Skip to content

Harden desktop-driver server recovery and agent operations - #4479

Merged
Jason Morse (JasonVMo) merged 15 commits into
mainfrom
user/jasonvmo/desktop-driver-server-hardening
Nov 10, 2026
Merged

Jason Morse (JasonVMo) merged 15 commits into
mainfrom
user/jasonvmo/desktop-driver-server-hardening

Conversation

@JasonVMo

@JasonVMo Jason Morse (JasonVMo) commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor

Platforms Impacted

  • iOS
  • macOS
  • win32 (Office)
  • windows
  • android

Description of changes

Related to #4435.

This PR carries two groups of fixes in @fluentui-react-native/desktop-driver, merged because the agent operations need the server to report what a session can do, and a Windows CI follow-up:

  1. Server, protocol and recovery: SRV-1 to SRV-7, SRV-9 (bug part), SRV-10, SRV-11 and SRV-13.
  2. Agent operations: OPS-2, 3, 4, 5, 6, 9, 10, 11, 12, 15, 16 and 17.
  3. Integration and a Windows contract fix that only exist in the combination (see the last section of the description).

Each group was validated alone before the merge, and those results are kept below, separate from the validation of the combined head.

Server, protocol and recovery

Changes are in src/server/*, src/protocol/*, src/host/types.ts, the matching tests, and references/{security,service,protocol,current-status,remaining-work}.md.

  • Named accessibility actions skipped the preview fence. POST /element/{id}/furn/accessibility-action did not enforce the Storybook preview-context check that click, clear and value enforce, so a stale preview element could be acted on. All four routes now share one "preview mutation" predicate.
  • Story select and reset from a popup. The story root marker was looked for only in the session's current window, so from a popup the call polled to the render deadline and then quarantined the target. The root is now searched in every session window.
  • Closing a session with the journal down. Input release ran through the journaled mutation path, so a journal outage left keys held, still closed the application and left the session draining. The release now runs as an unrecorded recovery mutation and raises an in-memory quarantine when its outcome is unknown. The application is left open when the release never ran.
  • Unknown furn: capabilities. A typo such as furn:inputmode silently became compatibility, and a non-string furn:target fell back to implicit selection. Request-side furn: keys are allow-listed with their types (clickMode, endpoint, inputMode, launchMode, renderer, target), anything else is invalid argument, and furn:launchMode is returned.
  • Browser rejection. Only Origin was checked. A request is now refused when Host is not 127.0.0.1, [::1] or localhost with the connection's own port (DNS rebinding), or when it carries Sec-Fetch-Site, Sec-Fetch-Dest, Sec-Fetch-User or a Sec-Fetch-Mode other than cors. Node's fetch always sends Sec-Fetch-Mode: cors, so blanket rejection of Sec-Fetch-* broke every client; a browser always sends the others. The listener closes and startup fails if the bound address is not loopback, and a permanent error handler keeps a later accept error from ending the process.
  • Settlement journal recovery. Storage that created a record and then failed, or a remove that returns false for an absent record, wedged the target. The journal now reads back and adopts an exact match, treats absent as removed, and bounds every storage and probe call with a 10 second deadline.
  • Timeouts. Values were assigned one by one before validation, 1e300 was accepted, and script: null was rejected. All values are validated first, must be safe integers of at least zero, script may be null, and an implicit-wait search stops when the session is deleted, so DELETE no longer waits out the deadline.
  • Routing bugs. Malformed percent-encoding is invalid argument, an unknown /furn/* route is unknown command, and POST /window/new is unsupported operation, instead of HTTP 500.
  • Startup options failed open. content: 'Deny' behaved as allow and maxBodyBytes: NaN meant unlimited. Options are now validated strictly (unknown keys, content, port, maxBodyBytes, host, recovery). A target registered after the server started reads its journal on registration, and /status fails instead of reporting a target whose records could not be read.
  • Content policy scope (documented). The W3C routes (element text, attributes, properties, source, screenshots, furn/tree) ignore operations.content. It governs furn/ops only, and grants have no scope that separates the two, so this is documented in service.md and the option's JSDoc rather than changed.
  • Closing a launched application lifted its quarantine too early. The quarantine was cleared as soon as closeApplication resolved, but the host contract does not promise the process exited. It is now cleared only when the probe reports exited; host/types.ts documents the contract. Without a probe the quarantine stays until the operator calls recoverTarget.
  • Stale docs. security.md said neither shipped supervisor supplies recovery storage, and current-status.md said WDIO had no named-action invocation. Both are corrected.

