The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
feat: integrate twelve reviewed catalog and documentation PRs
haelyra2a69663c5c5c
diff --cc docs/design/context-profile-delivery.mdindex 96939a62a,3b896f1da..7f6afcbb3--- a/docs/design/context-profile-delivery.md+++ b/docs/design/context-profile-delivery.md@@@ -120,4 -90,6 +120,8 @@@ The store recovers actual process exitSmall authenticated Claude pilots now provide task and token observations, but they are descriptive and the evaluation gate remains `review-required`. Adequately powered task-quality canaries and whole-context measurements need additional evidence. The opt-in interactive bootstrap has local source, discovery and terminal-start evidence, but authenticated task behavior and native skill invocation remain unobserved. Isolated Codex registration, switching, refresh and rollback have local native evidence; changing a live user installation still requires its own ownership and recovery contract. Fresh-install default changes, existing-user migration, other-provider activation, hook plans, ECC Tools compatibility, hosted rollout and package publication remain outside this local preview.+The optional routing index binds its advisory metadata, including its trigger snapshot, to the current managed receipt. Trigger content is not part of the carrier generation itself, so rebuilding can use newly curated triggers without changing that generation. No skill body or permission is admitted by the index; the resolver re-verifies canonical sources. Suggestions also require an exact name or the fallback's absolute evidence floor, plus the routing policy's name/trigger anchor, so weak ranking matches remain silent.+++ ### Static carrier dependency gate++ `npm run carrier-closure:check` parses planned JavaScript files with the pinned TypeScript development parser and checks literal module references against each carrier's planned files. It executes no carrier code. Invalid source syntax fails closed; dynamic module expressions and package imports are reported separately. Literal CommonJS and ESM resolution use their distinct path rules, including ESM file URL queries and fragments. Arbitrary loader aliases and shadowed bindings are outside this static check. A successful receipt is a static closure result, while provider-native loading and execution require their own runtime evidence.diff --cc docs/design/context-profile-delivery.tdd.mdindex e1c95b8ff,0d4dd7f53..e8e51d2fb--- a/docs/design/context-profile-delivery.tdd.md+++ b/docs/design/context-profile-delivery.tdd.md@@@ -88,49 -88,12 +88,57 @@@ Current combined verification after rec- The default sandbox checkout's 5,000-path capture limit truncated a real Tier 1 install diff and failed closed. The reviewed ECC-029 sandbox implementation raises the bounded cap to 50,000, passes its 26-case boundary suite, and produced both final reports. The driver receipt binds its 51-file implementation digest `a84e09ab848b8cd05f33792c13734f7aabe16bfe16d50d8f8292eb5261a93c3a`.- No real AI outcome call ran because `CODEX_API_KEY` was absent. Host ChatGPT authentication was neither copied nor exposed to the disposable evaluator.-These boundaries keep the shipped behavior distinct from the M1 release gate. Authenticated outcome observations, a complete Tier 2 disk diff, live-install migration, other-provider activation, whole-context token truth and release defaults remain unverified until their explicit prerequisites are available.+## Auto-admission anchor (routing policy v5)++The 77-prompt routing corpus from #2945 (52 direct, 25 paraphrased) found one wrong implicit load. "two services both think they own the same record" admitted `skill:santa-method` on four description-only words (`two`, `both`, `they`, `same`) with BM25 24.3 and a 1.94 margin, clearing every v4 bar.++- RED: the prompt and the corpus check fail against v4 (`skill:santa-method` loaded).+- GREEN: auto admission and the tier-2 fallback also require one matched term in the skill name or curated triggers. Exact directive citations are unchanged.+- Corpus result: zero loads outside the expected set; 19 implicit admissions become 17. The other dropped admission, `skill:cost-tracking`, now defers to the bounded proposal instead of loading.+- The existing nine-query auto/agent corpus, fallback, launch, evaluation and retrieval suites pass unchanged.