Initial work on an AGENTS.md file and associated skills []

Author: wchilders-nvidia

.agents/README.md

@@ -0,0 +1,15 @@
+# Portable agent skills
+
+Project skills live here in the open
+[Agent Skills](https://agentskills.io/) format. Each subdirectory contains a
+`SKILL.md` with YAML frontmatter (`name`, `description`) and instructions.
+
+```
+.agents/skills/<skill-name>/SKILL.md
+```
+
+Do not add parallel copies under tool-specific paths (`.cursor/skills/`,
+`.claude/skills/`, etc.); keep a single tree under `.agents/skills/`.
+
+Always-on project guidance for all agents: [`AGENTS.md`](../AGENTS.md) at the
+repository root. Changelog encoding: skill `edg-changes-example`.

.agents/skills/acknowledg-bench/SKILL.md

@@ -0,0 +1,88 @@
+---
+name: acknowledg-bench
+description: >-
+  Reads Cachegrind .cgout Ir totals, compares benchmark baselines to runs,
+  and supports AcknowlEDG bench review / update-baseline flows. Use when
+  editing edg-bench-run-delta, edgacknowledg.client.bench, .cgout summary
+  parsing, or benchmark UI ops fields.
+---
+
+# AcknowlEDG / Cachegrind benchmarks
+
+## Metric
+
+**Ir** (instruction reads) is the first `summary:` integer in a `.cgout`.
+CLIs and the UI call it **OPs**. Lower run Ir vs baseline = faster.
+
+```text
+events: Ir I1mr ILmr Dr ...
+...
+summary: 3696685 6148 4719 ...
+```
+
+With cache-sim enabled there are multiple counts; **always use the first**
+(Ir). Do not sum fields or take a later column.
+
+Shared regex (keep both call sites in sync):
+
+```python
+_REGEX_CGOUT_SUMMARY = re.compile(r'^summary: ([0-9]+)(?: [0-9]+)*\n$')
+```
+
+## Readers
+
+| Location | Function |
+| --- | --- |
+| [`dev_tools/bin/edg-bench-run-delta`](../../../dev_tools/bin/edg-bench-run-delta) | `read_program_totals` |
+| [`dev_tools/pylibs/edgacknowledg/client/bench/__init__.py`](../../../dev_tools/pylibs/edgacknowledg/client/bench/__init__.py) | same |
+
+Readers scan the last ~10 lines for `summary:`. Missing/malformed → skip that
+benchmark. Do not invent a third independent parser.
+
+Percent change:
+
+```text
+((baseline - run) / max(1, baseline)) * 100
+```
+
+## Layout
+
+Bench home: `$EDG_BENCH_HOME` or `<repo>/benchmarks`.
+
+| Path | Role |
+| --- | --- |
+| `benchmarks/benchmarks/**/*.bnch.cpp` | Sources |
+| `benchmarks/runs/<YYYY.MM.DD-HH.MM.SS>/…/*.bnch.cgout` | Runs |
+| `benchmarks/baselines/project/<tag>/` | Project baselines |
+| `benchmarks/baselines/user/<name>/` | User baselines (`@name` tag) |
+| `benchmarks/.acknowledg-cache/` | `cg_annotate` disk cache |
+
+Relative `.cgout` paths must exist under both run and baseline dirs to compare.
+
+## Review / UI principles
+
+Wire payloads should carry raw `baseline_ops` / `run_ops` (and baseline tag
+when needed). Derive in the client:
+
+- op delta, percent change, faster/slower badges
+- mean / median / population stdev over the list
+- sort order
+
+Do **not** ship redundant counts, samples arrays, or precomputed change
+comments when the UI already has the ops list.
+
+Bench annotate detail uses `cg_annotate --no-annotate --diff` (delta output,
+not a line-oriented source diff).
+
+## Update baseline
+
+Bench editor copies the selected run `.cgout` into the session’s baseline tag
+directory and refreshes statuses (`BenchEditHandler` /
+`edg-bench-review`).
+
+## Protocol
+
+Tags `bench.src.*` / `bench.edit.*`: see
+[PROTOCOL.md](../../../dev_tools/services/acknowledg/PROTOCOL.md).
+Client architecture: skill **acknowledg-client**.
+Python under `dev_tools/`: skill **edg-python-style**.