Behavior changes to know about (server group): a launched application's quarantine is no longer lifted on close when no probeApplication is configured, and unknown option keys, unknown furn: capabilities and invalid timeouts that were previously ignored are now rejected.

Agent operations

Observation paging was quadratic, sequence could refuse halfway, act and check misreported and mis-retried, delta observation issued and evicted references, caller and provider text could reach the journal and errors, session capacities were clamped, and complete meant different things on different operations. This is the platform-neutral Node operation layer shared by both desktop endpoints; the native helpers are untouched.

By file:

  • src/operations/projection.ts, observe.ts: pages fill incrementally and sort once, and canonical delta items are cached per snapshot. Output is byte-identical for the same input.
  • src/operations/observe.ts: delta mode skips fresh projection and reads removed elements with refs.peek. A delta element's ref is optional. The truncation of the canonical 5,000-node budget is carried into the delta.
  • src/operations/act.ts: interactability names the entry's reference and reports an unsupported enabled or visible as state-unsupported. Provider failures are mapped to fixed text per error code. A click hit-tests, activates, then hit-tests again.
  • src/operations/check.ts, limits.ts: unsupported is terminal, and evaluations are at least checkPollMs.minimum (100 ms) apart or as long as the last evaluation took.
  • src/operations/sequence.ts, selector.ts, runtime.ts: preflight covers selector content predicates in act and check steps, and a strict-mode clear.
  • src/operations/envelope.ts, refs.ts: a ref must match the reference grammar, so only well-formed references reach the journal. A malformed ref is ref-malformed without its text. An unknown key is Unrecognized key. without its name.
  • src/operations/limits.ts, refs.ts, journal.ts, proofs.ts, outcomes.ts: requireCapacity rejects a capacity that is not an integer in range. Proof and outcome capacities gain ceilings.
  • src/operations/find.ts, story.ts, journalOperation.ts, shapes.ts: complete means nothing reachable was left out. find with cut matches, journal, and stories report it and a truncation reason (limit or dropped). Clipped story fields are listed in clipped.
  • src/testing/operationRuntime.ts, src/operations/*.test.ts: the stub runtime accepts host overrides; new act, sequence and check suites and cases in observe, projection, find, journal, story, envelope, executor, refs and sessionState.
  • references/operations.md and the README /agent row: add the missing reasons and a furn:inputMode description, and document each fix.

Fixes:

  • OPS-2: every page candidate rebuilt its ancestors, recounted omitted children over all nodes and re-serialized the page. A 5,000-node delta page took 7.7 s. It now takes about 10 ms. A guard test counts 4,801 nodes against a 1,000 ms bound.
  • OPS-3: sequence preflight ignored selector content predicates and strict-mode clear, so a refusal landed after earlier steps had acted. It now checks every step first. See the integration section for the server side.
  • OPS-4: act by selector said Reference "undefined" is not enabled., and an unsupported enabled or visible counted as false. It now names the reference and reports state-unsupported.
  • OPS-5: check retried unsupported-fact until the wait expired, and event bursts ran back-to-back tree walks. It now ends as unsupported after one attempt, and walks are spaced.
  • OPS-6: observeDelta returned 41 references for 3 elements with zero changes, and minted references for removed elements, which could evict references the agent holds. It now issues none for unchanged or removed elements.
  • OPS-9: any 1 to 32 character ref was journaled, and malformed refs and unknown keys were echoed in validation messages. Both are now fixed text.
  • OPS-10: dispatch failures bypassed interpretFailure, so provider text such as window titles and element ids passed through. They are now fixed words per code, keeping only reason, nativeCode and the rebuilt mutation report.
  • OPS-11: click reported rejected with no effect after it had already raised the window, and a hit-test timeout became obscured. A cover in the element's own window is refused before activation. After activation, a failed or covered hit test is partial and indeterminate. Timeouts and cancellation propagate.
  • OPS-12: the missing reasons (observation-cursor-expired, story-context-stale, input-policy, input-unsupported, input-mechanism-mismatch, bridge-unavailable, relaunch-required, state-unsupported) and the furn:inputMode description are documented. The stale catalog.ts comment in shapes.ts now says registry.ts.
  • OPS-15: capacities were clamped, ceilings unchecked, and NaN disabled trimming and eviction. They are now validated as integers in range and rejected otherwise.
  • OPS-16: the README called /agent and agent describe bounded. DesktopAgent.describe reads the whole tree before trimming, so the README now says so.
  • OPS-17: stories and journal had no complete or truncation, and find reported complete: true after limit cut matches. All three now report them, with one meaning (see the integration section for stories).

