feat(plugin): plugin-suite improvements — grep, hints, skills, steering, extractor, markdown, auto-sync

[ScriptedAlchemy]

Consolidated plugin-suite improvements for Claude/Codex/Cursor, from tonight's audits, evals, and pattern research. Contains PR #278's permissions work (built on that branch) — merge this alone for everything, or merge #277/#278 first as smaller chunks and this rebases clean.

What's in it (911/911 tests green under nextest, the CI harness)

Adoption root-cause fixes

• Installer writes plugin-namespace permission allowlist (mcp__plugin_tracedecay_tracedecay__*) + migrates legacy entries — the fix that turns denied/stalled tool calls into working ones (evals: sonnet went 2/4-stalled → opus 4/4-substantive once permissions flowed)
• update_plugin refreshes the managed CLAUDE.md steering block (reaches subagents); rewritten from anti-Explore polemic to moment-triggers
• Claude SubagentStart hook + widened PostToolUse matcher (Grep|Glob|Read) — subagents finally receive steering

New capability closing the #1 native-tool leak

• tracedecay_grep: gitignore-aware literal/regex content search, graph-enriched (enclosing symbol + node id per hit) — rg was the top shell command (683 calls); nothing in the graph answered content search before

Steering that actually triggers

• Skills consolidated 30→13 (deleted 13 duplicate dispatchers; merged the 5-skill memory cluster); pattern-researched gateway skill (Iron Law, 1% rule, rationalization table keyed to observed failure modes) injected at SessionStart; descriptions at the 320/45 contract with Do-NOT boundaries + sibling routing; content/symbol/concept tool routing
• Hint engine: telemetry integrity (hint_id + terminal-outcome invariant + project_id), per-session budget + escalation, new build-diagnostics and memory-store categories

Correctness / quality

• TS extractor test-attribution: extracts describe/it callback nodes + ArrowFunction coverage universes (fixes rstest 3/62-attribution blindness across all callback-style JS frameworks)
• Markdown render pass: CLI multi-block P0 (was dropping payloads behind warnings), no JSON-in-cells, column pruning, cycle rendering, humanized timestamps
• diagnostics_prewarm config knob (env-wins precedence) for the cold-start blocker

Index freshness (auto-sync D1–D7)

• git-metadata watcher (bounded, debounced), serve-stale-then-refresh, session-start sync + harness-worktree auto-tracking, branch-store lifecycle (reflink clone + dead-store GC), [sync] config table, warning UX ("refresh in progress" vs "run tracedecay sync"), the !Send fix that unblocked the whole library

Known test-harness note

14 tests fail under in-process cargo test (shared env-lock poison cascade under load) but all pass in isolation and 911/911 pass under nextest (process-per-test = what CI runs). Two pre-existing flakes worth standalone fixes: install-family tempdir rename race; git_watch debounce timing sensitivity.

Recovery context

Session 2c51d204-3565-4a10-833d-d8fbd51620c3 · workflows wf_785c5851-aee/wf_7f23bed7-803 (plugin), wf_d002080a-6e2 (auto-sync) · facts 17–39 (tracedecay tool fact_store --action get --fact-id 39 = full manifest)

🤖 Generated with Claude Code

[changeset-bot]

⚠️ No Changeset found

Latest commit: 3326a07

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[ScriptedAlchemy]

Full recovery manifest (all session + workflow ids)

Main orchestrating session: 2c51d204-3565-4a10-833d-d8fbd51620c3 (claude, /home/zack/projects/tracedecay) — tracedecay tool lcm_load_session --provider claude --session-id 2c51d204-3565-4a10-833d-d8fbd51620c3

Workflow runs (transcripts: session dir → subagents/workflows/<run_id>/):

Run	Purpose	Status
wf_785c5851-aee	plugin-suite wave 1 (grep, telemetry, Claude surfaces, steering)	completed; E/F/H died on session limit
wf_7f23bed7-803	triple-plugin completion (all 8 tasks + review + verify)	wave 1 cached; resumed
wf_bcfa3aa1-ac9 / wf_2ede3e1d-75a / wf_d002080a-6e2	auto-sync D1–D7 (attempts 1/2/3)	3rd attempt active
wf_d0bf6fa4-48f	36-scenario eval corpus (sonnet+codex)	done → fact 28
wf_d0ae2099-bfe	obscure-tool evals + missed-opportunity hunt	done → fact 33
wf_f46d3a1c-ccb	session-git-correlation build (this feature)	active