.agents/skills/acknowledg-client/SKILL.md

@@ -0,0 +1,114 @@
+---
+name: acknowledg-client
+description: >-
+  Works on the AcknowlEDG Python client: BatchClient vs DirectClient,
+  ReviewClient Protocol, resilient websocket reconnect, and test/bench
+  provider handlers. Use when editing edgacknowledg.client, edgy-review,
+  edg-bench-review, edg-acknowledg-cli, or section subscribe/reconnect logic.
+---
+
+# AcknowlEDG client
+
+Canonical package: [`dev_tools/pylibs/edgacknowledg/`](../../../dev_tools/pylibs/edgacknowledg/).
+`dev_tools/services/acknowledg/server/edgacknowledg` is a **symlink** — edit
+pylibs only.
+
+Wire protocol (tags, routes, auth, occupancy):
+[PROTOCOL.md](../../../dev_tools/services/acknowledg/PROTOCOL.md). Keep
+`PROTOCOL_REVISION` in sync with
+`dev_tools/services/acknowledg/client/src/sockets.js`.
+
+## Public API
+
+From `edgacknowledg.client` (`__all__`):
+
+- `BatchClient` / `DirectClient` — async context managers; constructor takes
+  endpoint settings and optional `status_listener`
+- `BatchClientSession` / `DirectClientSession` — yielded by client `__aenter__`;
+  `create_review` / `remove_review`, `close()`, `stopped()`, `start()`, `stop()`
+  (client `__aexit__` awaits `stopped()` then `stop()`)
+- `ConnectionStatusListener` — `connected` / `attempting_reconnect` /
+  `disconnected`
+- `ReviewClient` (structural **Protocol**): `review_id`, `name`, `timestamp`,
+  `sections()`, `add_section` / `remove_section`
+- URL/env helpers: `add_endpoint_arguments`, `form_section_http_url`,
+  `default_*` host/port helpers
+- Shared: `BackgroundWorkerPool`, errors
+
+Usage:
+
+```python
+async with BatchClient(
+  api_host=..., api_port=..., status_listener=listener
+) as session:
+  review = await session.create_review(...)
+  # Block ends when session.stop() is called, or on cancellation.
+```
+
+Private (leading `_`): `_ClientReview`, `_BatchClientReview`,
+`_DirectClientReview`, channel helpers. Callers should type against
+`ReviewClient`, not private classes.
+
+## Batch vs Direct
+
+| | Batch | Direct |
+| --- | --- | --- |
+| Socket | One resilient `/batch` | Per-section `/source|edit/<uuid>/<kind>/<slug>` |
+| Mux | `batch.subscribe` / `unsubscribe` / `event` | Dedicated WS per section |
+| CLI | `edg-acknowledg-cli` | `edgy-review`, `edg-bench-review` |
+| Reconnect | Drop local subs, `_restore_subscriptions()` | Each section's `ResilientWebsocket` loop |
+
+Direct URL mapping uses handler `subscription_type` (e.g. `test.src` →
+`/source/.../test/<slug>`, `bench.edit` → `/edit/.../bench/<slug>`).
+
+## Reconnect
+
+[`client/websocket.py`](../../../dev_tools/pylibs/edgacknowledg/client/websocket.py):
+`ResilientWebsocket`, exponential backoff + jitter on connect failure;
+jitter-only delay after a drop. Pattern:
+
+```python
+async for websocket in ResilientWebsocket(...):
+  try:
+    ...
+  except ConnectionClosed:
+    continue
+```
+
+Auth header: `X-AcknowlEDG-API-Key` (never put the key in query strings).
+`RoleOccupied`: discard that handler / stop reconnecting that section; keep
+other sections running.
+
+## Batch subscription maps
+
+`_BatchClientReview` tracks:
+
+- `_subscriptions: Dict[int, uuid.UUID]` — `id(handler)` → server subscription id
+- `_subscription_tasks: Dict[uuid.UUID, asyncio.Task]` — subscription id → task
+
+Unsubscribe needs the server id; the handler-key map provides it after
+reconnect bookkeeping.
+
+## Providers
+
+| Package | Handlers |
+| --- | --- |
+| `client/test/__init__.py` | `TestSourceHandler` (`test.src`), `TestEditHandler` (`test.edit`) |
+| `client/bench/__init__.py` | `BenchSourceHandler` (`bench.src`), `BenchEditHandler` (`bench.edit`) |
+
+Shared: `ReviewClientHandler`, `Channel`, `announce_section`,
+`trunc_display_content` in `client/shared.py`.
+
+## Env / URLs
+
+- `EDG_ACKNOWLEDG_API_{HOST,PORT}`, `EDG_ACKNOWLEDG_UI_{HOST,PORT}`,
+  `EDG_ACKNOWLEDG_API_KEY`
+- Defaults: API `localhost:2727`, UI `localhost:5173`; non-localhost hosts
+  default port 80
+- Print UI section URLs as soon as the section exists (`form_section_http_url`)
+- Timestamps: `utc_now_iso` / `is_iso_timestamp` — no `Z` / `+00:00` munging
+
+## Python style here
+
+Follow skill **edg-python-style** (2-space indent, ` = ` in kwargs,
+`f"..."` / `'...'`, match neighbors).