Integration of the two groups, and a Windows contract fix

These changes exist only because the groups are combined.

  • clear is always reported by the server (finishes OPS-3). The operation group added OperationCapabilities.clear as optional, and the real server never set it, so sequence preflight could not see a strict clear the target cannot perform, and an omitted capability read as permissive. clear is now required. src/server/sessionOperations.ts derives it from strictClearMechanism, the same rule that guards the host call, so it is true for a compatibility session and, for a strict session, only when the mode allows the clear mechanism and the target advertises strictClear and that mechanism. A refusal there is false. A new sessionInput.test.ts case drives a real sequence through the server: an accessible or physical session on a target without strictClear (or with it omitted) refuses a later clear step before the earlier activate or type step touches the host, with no host actions and no receipts. The same sequence runs when the target can clear, and a compatibility session is still left to the host. A strict act clear on such a target is now action-unsupported before dispatch.
  • stories always reports complete and truncation (finishes OPS-17). They are true and [] when nothing after offset was left out, and false with a limit entry otherwise, as find and journal do. The server expectation in storyOperations.test.ts and operations.md are updated.
  • references/architecture.md and references/contributing.md no longer call src/agent "Bounded", matching the README.
  • Windows CI, protectedService.contract.test.ts:233. CI for the server-only head 26f68768e failed in Windows Storybook job 129833452854 (run 42817279330): quarantines the target for every grant when the provider cannot settle, until the application is closed found the new session still quarantined after the close. This is SRV-11 working as designed, not a recovery regression. A launched application's quarantine is now lifted only when probeApplication reports the exact incarnation exited, and the contract served the target with no recovery, so nothing could prove the exit. The contract now serves that case with createWindowsApplicationProbe over the helper's --application-incarnation mode and Windows private storage on a scratch root, as a Windows supervisor does. A second case keeps the fail-closed expectation: with no probe, closing the application never ends the quarantine. Recovery is not weakened.

Not in this PR

  • SRV-8: hit-ownership logic is unchanged beyond OPS-11.
  • OPS-1, OPS-7, OPS-8: CLI usage errors, owned by the cli fixes. OPS-13, OPS-14: refactors, deferred.
  • SRV-12, the route-table refactor; only the SRV-9 bug part is fixed.
  • MAC-2, the macOS helper side of the close-and-exit proof, handled by the macOS change. This PR relies only on the probe verdict.
  • Sealing the target registry after listen; late registration is handled instead, because existing tests and integrations register targets after listen.
  • Making the content policy cover the W3C routes (documented, see above).
  • A DELETE does not cancel in-flight native commands; only client-sized waits end.

Verification

Validation is reported in three separate parts: the server group alone, the operations group alone, and the combined head. Only the combined head is the code this PR contains.

Server group, standalone (on its own branch, before the merge)

Run in packages/agentic/desktop-driver unless noted, on macOS. No GUI or native runs were needed.

