialization and HMR counts are not observed. See [Lifecycle parity](#lifecycle-parity). Tracking: [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Visual flash claims:** Deferred. No flash-reduction claim is made. See [Visual flash claims](#visual-flash-claims). Tracking: [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449). + +- **Native qualification:** Qualified (macOS Fabric and macOS Paper direct selection; Windows Fabric and Win32 Paper limited to the connected-session fencing and secondary-window cases). Windows refresh-reduction, console and display-scale lanes are not run. See [Native qualification](#native-qualification). Tracking: [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441), [#4451](https://github.com/microsoft/fluentui-react-native/issues/4451). + +## Preview fencing and context loss + +Scope: Win32 Paper stale-ref rejection after the Storybook preview context is lost. + +### Behavior and limits + +- After provider-context loss, a ref from the old preview is rejected with `ref-region-changed`. +- A freshly found native ref in the new preview epoch is rejected with `story-context-stale` until the story context is re-established. +- Both rejections are effect-none and do not quarantine the session. +- Chrome refs established before the loss survive. +- Qualification injected the loss through the orchestrator's bridge-retirement path while keeping the socket and native preview. +- The qualification harness reads current-story state before injecting loss; it is retained local tooling, not a published CLI or fault-injection API. +- Not claimed: native crash, actual socket disconnect and reconnect, transparent upstream reconnection, chrome mutation, or quarantine recovery. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): requalify on later revisions and the remaining lost-context variants. + +## Secondary windows + +Scope: separately registered popup windows such as the modern Callout. + +### Behavior and limits + +- Windows Fabric modern Callout registers as a distinct window with an opaque handle; a complete observation, a unique focus target and a physical-pointer click with `focus: confirmed` passed. +- Windows legacy-only semantic activation is refused before dispatch with `activation-unsupported`, because an accessibility default action does not prove the Pressable ran. +- Win32 Paper registers an untitled tool popup only after exact-process ownership, native identity, visible bounds and bounded content checks. +- The Paper legacy story publishes no native focus-command outcome, so its physical-pointer receipt proves native delivery, not the React Native focus command. +- Paper's modern Callout is absent from the Paper catalog, so Paper modern-Callout secondary coverage does not exist. +- Secondary windows are inspected without activation through the read-only window-tree route. +- Not run: 200% scaling, mixed DPI and negative-origin layouts. + +### Open work + +- [#4441](https://github.com/microsoft/fluentui-react-native/issues/4441): the unrun display-scale, negative-origin and console cases. + +## Story selection and refresh reduction + +Scope: the driver-owned channel factory and the run-owned extraction cache. + +### Behavior and limits + +- Agent sessions select stories directly through the authenticated channel without an upstream HTTP mutation. +- Controlled policy remains independent: protected story args are refused and foreign mutation relaying is refused. +- Extracted export maps are cached by source content and location within one owning run, including across smoke's per-story calls. +- Cache hits are bound to digests of every helper module read by input-policy validation, including transitive helpers and failed reads; a changed helper reruns extraction. +- Manifest digest checks, per-case remounts, fresh sessions, fresh workers and native marker verification remain on every callback. +- Phase durations include failed attempts; a missing duration does not mean success or zero cost. +- `processEventsTruncated` reports the native process-lifetime flag; `eventsTruncated` reports per-case history loss only when the history no longer covers the case cursor. +- Session reuse, worker sharing, preview reuse, new isolation descriptors and runtime protocol expansion are not implemented. +- Attach is the dominant remaining cost, so a session or worker investigation is eligible but needs its own paired benefit and ownership evidence. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): session and worker investigation and the Windows measurements. + +## Lifecycle parity + +### Behavior and limits + +- Each lifecycle has one verified app incarnation and one authenticated bridge connection; macOS Fabric accepts one hello and macOS Paper two on the same connection. +- Each callback selects a fresh preview and context generation. +- Keyed remounting is inferred from the rendering policy, not counted; the duplicate `mountGeneration` field no longer exists. +- JS initialization and HMR counts are not observed, and generations are not React effect-mount counters. +- Native activity reports are advisory fixed counters; AX element-destruction counts describe preview elements, not closed windows. +- Window-oriented test helpers can activate the app without physical input; this does not identify a painted flash or justify removing foreground guarantees. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): an independent JS-initialization counter, if needed. + +## Visual flash claims + +### Behavior and limits + +- Window-only ScreenCaptureKit recordings exist only as local artifacts. +- The near-uniform-luma detector is not a validated blank-preview or activation-flash classifier; chrome, variable-rate recordings and window-only pixels limit it. +- No speedup or flash-reduction claim is made without paired candidate and renderer evidence plus visual analysis. + +### Open work + +- [#4449](https://github.com/microsoft/fluentui-react-native/issues/4449): validated visual analysis, if a flash claim is wanted. + +## Native qualification + +Results are per endpoint and tied to the revision that produced them; skipped checks are not passes. +Full per-run evidence (commands, counts, run ids, artifact locators, cleanup outcome) lives in the linked pull requests and issues. + +### macOS Fabric + +- **Direct selection, accessible lane:** full-index run rendered 223 stories and passed five callbacks (228 passes, no failures or unsupported results) at candidate `06446c276`. Evidence: [#4496](https://github.com/microsoft/fluentui-react-native/pull/4496). + +- **Direct selection, physical lane:** 223 stories plus 15 callbacks passed (238 passes) at `06446c276`; sparse regression evidence, not the full physical suite. Evidence: [#4496](https://github.com/microsoft/fluentui-react-native/pull/4496). + +- **Review corrections:** not requalified natively; earlier qualification is historical for `06446c276`. + +### macOS Paper + +- **Direct selection, accessible lane:** 223 stories plus five callbacks passed (228 passes) at `06446c276`. Evidence: [#4496](https://github.com/microsoft/fluentui-react-native/pull/4496). + +- **Direct selection, physical lane:** 223 plus 15 callbacks passed (238 passes) at `06446c276`. + +- **Review-correction smoke:** full-index accessible smoke with Button, Dropdown and modern Callout selection passed 228 checks; ten physical callbacks were excluded by input policy. + +### Windows Fabric + +- **Secondary window:** at `6c626602c` the modern Callout physical-pointer click passed (`focus: confirmed`) and semantic activation was refused before dispatch with effect none and no fallback. Evidence: [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Environment:** Windows 11 Enterprise build 26300 x64, remote RDP session (not console), one 1728 x 1084 monitor at 96 DPI. + +- **Not run:** 200%, mixed DPI and negative-origin layouts. Refresh-reduction measurements not run. + +### Win32 Paper + +- **Fresh-preview fencing:** passed for injected provider-context loss at baseline `505a2187b` (REX 0.81.1). Evidence: [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Secondary tool window:** the owned untitled tool popup registered, a complete eight-node sample was read, and a physical-pointer click passed at `6c626602c` (native delivery only). Evidence: [#4498](https://github.com/microsoft/fluentui-react-native/pull/4498). + +- **Not run:** refresh-reduction measurements. + +### Hosted CI and portable evidence + +- Portable Storybook integration, driver and root checks passed for the refresh-reduction work; JavaScript tests are not native qualification. +- The cache-cleared `yarn lage buildci` aggregate failed on timing-sensitive driver tests (`boundedCommand`, `privateStorage`) on the candidate and on base `aadc2f359`, so it is not hosted-CI parity; tracked in [#4349](https://github.com/microsoft/fluentui-react-native/issues/4349). + +## Measurements + +Revision: candidate `06446c276` against its instrumented parent; stacked on PR [#4490](https://github.com/microsoft/fluentui-react-native/pull/4490) (`aadc2f359`). +Platform: macOS, accessible lane, Fabric and Paper separately. +Method: eight independently launched lifecycles per renderer and side, run through the public `DesktopStorybookCli` API with the traversal index narrowed to `components-listboxitem--native-selection`; three callbacks passed and one declared macOS-unsupported callback was reported per lifecycle. +Samples: phase samples are 24 callbacks; extraction samples and totals are eight lifecycles; values are p50 / p95 milliseconds (inclusive linear interpolation). +Caveats: rows omitted for unchanged phases and lifecycle totals (which include startup and traversal); sequential before/after blocks (not randomized), small cohort, no system-load control, first lifecycle may include cold caches, navigation subphases overlap and must not be added, no reliability or startup claim. +Evidence: [#4496](https://github.com/microsoft/fluentui-react-native/pull/4496). + +### macOS Fabric | Phase | Baseline | Candidate | | ---------------------------- | ----------------- | ----------------- | | Session attach | 1192.22 / 1612.56 | 1088.67 / 1611.77 | -| Manifest | 2.62 / 3.37 | 2.90 / 3.88 | -| Input policy | 0.01 / 0.04 | 0.01 / 0.02 | -| Preparation | 0.13 / 0.17 | 0.15 / 0.20 | | Upstream HTTP | 503.23 / 504.56 | Not dispatched | | Correlated runtime readiness | 28.77 / 562.62 | 31.78 / 538.64 | | Native marker verification | 75.07 / 168.00 | 72.32 / 546.11 | | Complete navigation | 753.13 / 812.97 | 274.20 / 788.37 | -| Worker startup | 203.00 / 215.85 | 201.50 / 212.85 | -| Worker body | 505.18 / 533.74 | 510.90 / 533.80 | -| Settled worker | 723.34 / 750.74 | 729.28 / 758.36 | -| Worker settlement | 18.12 / 20.96 | 19.63 / 21.90 | -| Receipts | 2.62 / 2.93 | 2.67 / 2.98 | -| Session cleanup | 115.24 / 126.77 | 118.25 / 126.46 | -| Diagnostic native reads | 13.33 / 18.08 | 7.81 / 14.71 | -| Extraction | 6.50 / 10.34 | 6.50 / 8.67 | - -| Fabric total | Baseline | Candidate | -| -------------------------- | ------------------- | ------------------- | -| Runner p50 / p95 ms | 7939.07 / 8845.37 | 6680.04 / 6995.31 | -| Runner aggregate ms | 63950.71 | 53173.56 | -| Runner sample variance ms² | 306828.83 | 91557.04 | -| Runner range ms | 7478.77–9245.48 | 6192.64–7031.09 | -| Lifecycle p50 / p95 ms | 19887.17 / 28433.88 | 18461.08 / 28733.43 | -| Lifecycle aggregate ms | 171994.83 | 162636.05 | -| Lifecycle variance ms² | 19950017.41 | 30801394.39 | - -The bounded Fabric runner median was 15.9% lower. Matching repetitions by -index gave a median candidate-minus-baseline delta of −1141.99 ms; a -10,000-resample paired-median bootstrap interval was -−1887.49 to −782.75 ms (95%, fixed seed). This does not account for -temporal/system-load bias or establish reliability. Navigation's median -improved while native-verification p95 worsened; no universal tail-latency -or startup claim follows. This single-source cohort parsed once per run on -both sides and does not measure the cross-story cache benefit. - -### Paper before/after - -The same eight-plus-eight design, callback selection/order, and accessible -lane passed 24 callbacks and eight declared unsupported on each side. -Values are p50 / p95 milliseconds, with 24 callback samples or eight -extraction/total samples. +| Runner total | 7939.07 / 8845.37 | 6680.04 / 6995.31 | + +Runner median was 15.9% lower; paired-by-index median delta −1141.99 ms (10,000-resample bootstrap 95% interval −1887.49 to −782.75 ms, fixed seed). +Navigation median improved while native-verification p95 worsened. + +### macOS Paper | Phase | Baseline | Candidate | | ---------------------------- | ----------------- | ----------------- | | Session attach | 1094.93 / 1609.26 | 1434.73 / 1621.30 | -| Manifest | 2.60 / 3.14 | 2.69 / 3.11 | -| Input policy | 0.01 / 0.04 | 0.01 / 0.01 | -| Preparation | 0.13 / 0.24 | 0.14 / 0.20 | | Upstream HTTP | 503.08 / 503.95 | Not dispatched | | Correlated runtime readiness | 25.86 / 458.94 | 26.28 / 456.76 | | Native marker verification | 97.40 / 170.07 | 167.26 / 582.90 | | Complete navigation | 756.77 / 792.62 | 340.59 / 787.83 | -| Worker startup | 196.50 / 207.40 | 193.00 / 201.85 | -| Worker body | 523.32 / 556.19 | 529.44 / 547.30 | -| Settled worker | 736.00 / 789.32 | 741.24 / 756.44 | -| Worker settlement | 15.94 / 18.23 | 16.09 / 17.49 | -| Receipts | 2.46 / 2.76 | 2.44 / 2.62 | -| Session cleanup | 108.93 / 115.69 | 106.57 / 114.07 | -| Diagnostic native reads | 6.79 / 9.11 | 7.09 / 12.84 | -| Extraction | 6.44 / 8.97 | 6.64 / 9.77 | - -| Paper total | Baseline | Candidate | -| -------------------------- | ------------------- | ------------------- | -| Runner p50 / p95 ms | 7915.99 / 8755.09 | 7218.78 / 7811.59 | -| Runner aggregate ms | 60909.13 | 57610.90 | -| Runner sample variance ms² | 2199477.39 | 210172.97 | -| Runner range ms | 4123.59–8794.34 | 6440.61–7996.85 | -| Lifecycle p50 / p95 ms | 19928.62 / 62342.92 | 18921.30 / 60994.39 | -| Lifecycle aggregate ms | 224024.71 | 215170.59 | -| Lifecycle variance ms² | 529506165.46 | 519947445.91 | - -The bounded Paper runner median was 8.8% lower. Paired-by-index median -candidate-minus-baseline delta was −760.17 ms; its 10,000-resample bootstrap -95% interval was −1120.33 to −192.13 ms. One pair regressed by 3245.12 ms; -no samples were dropped. Cold first-lifecycle totals, including startup, were -retained on both sides (baseline 84945.44 ms, candidate 83318.31 ms). -Session-attach and native -verification medians worsened despite faster median navigation. The same -non-randomization, small-cohort, system-load, reliability, and cache-benefit -limits apply as for Fabric. - -### Candidate lifecycle parity - -Across each renderer's 24 passing candidate callbacks, preview/context -generations remained 2/3/4 per lifecycle. Fabric matched its native marker on -the first poll in all 24 callbacks, as did both renderer baselines. Paper's -candidate needed one poll in 12 callbacks and two polls in 12, contributing -to its slower native verification; fresh verification was never bypassed. -All eight lifecycles per side had distinct -verified app incarnations, one authenticated socket per lifecycle, and no -within-lifecycle app replacements or extra accepted hellos during callbacks. -Fabric accepted one hello and Paper two on the same connection. No independent -JS-init/HMR counter was added, and keyed generations do not count React -effect mounts. - -Candidate activity retained zero activation commands, provider activations, -application activation/deactivation, main/focused-window changes, or window -creation in the bounded observational cohort. Generic AX element-destruction -counts matched baseline (Fabric 96; Paper 240), not closed-window counts. -The earlier 16/17-of-24 truncation comparison was invalid: sticky -process-history flags and globally allocated sequence gaps do not establish -per-case loss. That comparison is withdrawn. These are advisory observations, not -proof that no activation or visual flash occurred. - -### Window-only recording limits - -Separate synchronized recording lifecycles used only the target Storybook -window, outside the performance cohorts. All recordings were 1864×1564: - -| Renderer / revision | Duration seconds | Encoded frames | Last encoded timestamp seconds | Uniform-content candidates | -| ------------------- | ---------------: | -------------: | -----------------------------: | -------------------------: | -| Fabric baseline | 8.037 | 4 | 0.102 | 0 | -| Fabric candidate | 7.273 | 4 | 0.103 | 0 | -| Paper baseline | 8.590 | 4 | 0.103 | 0 | -| Paper candidate | 7.862 | 4 | 0.100 | 0 | - -The candidate detector measured near-uniform luma in a padded window-content -crop; it is **not a validated blank-preview or activation-flash classifier**. -The crop includes persistent Storybook chrome, variable-rate recordings hold -their last frame, and window-only pixels cannot establish foreground/window -raising behavior. There is **no measured flash-reduction claim**. Raw media -and derived image samples remain local, ignored artifacts. - -Phase 2 therefore ships measurement and a source-traced compatible proposal, -not an activation change. Phase 3's investigation gate is eligible because -attach is material, but pooling/startup changes remain deferred without their -own benefit and ownership/settlement evidence. Phase 4, new isolation -descriptors, preview reuse, and runtime protocol expansion remain excluded. - -### Focused native qualification and cache ownership - -The Fabric full-index accessible run rendered all 223 real catalog stories -and passed five selected callbacks (228 canonical passes, no failures or -unsupported results; run `57cf8594-addf-4cd3-9e60-5e1907a46b09`). -Selected stories were Button default, Dropdown default, and modern Callout -default/local-point/local-rectangle. This is not a full authored-test-suite -pass. The selected callbacks produced three source extractions and two -same-source cache hits across smoke's separate per-story runner invocations. -Each callback still had its own session, worker, and fresh