.agents/skills/edg-changes-example/SKILL.md

@@ -0,0 +1,108 @@
+---
+name: edg-changes-example
+description: >-
+  Encodes and decodes non-UTF-8 example bytes in Changes changelogs with
+  edg-changes-example («HHHH...» markers) and writes Changes files as UTF-8.
+  Use when editing any Changes file (src/Changes, util/Changes, …), adding
+  examples with Latin-1/EUC-JP/Shift-JIS/other non-UTF-8 bytes, or verifying
+  changelog encoding.
+---
+
+# `edg-changes-example` and Changes files
+
+## Policy (all `Changes` files)
+
+Any path named `Changes` under the monorepo (for example `src/Changes`,
+`util/Changes`, `lib_src/Changes`, `include_c++/Changes`, `nt_util/Changes`)
+must remain **valid UTF-8** on disk.
+
+- Write and save Changes entries as **UTF-8**.
+- Ordinary Unicode that is valid UTF-8 may appear as real characters (or as
+  C/C++ escapes such as `\u00b6` in examples). Do **not** wrap valid UTF-8
+  in `«…»` unless you intentionally need a raw-byte example
+  (`encode --escape-all-non-ascii`).
+- Examples that need **non-UTF-8** bytes (legacy Latin-1, EUC-JP, Shift-JIS,
+  etc.) must use `«HHHH…»` markers produced by `edg-changes-example`, never
+  raw non-UTF-8 bytes in the committed file.
+
+CLI: `dev_tools/bin/edg-changes-example` (on `PATH` when direnv is active).
+
+## Note line (required for encoded examples)
+
+When an entry’s example contains one or more `«…»` hex markers, add this
+note on its own line after the example, matching existing changelog style:
+
+```text
+Note: The above example must be decoded using the edg-changes-example tool.
+```
+
+Use that exact wording. Place it immediately after the example block (blank
+line before the next dated entry is fine).
+
+## Encode (raw → UTF-8 Changes text)
+
+Given a fragment or file that still contains raw non-UTF-8 bytes:
+
+```bash
+edg-changes-example encode raw-fragment.txt > utf8-fragment.txt
+# or:
+edg-changes-example encode -o utf8-fragment.txt raw-fragment.txt
+printf '...' | edg-changes-example encode
+```
+
+Default encode keeps maximal runs of bytes `>= 0x80` that form valid UTF-8,
+and replaces only non-UTF-8 runs with one marker `«` + hex + `»` (two hex
+digits per byte, no separators). `\uXXXX` in the text is left alone.
+
+Optional: `--escape-all-non-ascii` escapes every byte `>= 0x80`, including
+valid UTF-8 sequences (rare; only when the example must show raw bytes).
+
+## Decode (UTF-8 Changes text → raw bytes)
+
+To rebuild the real example bytes (for local reproduction, not for commit):
+
+```bash
+edg-changes-example decode utf8-fragment.txt > raw-fragment.txt
+edg-changes-example decode -o raw-fragment.txt utf8-fragment.txt
+# Whole file:
+edg-changes-example decode src/Changes > /tmp/Changes.raw
+```
+
+Decode replaces each `«HHHH…»` (or legacy `«HH»«HH»`) with the corresponding
+bytes. No other escapes are interpreted.
+
+## Workflow for a new changelog entry
+
+1. Draft the entry as UTF-8 prose.
+2. If an example needs non-UTF-8 bytes, author the example in a small raw
+   file (or produce the markers by hand only when the bytes are already
+   known), then run `edg-changes-example encode` and paste the UTF-8 result
+   into the entry.
+3. Immediately after that example, add:
+
+   `Note: The above example must be decoded using the edg-changes-example tool.`
+
+4. Insert the entry into the appropriate `Changes` file with a **UTF-8**
+   write (keep the rest of the file valid UTF-8).
+5. Verify the file still decodes as UTF-8 (raises on failure):
+
+   ```bash
+   python3 -c "open('path/to/Changes', encoding='utf-8').read()"
+   ```
+
+   Use the real path (for example `src/Changes`). A clean exit means the
+   whole file is valid UTF-8; a `UnicodeDecodeError` means the write
+   introduced non-UTF-8 bytes and must be fixed before continuing.
+
+## Do not
+
+- Commit raw Latin-1 / Shift-JIS / etc. bytes inside a `Changes` file.
+- Invent alternate note wording; match the existing sentence above.
+- Run `encode` on the entire historical `src/Changes` and rewrite it unless
+  that is an explicit, reviewed task — encode **new example fragments**, then
+  splice the UTF-8 text into the entry.
+
+## Related
+
+Always-on summary: [AGENTS.md](../../../AGENTS.md). Front-end context:
+skill **edg-cpfe-frontend**.