Command Result
yarn format passed
yarn lint passed
yarn build passed
yarn test passed, 60 suites, 944 tests (8 suites skipped for platform or opt-in)
yarn test in packages/agentic/desktop-test passed, 15 suites, 285 tests (1 skipped)
yarn test in packages/agentic/storybook-desktop passed, 22 suites, 333 tests (4 skipped)
yarn oxfmt --check on the changeset, the edited references, src/server and src/protocol (root) passed
yarn markdown-link-check -c .github/markdown-link-check-config.json packages/agentic/desktop-driver/references/<file>.md for the five edited references (root) passed
yarn clean-lage-cache && yarn lage buildci --concurrency 6 (root) passed, 178 tasks succeeded, 0 failed

New tests: serverBoundary.test.ts (Host and Fetch Metadata rejection, malformed paths, loopback bind assertion, listener error survival, strict startup options), and cases in desktopDriver.test.ts (invalid capabilities, furn:launchMode, atomic and safe-integer timeouts, script: null, DELETE ending an implicit wait, unknown /furn/*, /window/new), storyOperations.test.ts (named-action fence, popup story root, unknown furn/story/* with an orchestrator), recovery.test.ts (adopt, absent-as-removed, storage deadline, unrecorded release, launched-quarantine proof, late registration) and timeouts.test.ts (withDeadline). 35 of the new cases in serverBoundary, desktopDriver and recovery fail against the previous createDesktopDriverServer.ts.

Not run: native helper suites (FURN_NATIVE_DRIVER_TEST) and any Windows or GUI qualification; this change does not touch native code.

Operations group, standalone (on its own branch, before the merge)

Run in the operations worktree on macOS, with no GUI or native runs.

Command Result
DD focused yarn test <name> runs while iterating passed
Regression check: each fix stashed, then its new tests rerun (OPS-3, 4, 5, 10, 11) the new tests failed without the fix
DD yarn format passed (206 files)
DD yarn lint failed once (consistent-type-imports in two new tests), fixed, then passed
DD yarn build passed
DD yarn test passed: 62 suites (8 skipped), 913 tests passed, 64 skipped, 0 failed
SBD yarn test (consumes agent operations) passed: 22/22 suites, 333 tests passed, 4 skipped
root yarn oxfmt --check passed (3122 files)
yarn markdown-link-check -c .github/markdown-link-check-config.json on operations.md and README.md passed (16 links)
yarn clean-lage-cache then yarn lage buildci --concurrency 6 passed on the fifth run, see below

buildci ran five times in that worktree because the machine was loaded by parallel agent builds (load average above 80). The three clean-cache runs that failed did so in packages this group does not touch, and each failing test passed when rerun alone: desktop-test contracts.test.ts; storybook-desktop wdioRunner.test.ts (7 s process deadline); and desktop-test supervisor.test.ts and contracts.test.ts on the machine-wide furn-desktop-workload interactive.lock. The two reruns that kept the cache passed with exit 0.

Combined head (this PR), run in furn-fix-srv on macOS

Command Result
yarn test sessionInput story sequence act in packages/agentic/desktop-driver (focused) passed; the new strict-clear cases fail (3 of 3) when the server reports clear: true
DD yarn format, yarn lint passed
DD yarn build passed after adding the now-required clear to two capability fixtures in registry.test.ts and a satisfies type in a new test
DD yarn test passed: 63 suites (8 skipped), 986 tests passed, 65 skipped, 0 failed
yarn test in packages/agentic/storybook-desktop passed: 22 suites, 333 tests passed, 4 skipped
yarn test in packages/agentic/desktop-test passed: 15 suites, 285 tests passed, 1 skipped
yarn test in packages/agentic/desktop-driver-mcp passed: 2 suites, 26 tests
root yarn oxfmt --check on packages/agentic/desktop-driver and .changeset passed
yarn markdown-link-check -c .github/markdown-link-check-config.json on operations.md, architecture.md, contributing.md passed (4 links)
yarn clean-lage-cache, then native-lock.sh srv-ops-buildci yarn lage buildci --concurrency 4 (root, one fresh-cache run) passed, exit 0: 178 tasks succeeded, 0 skipped, 0 failed, 1 m 8.8 s, first try, with the lock acquired after 0 s