Research agents (task transcripts in the session's tasks/ dir): a24542ee4f08eee98 markdown audit→fact 24 · a143ce24edd3525f7 adoption research→fact 21 · a07f2677c93ebc313 rstest diagnosis→fact 23 · af4d3dac14e17f59c eval baseline+opus rerun→fact 25 · a3cfc66f0ca02f263 freshness design→fact 22 · ab49d9ae2d3a5f72d skill-pattern study (children af2210c6ae0c73588, af50cc51138da9bb2)→fact 29 · a546e5bb4d17b8ac6 hermetic harness

Durable decisions: facts 17–39 in this project's fact store (tracedecay tool fact_store --action search --query 'recovery manifest' → fact 39 has this table).

[ScriptedAlchemy]

Status update (15:28 UTC)

Done since the description was written:

• Triple-plugin workflow completed: all 8 tasks landed (incl. T5: Claude SubagentStart hook, hint budget/escalation, SessionStart root-detection fix), 5 review blockers fixed, verify passed
• Auto-sync workflow completed: 5 implementations + reviews, watcher/session-start majors fixed, verify passed
• Skill-drafts polish applied (04c442a8): pattern-researched gateway + 6 rewrites, descriptions at the 320/45 contract, 23/23 skill-content tests green. Gateway-injection unification is satisfied by construction — SessionStart include_str!s the gateway SKILL.md
• diagnostics_prewarm config knob (env-wins precedence) + the !Send gix fix that unblocked lib compilation
• Independent test adjudication: agent_suite 432/433, lib 909/911 at bounded parallelism; all remaining failures pass in isolation

Known flakes (pre-existing / test-only, not blocking):

• daemon::git_watch debounce tests are timing-sensitive under CPU load (pass isolated; CI's nextest process-per-test reduces exposure)
• install-family tempdir race (memory_digest_targets.json.new atomic-rename) — worth a small standalone fix

Remaining before ready-for-review: Fable integration review of the assembled diff → rebase onto master + #277 + #278 (dedupe) → sccache/mold enablement during the rebuild window → hermetic clean-build re-smoke → eval corpus re-run (before/after vs sonnet-50%/codex-90% baseline) → reorganize wip checkpoints into logical commits.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[ScriptedAlchemy]

CI correction

CI surfaced 7 failures I missed locally — my local gate was cargo nextest run --lib (911 green), but CI runs the full workspace + clippy --all-targets under a blocking policy. The gaps, all in the auto-sync feature code (not the plugin/skills/hints work, which is solid):

• Clippy lints in src/daemon/git_watch.rs and src/branch.rs
• git_watch_test::deleted_branch_store_is_garbage_collected — GC returns empty (real logic/wiring bug)
• git_watch_test::concurrent_syncs_are_single_flight — two syncs both indexed (single-flight defect or racy test)

These are in tests/daemon_suite/ integration tests that a --lib run doesn't execute. Fix in progress (follow-up commit, no force-push). The plugin-suite core (grep, hints, skills, permissions, extractor, markdown) is unaffected and validated — the hermetic smoke on the assembled binary passed end-to-end.

[ScriptedAlchemy]

Before/after proof — adoption fix works on the assembled binary

Ran the 6 baseline-failure scenarios (the ones where Sonnet showed zero tracedecay usage) against this branch's binary in a hermetic isolated env (sonnet, tracedecay repo). 6/6 pass, and every silent-bypass scenario now fires the right tool:

scenario	baseline	now (tracedecay / native)
literal search	grep only, 0 tracedecay	1 / 0 (tracedecay_grep)
construction sites	grep, silent bypass	1 / 2 (constructors)
struct derives	grep + whole-file Read	3 / 2 (derives)
module public API	whole-file Read of mod.rs	2 / 1 (module_api)
trace call path	grep	3 / 4 (call_chain)
any compile errors?	shell cargo check	1 / 2 (diagnostics)

The five symbol-metadata tools that never triggered on tailor-made prompts in the baseline (fact 33) all trigger now. This validates the full stack end-to-end on the real binary: installer plugin-namespace permissions (calls flow instead of being denied) + tracedecay_grep (content-search gap closed) + moment-trigger skill/steering rewrites. Native tools still appear for verification, but tracedecay now leads.

(The 7 CI failures noted above are all in the bundled auto-sync tail — being fixed separately; they don't touch this adoption path.)