.agents/skills/edg-cpfe-bisect/SKILL.md

@@ -0,0 +1,65 @@
+---
+name: edg-cpfe-bisect
+description: >-
+  Finds the commit that introduced a test regression or improved behavior
+  between known good and bad front-end revisions using
+  edg-docker-test-bisect (and edg-bisect for manual steps). Use when
+  troubleshooting a failing or newly-passing test across two commits,
+  searching for the first bad commit, or locating a fix commit.
+---
+
+# EDG cpfe test bisection
+
+Prefer Docker. Put `dev_tools/bin` on `PATH` (direnv /
+[HACKING.md](../../../HACKING.md)). Project root = directory containing
+`src/cfe.c` and `dev_tools/`.
+
+Full tutorial: [dev_annex/tutorials/TEST_BISECT.md](../../../dev_annex/tutorials/TEST_BISECT.md).
+Build/test/record recipes: skill **edg-cpfe-build-test**.
+
+## When to use
+
+You need a **test that differs** between two revisions, plus:
+
+| Goal | Newer tip | Older tip |
+| --- | --- | --- |
+| Find the **regressing** commit | bad (fails) | good (passes) |
+| Find the **fixing** commit | good (passes) | bad (fails) |
+
+The CLI detects which situation you mean after verifying that the test
+behaves differently on the two commits.
+
+## Automatic bisect
+
+Canonical CLI: **`edg-docker-test-bisect`** (not `edg-docker-bisect`).
+
+```bash
+edg-docker-test-bisect [--posture POSTURE] <bad_commit> <good_commit> <test_path>
+```
+
+- **Overwrites** front-end sources during the run; restores them to **HEAD**
+  afterward (HEAD itself is not rewritten as history).
+- Tip: `--posture=quick-c-fe` for speed (C-generating `edg_x86_64` only).
+
+### Test path forms
+
+- `sandbox/foo.sft.cpp` (short form; preferred)
+- `tests/sandbox/foo.sft.cpp`
+- `tests/tests/sandbox/foo.sft.cpp`
+- Or a cwd-relative / absolute path into a valid suite
+
+If the reproducer is throwaway, put it under `sandbox`.
+
+## Manual bisect
+
+Wrapper around git-bisect style steps: `edg-bisect`
+(`start|good|bad|skip|reset|…`). Use when automatic mode is too heavy or you
+need to mark skips yourself.
+
+## Checklist
+
+1. Confirm the test **fails on bad** and **passes on good** (or the inverse
+   for a fix hunt) with `edg-docker-test --non-interactive …`.
+2. Run `edg-docker-test-bisect` with those commits and the test path.
+3. After it finishes, sources should match **HEAD**; verify with `git status`.
+4. Inspect the identified commit; do not commit or push unless asked.