Windows: the native contract suites run only on Windows with FURN_NATIVE_DRIVER_TEST=1, so protectedService.contract.test.ts was not run on this macOS host. The change to it was checked with tsc -b, lint and format only. The server-only head 26f68768e failed that suite in Windows Storybook job 129833452854 (run 42817279330), as described above, and the fix was not run on Windows locally. On the combined head 684ff5bf4 the Windows Storybook job and the other native and platform CI jobs passed, which is the evidence that the fix works on Windows. If it regresses, the quarantine will have stayed because the probe did not report the closed application exited, and the next step is to look at that verdict, not to weaken recovery.

Linux CI, desktop-test contracts.test.ts › recovery contract. The JS PR job at 684ff5bf4 (job 129845427368, run 42820692741) failed once in that contract with AggregateError: Desktop supervisor cleanup failed. from await seed.close() at contracts.contract.cjs:1585; the other 14 suites passed. The nested cleanup errors were not captured: the contract's main().catch writes only error.stack, which omits error.errors. The cause is therefore not identified, and the failure did not reproduce here. What is established:

  • I found no changed cleanup path in this PR, which does not prove the PR is not the cause. The recovery contract drives only WebDriver sessions, never agent operations, and the files changed between the server-only head 26f68768e (JS PR job passed) and 684ff5bf4 are src/operations/*, server/sessionOperations.ts, testing/operationRuntime.ts, docs and the Windows contract test. desktop-test has no change in this branch.
  • On macOS, Node 24 and Node 22 (CI's version), the recovery contract passed in about 600 direct runs under CPU saturation (load average above 130), random main-thread stalls of up to 200 ms, and GITHUB_ACTIONS and CI set, and in all 18 whole-file or whole-package jest runs under load. Under the same load other desktop-test tests failed on timing, such as lane-cli, reporter-cli and the 5 s service-readiness deadline in owned process lifelines, so that package does flake when the machine is saturated. That is consistent with a load-dependent failure, but it does not show that this one was.
  • This PR did not weaken any server deadline or exit-proof rule, and none is proposed to be relaxed for this failure.

The single failed-job retry (gh run rerun 42820692741 --repo microsoft/fluentui-react-native --failed) passed on unchanged head 684ff5bf4. All PR checks, including JS, Windows Storybook, macOS Paper/Fabric Storybook, and Win32 accessible/physical Storybook, are now successful. This does not identify the original cleanup failure's cause. Nested AggregateError diagnostics are being added in the separate desktop-test lifecycle/reporting PR #4478; any recurrence should be investigated from those causes rather than by relaxing deadlines or exit proof.

No macOS native, GUI or Storybook run was needed for this change.

Before After
A foreign Host, a typo in a furn: capability or maxBodyBytes: NaN was accepted; DELETE could leave keys held. Each is refused or handled with a typed error, and DELETE releases input and closes the session even when the journal is down.
On an accessible or physical session whose target cannot clear with one mechanism, a sequence of activate then clear ran the activate, then the clear failed (the server never reported clear). The whole sequence is refused action-unsupported before the activate; the host sees no action and no receipt is recorded.
stories omitted complete and truncation unless stories were cut. complete: true and truncation: [] when the page reaches the end; complete: false and a limit entry otherwise.
Canonical delta projection of 5,000 nodes, view=all: 7,733 ms; observe then an empty delta issued 3, then 41 references. 10 ms with the same 485,002 bytes; 3, then 3 references.
sequence with denied content in step 2 clicked step 1, then stopped. A ref of a man, a plan, a canal was journaled as typed, and an unknown key was quoted in the error. The request throws content-denied before any click. The ref is ref-malformed and nothing is journaled, and the key is Unrecognized key.
Windows contract: new session still quarantined after session.delete(), because nothing proved the application exited. The case serves the real incarnation probe and private storage, and a second case asserts the quarantine stays without a probe. Passed in Windows CI on 684ff5bf4; not run locally.