++## Isolated Claude native generations++Session-only Claude discovery adapts the ownership and receipt intent of #2788 to the existing native store; it never writes a marketplace or user settings.++- RED: 22 cases fail before the provider adapter (preview, prepare, version gate, nine discovery corruptions, details mismatch, static integrity, runtime-state tolerance, provider-home roots, Codex-to-Claude staleness, task launch and interactive start).+- GREEN: all 22 pass. Codex native (39), interactive (8), launch (12), auto launch (5), evaluation (31), store (31) and profile CLI suites pass unchanged; Codex receipts keep their shape, and evaluator Claude launches without a plugin directory keep their arguments and environment.+- Credential-free real run on Linux x64 with Claude Code 2.1.292 and the actual registry: Lean prepared ready with three skills and a 398-token always-on projection; Full prepared ready with 293 skills and 35,658 tokens; Full excluding `skill:python-patterns` verified 292 skills without it at 35,561 tokens. Switching profiles reported `stale`, and native rollback restored the Lean generation. `run --dry-run` resolved the pinned binary with `--plugin-dir`; `start --dry-run` proposed the Claude generation.+- Claude Code auto-updated to 2.1.293 during review, and the exact 2.1.292 pin rejected it. RED: tests for later 2.1 patches fail. GREEN admits 2.1.292 and later 2.1 patches while rejecting 2.1.291, 2.2.0, prereleases and malformed output. A real 2.1.293 run prepared Lean ready at the same 398-token projection; preparation took 2.4 s (provider calls 1.2 s), an unchanged re-preparation 0.9 s and status 11 ms.+- No authenticated model call, interactive turn or native skill invocation ran. The projection is the host's estimate for the plugin listing, not whole-context truth.+- Maintainer credential-free macOS arm64 check, October 10: Claude Code 2.1.296 prepared Lean with three skills (~289 listing tokens), Full with the then-current 302 skills (~25,982 listing tokens), reported stale after each managed profile switch, and restored Lean through managed and native rollback. These are observed plugin-listing projections for that snapshot, not whole-context measurements.++## Prompt suggestions++The suggest-only prompt hook from #2945 now reads a generation-bound metadata index instead of its own catalog cache.++- RED: routing-index and hook cases fail before the modules exist. GREEN adds a stored-vector key to the entry shape, a stale-source refusal and the session-start builder; all 16 cases pass, with a retrieval case proving stored vectors rank identically to computed ones.+- The resolver path took 300 to 500 ms per prompt from a source checkout, almost all in rehashing 293 skills and 584 files. Loading Ajv lazily and storing vectors moved the hook's own work to about 95 to 110 ms on a Full Claude index (276 entries, 1.8 MB); wrapper invocations measured 140 to 180 ms against a 71 ms disabled baseline on Linux x64.+- RED: on the full registry the function-word prompt still produced three unanchored suggestions. GREEN applies the routing policy v5 anchor to suggestions; the prompt now yields none and a React keyboard-focus prompt still leads with `skill:frontend-a11y`.+- Stored vectors moved from JSON pairs to little-endian base64 (RED: format test fails; GREEN: exact decode). The Full index shrank from 1.8 MB to 0.97 MB, its parse from 26 ms to 6 ms, and hook work to about 89 to 98 ms; wrapper invocations measured 149 to 190 ms against a 61 ms disabled baseline. The resolver's two registry reads are left intact as its source-consistency check.+- Selection, retrieval, profile CLI and hook-wrapper suites pass unchanged.+- No interactive Claude or Codex session loaded the hook; suggestion quality is bounded by the resolver ranking measured above.++Review follow-up, October 8:++- RED (15 index, hook and selection cases fail): a rollback to an indexed generation threw instead of reporting the index missing; digest-consistent entries with a non-string ID, name, description, owner or trigger, or a non-skill ID, were accepted; an oversized index was read in full; an anchored skill ranked fourth was hidden behind three unanchored matches; a skill edited mid-build was published under the stored generation; manual mode read the index before going silent; suggest mode advertised `resolve --load`, which returns no bodies there; admission denials were matched by message text, so a non-UTF-8 skill silently left the index.+- GREEN: a valid pointer for an earlier receipt is retired (`missing`) and the SessionStart builder rebuilds it; entries are validated against a closed key set before decoding and returned as new objects; pointer and index reads are bounded at 4 KiB, 4 MiB and 2,048 entries; suggestions filter for anchors before trimming to three; a build refuses plan or registry digests that differ from the checked carrier; the hook reads only the state binding before its manual-mode exit; suggest mode prints `ecc profile mode auto` with the current revision; denials carry `ECC_CONTEXT_*` codes and everything else propagates.+- Budget: a worker thread would make the 150 ms budget a hard bound but measured 135 to 150 ms for the same work (about 85 ms in process), so it would suppress most suggestions. The hook keeps in-process checks between steps, the read bounds cap the work at a real registry's size, and the documented snippet sets Claude Code's `timeout` as the outer bound.+- Full Claude index, auto mode, Linux x64 Node 22.22.0: wrapper invocations 108 to 122 ms against 29 to 45 ms with the hook disabled.++Review follow-up, October 9:++- RED (1 case fails): a build interrupted while writing the entries file left partial bytes under its digest name, and every later build refused them, so the index could not be restored. GREEN writes the file through the store's temporary-file-and-rename path and replaces a digest-named file whose bytes do not hash to its name.+- Sources edited after the entries are read: a new case pins that the published index still matches the stored generation and that the next build reports the generation stale. A final source recheck before the pointer would only discard a correct index. The triggers manifest, though, is not bound to the generation: an edit to it after `ecc profile set` reaches the next index unnoticed. Binding it is a store change, left open.+- Schema validation stays manual on the read path. On Windows x64, Node 24, loading Ajv took about 44 ms and compiling an index schema about 20 ms in a fresh process, against 22 ms for loading the whole routing module. That would cost the prompt-time read about 64 ms of its 150 ms budget, so Ajv stays off the prompt path, as `validateSchema` already documents.+- Budget: the October 8 decision stands. On the same Windows machine a prototype worker measured 64 to 77 ms against about 55 ms in process, and stopped a simulated 600 ms read at 152 ms. That gap is smaller than the Linux measurement above, so the trade-off may be worth revisiting with numbers from both platforms.++These boundaries keep the shipped behavior distinct from the M1 release gate. Authenticated outcome observations, a complete Tier 2 disk diff, live-install migration, additional-provider activation, whole-context token truth and release defaults remain unverified until their explicit prerequisites are available.++ ## October 10 carrier closure parser follow-up++ Independent review reproduced literal loads hidden after object-literal division and postfix increment, native `module.require` calls being skipped, and malformed source yielding a successful empty scan. The original hand-written lexer returned no dependencies for all four RED fixtures.++ The maintained parser in the existing pinned TypeScript development dependency now reads source without evaluating it. The gate traverses literal `require`, native `module.require`, dynamic `import`, static imports/re-exports, and literal `path.join(__dirname, ...)` calls. Existing regex, template, escape and nested-dynamic cases remain covered. Parse diagnostics fail the projection closed. ESM URL queries and fragments identify the same planned file; encoded separators and malformed escapes fail rather than normalize into a safe path.++ Focused validation passes 67 cases under Node 24.9.0 and Node 18.20.8. The actual CLI projects the current 304-skill registry into 10 profile/layout combinations, with 3,016 planned files, 65 JavaScript scripts and 35/35 literal relative specifiers resolved, zero escapes and zero warnings. Script execution remains deferred. Dynamic expressions are reported as warnings; arbitrary loader aliases, shadowed bindings and provider-native runtime behavior remain outside this literal-dependency gate.