preview/context -generation (224–228); the application incarnation and authenticated -connection stayed stable. - -Fabric physical qualification also rendered all 223 stories and passed -15 selected callbacks (238 canonical passes, no failures or unsupported -results; run `6ba319db-cada-4e66-a000-513015b2610b`). Selection covered -Button default, Dropdown default, modern Callout default, and native -Callout default/window-commands. Four source extractions and one cache hit -were recorded. Dropdown's open-list keyboard case preceded its -initially-collapsed semantic case; fresh per-case remounting prevented the -open popup from contaminating the latter. Native and modern Callout -dismiss/reopen, outside-click, anchor input, and Escape cases passed. -This is sparse physical regression evidence, not the full physical suite -or a qualification of concurrent physical-input fixes on another branch. - -Paper passed the same full-index selections: 223 rendered stories plus five -accessible callbacks (228 canonical passes; run -`c6e2032d-76ee-4021-bc96-3e1c73be60ff`) and 223 plus 15 physical callbacks -(238 passes; run `9351f32d-29fe-4e6e-ae3d-cadd85a51989`). Neither lane had -failures, skips, unsupported callbacks, retries, or incomplete work. -Source extraction/cache-hit counts matched Fabric (3/2 accessible, 4/1 -physical). Fresh callback generations were 224–228 and 224–238 respectively; -each lane retained one verified app incarnation and authenticated socket, -with two accepted hellos. - -Fabric's three accessible Callout callbacks each observed two activation -commands -and two provider activations during their worker interval, plus focused-window -changes. This shows that window-oriented test helpers can activate even -without physical input; it does not identify a painted flash or justify -removing foreground guarantees. The timing cohort itself observed no -activation commands on either side. - -### Package validation - -| Workspace / command | Result | -| ---------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------- | -| Storybook integration: `yarn format && yarn lint && yarn build && yarn test --runInBand` | **Passed**; 405 tests passed, four skipped, 28 suites | -| Driver: `yarn format && yarn lint && yarn build` | **Passed** | -| Driver portable: `FURN_NATIVE_DRIVER_TEST=0 yarn test --runInBand` | **Passed**; 1223 tests passed, 66 skipped, 73 suites passed and eight skipped | -| Root: `yarn oxfmt --check` | **Passed** | -| Root: `yarn check-agent-instructions` | **Passed** | -| Root: `yarn change:check` after committing patch changesets | **Passed** | - -New regressions cover direct dispatch without HTTP, normal args versus -protected args refusal, correlated/stale/foreign readiness, uncertainty after -dispatch cancellation or disconnect, native-marker verification, bounded -diagnostic projection, source-content/location invalidation, run ownership, -cross-invocation cache hits, and manifest-digest rejection even on a cache hit. - -### Initial native and environment-overridden aggregate validation - -All native/GUI and aggregate commands were serialized through the shared -native lock, with project-local temporary files. - -From `apps/storybook`, the strict standalone-helper build **passed**: - -```sh -swift build --package-path ../../packages/agentic/desktop-driver/native/macos --scratch-path artifacts/refresh-reduction/native-strict-build -c release -Xswiftc -warnings-as-errors -artifacts/refresh-reduction/native-strict-build/release/furn-desktop-driver-host --self-test -``` - -The self-test **passed 448/448 checks**. From the driver workspace, -`FURN_NATIVE_DRIVER_TEST=1 yarn test nativeDriver nativeCache applicationIncarnation serviceIdentity --runInBand` -**passed 87 tests**, with 16 skipped, six suites passed and two Windows suites -skipped. No product Swift or React-module boundary changed; Office -module/bridging-header qualification is not applicable. - -The initial cache-cleared `yarn clean-lage-cache && yarn lage buildci` attempt -**failed** in scripts' negative instruction fixtures: a temporary directory -inside ignored artifacts inherited the outer Git repository. The scripts -workspace's `yarn test` then **passed all 76 tests** with Git discovery -bounded by the temporary directory's parent. The next environment-overridden aggregate -**failed** two timing-sensitive driver tests in parallel execution -(`boundedCommand` termination signal and `privateStorage` timeout). -`FURN_NATIVE_DRIVER_TEST=0 yarn test boundedCommand privateStorage --runInBand` -then **passed 73 tests**, with ten skipped. - -An earlier local aggregate **passed** with `TF_BUILD=1` serializing each -package's Jest run and with the fixture environment below, without another -cache clear. This was **not PR CI mode**: `.github/workflows/pr.yml` runs -parallel `yarn lage buildci` without these overrides. - -```sh -export TMPDIR="$PWD/apps/storybook/artifacts/refresh-reduction/tmp" -export GIT_CEILING_DIRECTORIES="$PWD/apps/storybook/artifacts/refresh-reduction" -TF_BUILD=1 yarn lage buildci -``` - -Lage reported 46 successful and 132 skipped tasks, zero pending/aborted, exit -zero. These task skips are not test-pass counts. No source changes were made -to the instruction validator, private storage, or bounded process runner. - -## 2026-11-12: refresh-reduction review corrections - -Part of #4435. This follow-up corrects the run-owned extraction cache and -diagnostic interpretation without changing dispatch, per-case workers, -context verification, activation, or native protocol semantics. - -- Cache entries now retain content digests for every helper module read by - direct-input validation, including transitive helpers and failed reads. - Hits recheck those digests; a changed helper reruns extraction and policy - validation even when the story source is unchanged. Regressions change a - direct or transitive helper to keyboard input after a confirmed cache hit. -- `processEventsTruncated` reports the native process-lifetime flag. - `eventsTruncated` reports possible per-case history loss only when a - truncated history no longer covers the case cursor. Globally allocated - sequence gaps alone do not establish loss. The unsupported - `commandsTruncated` inference was removed. Tests cover sticky lifetime - flags, global gaps, and lost cursor coverage. Historical 16/17-of-24 - truncation comparisons above are withdrawn, not relabeled: the original - bounded projections cannot reconstruct actual per-case loss. -- The duplicate `mountGeneration` field was removed. Context/preview - generations are observed; keyed remounts are inferred from the default - rendering policy, not independently counted. A runner contract requires - the derived field to be absent. -- Present phase durations include failed attempts because measurement runs - in `finally`. Missing durations do not establish success or zero cost. - A runner regression verifies attempted navigation time on failure and - absence of worker time when the worker was never started. - -### Review qualification - -The regression-first run failed as expected before these fixes. From the -Storybook integration workspace, -`yarn test extractWdioTests wdioExtractionCache runDiagnostics wdioRunner --runInBand` -then **passed 112 tests in four suites**, including a repeat with the normal -inherited environment. Package format/lint/build passed. - -From `apps/storybook`, `yarn storybook manifest --macos --paper` and -`yarn storybook bundle --macos --paper` **passed**. A full-index CLI-API smoke -with the app-owned configuration, Paper renderer, accessible input, and the -Button/Dropdown/modern-Callout selection **passed 228 checks**: 223 rendered -stories and five callbacks; ten physical callbacks were excluded by input -policy. The five callback reports showed three source extractions, two -cross-invocation cache hits, no synthetic mount field, and separate process -and case history flags. The first local harness invocation failed before -selection because plain `node` did not provision Yarn's tool PATH -(`spawn rnx-cli ENOENT`); rerunning through `yarn node` passed without a -product change. - -The plain aggregate commands below ran under the shared native lock with -the normal inherited environment: no `TF_BUILD`, `TMPDIR`, or -`GIT_CEILING_DIRECTORIES` overrides. Unlike the earlier locally serialized -run, they preserve parallel Jest execution. Local Node was 24.18.0; the PR -workflow declares Node 22, so this is not complete hosted-CI parity. - -| Scope | Exact command | Result | -| -------------------------- | -------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ | -| Candidate, first plain run | `yarn clean-lage-cache && yarn lage buildci` | **Failed**: `boundedCommand` expected `SIGKILL`, received `SIGTERM`; 122 Lage tasks succeeded, one failed | -| Detached base `aadc2f359` | `yarn clean-lage-cache && yarn lage buildci` | **Failed**: the same `boundedCommand` assertion plus `privateStorage`'s 10-second test timeout; 129 Lage tasks succeeded, one failed | - -A final plain repeat of the same cache-cleared candidate command also -**failed** the `boundedCommand` assertion (164 Lage tasks succeeded, one -failed). `privateStorage` passed in both candidate runs. In each candidate -driver suite, 1222 tests passed, one failed, and 66 were skipped. -`yarn oxfmt --check`, `yarn check-agent-instructions`, and -`yarn markdown-link-check -c .github/markdown-link-check-config.json ` -for every changed Markdown file **passed**. - -The detached base was installed independently after its first command -reported missing dependencies; it did not reuse candidate workspace -modules or outputs. This reproduces the aggregate failures before the -refresh-reduction commits. Neither unrelated test nor implementation was -changed. The base worktree was removed, and its orphaned Jest worker was -terminated by its verified PID. - -The paired performance and recording evidence above remains evidence for -`06446c276`, not a fresh benchmark of this correction. No new speedup or -flash-reduction claim is made. macOS Fabric and physical-input lanes were -**not rerun in this follow-up**; their earlier qualifications remain -historical. Windows Fabric and Win32 Paper were **not run locally**. +| Runner total | 7915.99 / 8755.09 | 7218.78 / 7811.59 | + +Runner median was 8.8% lower; paired-by-index median delta −760.17 ms (95% interval −1120.33 to −192.13 ms); one pair regressed by 3245.12 ms and no samples were dropped. +Session-attach and native-verification medians worsened; the candidate needed one marker poll in 12 callbacks and two in 12, and fresh verification was never bypassed. +The baseline's first lifecycle was unusually short (runner 4123.59 ms) and is retained. + +## Maintaining this document + +- Edit only the topic your change affects, and keep that topic's line in the status overview in step with it. +- Describe what is true now. Do not add dated entries, run logs, or "updated" stamps. +- Record per-run evidence (commands, counts, run ids, artifact locations, cleanup outcome) in the pull request description or the tracking issue, and keep only the latest result for each endpoint here with a link to that record. +- Put remaining-work detail in the tracking issue and link the issue from the topic's open-work list; remove the link when the issue closes. +- Write one sentence or fact per line, and keep a blank line between adjacent list entries that different changes are likely to edit, so concurrent changes merge independently. +- Historical results stay tied to the revision that produced them and are never reported as qualifying a later revision. +- Skipped checks are not passes; report each endpoint (macOS Fabric, macOS Paper, Windows Fabric, Win32 Paper) separately. +- The detailed per-run ledger that preceded this format is preserved in git history: `git show a7af2b4b2:packages/agentic/storybook-desktop/references/current-status.md`. diff --git a/references/agents/repository/package-guidance.md b/references/agents/repository/package-guidance.md index d19a499e7d..fcaf2ea072 100644 --- a/references/agents/repository/package-guidance.md +++ b/references/agents/repository/package-guidance.md @@ -22,6 +22,7 @@ add a package to the registry when it adopts this standard. - **R-PKGDOC-007:** Changes to behavior, commands, configuration, outputs, or contracts update the owning reference in the same change, and update `AGENTS.md` or `README.md` when their entry-point information changes. - **R-PKGDOC-008:** Published packages include their `AGENTS.md`, `README.md`, reference index, every reference document, and local linked assets. Validate the actual Yarn pack inventory; build caches and generated artifacts must not ship. - **R-PKGDOC-009:** Governed packages have no `PLAN.md` or other plan files outside `references/`. Unfinished or explicitly deferred work lives in an indexed `/references/remaining-work.md`. When work completes, move durable behavior into the owning reference and delete the item rather than marking it done; delete the file when no work remains. +- **R-PKGDOC-010:** A package's `/references/current-status.md` is a status overview followed by stable topic sections, not a dated log. It describes what is true now, keeps only the latest result for each native endpoint with a link to the pull request or issue that holds the full evidence, and links open work to its tracking issue instead of restating it. Per-run commands, counts, and artifact locators belong in the pull request or issue. Edit only the topic a change affects so concurrent changes merge cleanly. ## Package entry points