Pull request checklist

This PR has considered (when applicable):

  • Automated Tests
  • Documentation and examples
  • Keyboard Accessibility
  • Voiceover
  • Internationalization and Right-to-left Layouts

Jason Morse (JasonVMo) and others added 12 commits November 10, 2026 12:16
Page filling rebuilt each candidate's ancestors, recounted every omitted child and
re-serialized the whole page, so a 5,000 node sample took seconds per delta page.
Pages now grow by the candidate's missing ancestors, measure only the lines whose
size changed, and sort once. The canonical projection a delta compares is cached per
immutable snapshot.

A delta no longer projects a page nobody reads: it issued references it never
returned, and a removed element minted a reference that could evict one the agent
holds. A removed element now carries the reference the session already holds, or none.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
A reference of 1-32 characters was accepted whatever it contained and then
recorded in the journal summaries, and malformed references and unknown keys
were echoed back in validation messages. Both are caller-typed text.

- referenceSchema now requires the reference grammar, so only a well-formed
  reference reaches a summary. A malformed one keeps the ref-malformed reason.
- A malformed reference is reported without quoting it.
- An unknown key is reported as "Unrecognized key." at its path.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Journal, reference, outcome and proof capacities were clamped to a floor and
floored to an integer, so a ceiling was never checked and NaN made the
trimming comparison false forever, leaving a table unbounded. They now share
one validator that accepts only an integer inside the budget and throws
otherwise, like every other budget.

The outcome and proof capacities become budgets with a floor and a ceiling.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
A named accessibility action on a preview element skipped the Storybook
context fence that click, clear and value enforce. One predicate now names
the mutating element commands. Story verification searched only the
session's current window, so a popup waited out the render deadline and
quarantined the target; it now searches every window.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…nched quarantine

Input release runs without a record when the journal cannot publish, and a
release that never ran keeps the application open. The journal reads back
after a storage failure, treats an absent record as removed, and bounds
every storage and probe call. A launched application's quarantine is lifted
only when the probe proves the exact incarnation exited.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
A sequence now checks every step's selector for content predicates and every
clear against the session's strict-clear capability before the first step runs,
so a refusal no longer lands halfway. The optional `clear` capability is false
when the session's input mode demands a mechanism the target cannot perform.

A check ends as unsupported as soon as the search reports an unsupported fact
instead of retrying until the wait expires. A host event shortens a pause only
after a minimum interval, at least as long as the last evaluation, so event
bursts no longer trigger back-to-back walks of the application.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
`act` names the reference the selector proved instead of "undefined", and an
`enabled` or `visible` fact the provider cannot read refuses with the distinct
`state-unsupported` reason rather than reading as false.

A dispatch failure no longer passes provider text (window titles, element ids)
through: provider failures become fixed text per code, keeping only the machine
reason, native code and the settled outcome. A reference or window the provider
no longer has is dropped, as it is for a read.

A physical click hit-tests first. A point covered inside the element's own
window is rejected before the window is raised. Once the window was raised, a
failed or still-covered hit test is reported as a partial effect, and hit-test
timeouts and cancellations propagate instead of reading as `obscured`.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
`find` no longer reports `complete: true` when `limit` cut matches: complete now
means nothing reachable was left out, and the cut is listed in `truncation` as a
`limit`. `journal` reports `complete` and `truncation` for entries a page left
out or the journal no longer holds. `stories` reports them, with the matching
`next`, only when stories after the offset were left out, and marks the fields
of an entry that were cut short.

A delta now says which budget left its basis incomplete when the provider's own
sample was complete, instead of suppressing removals without a reason.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…d routing

Reject unknown or mistyped furn: capabilities and return furn:launchMode.
Require a loopback Host with the connection's port, refuse browser Fetch
Metadata, assert a loopback bind and keep a permanent listener error handler.
Validate startup options strictly and read the journal of a target registered
after startup before it is reported or opened. Validate timeouts as a whole
with safe integers, accept a null script timeout and end implicit waits when
the session is deleted. Answer malformed percent-encoding with invalid
argument, unknown furn routes with unknown command and window/new with
unsupported operation.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Describe the request boundary, strict startup options, timeouts, recovery
behavior and the content policy scope, and correct the stale statements about
the shipped supervisors' recovery storage and WDIO named actions.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add the failure reasons operations.md lacked (observation-cursor-expired,
story-context-stale, input-policy, input-unsupported, input-mechanism-mismatch,
bridge-unavailable, relaunch-required, state-unsupported), a furn:inputMode
description, and the behavior of the fixed projection, preflight, check,
redaction and complete/truncation semantics. Correct the README row that called
/agent bounded, and add the changeset.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Satisfy the consistent-type-imports lint rule in the new act and sequence tests.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@github-actions

github-actions Bot commented Nov 10, 2026 •

Copy link
Copy Markdown
Contributor

Bundle size report

Tree-shaken, minified production esbuild bundles with React and React Native runtimes externalized.

Scenario Modules-Mac (Δ) Modules-Win (Δ) Size-Mac (Δ) Size-Win (Δ)
components-button 79 (+1) 80 (+1) 72.78k (-842b) 72.71k (-842b)
components-catalog 228 (+1) 229 (+1) 228.46k (-855b) 228.38k (-855b)
design-themed-root 55 (+4) 56 (+4) 56.26k (+998b) 56.18k (+998b)
design-color-lib 6 (+0) 6 (+0) 6.27k (+0b) 6.27k (+0b)
design-theme-state-legacy-source 49 (+1) 50 (+1) 52.96k (-841b) 52.89k (-841b)
design-legacy-theme 42 (+0) 42 (+0) 39.11k (+0b) 42.91k (+0b)
design-legacy-theme-from-flex 58 (+4) 58 (+4) 60.70k (+824b) 64.51k (+822b)
design-fluent-light-theme 7 (+0) 7 (+0) 18.74k (+0b) 22.59k (+0b)
design-fluent-theme 35 (+0) 35 (+0) 27.46k (+4b) 31.25k (-1b)
design-fluent-hc-theme 32 (+0) 32 (+0) 27.34k (+0b) 31.13k (+0b)
design-flex-theme 29 (-4) 29 (-4) 14.30k (-9.07k) 14.29k (-9.07k)
design-flex-cool-theme 26 (New) 26 (New) 12.51k (New) 12.50k (New)
design-flex-warm-theme 26 (-7) 26 (-7) 12.28k (-11.15k) 12.27k (-11.15k)

The job is advisory: size changes are reported but do not fail the pull request. Bundle or analysis errors still fail.


View bundle-size workflow run

@JasonVMo
Jason Morse (JasonVMo) enabled auto-merge (squash) November 10, 2026 12:42
Keep the operation capability contract and the server implementation in one PR so strict-clear preflight and result-completeness fixes ship together.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…ication probe

Lifting a launched application's quarantine now requires the probe to report the
exact application incarnation exited, because closing the application is no
proof. The contract served the target without recovery, so the new session stayed
quarantined after the close and the case failed on Windows CI. The case that
closes the application now serves with the helper's incarnation probe and
Windows private storage, and a second case keeps the fail-closed expectation:
with no probe, closing the application never ends the quarantine.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…d stories

The service now sets the required clear capability from the same strict-mode
rule that guards the host call, so sequence preflight refuses a clear the
session's input mode or target cannot perform before any earlier step acts,
instead of staying permissive when the capability was omitted. stories always
reports complete and truncation, as find and journal do. The architecture and
contributing references no longer call src/agent bounded.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
@JasonVMo Jason Morse (JasonVMo) changed the title Harden desktop-driver server fences, capabilities and recovery Harden desktop-driver server recovery and agent operations Nov 10, 2026
@JasonVMo
Jason Morse (JasonVMo) merged commit 6404535 into main Nov 10, 2026
33 of 35 checks passed
@JasonVMo
Jason Morse (JasonVMo) deleted the user/jasonvmo/desktop-driver-server-hardening branch November 10, 2026 23:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants