grayhook
1e5cd07784
Post-fix script and doc: automated tests cover loose defer path; short RPC session keeps skill in kept so compact reinject not triggered. Co-authored-by: Cursor <cursoragent@cursor.com>
1.6 KiB
B-002 post-fix E2E (Phase 14)
Run: node scripts/b002-repro-post-fix.mjs (LiteLLM Eltex-Coder-Senior, skill at ~/.cursor/skills/fup-blame-commits).
Patches ~/.pi/agent/settings.json temporarily (skillReinject.debug: true, requireRegistered: false, autoCompactIntegration: defer).
2026-06-17 — RPC post-fix
Flow: /skill-reinject on → /skill:fup-blame-commits → 6 filler turns → /skill-reinject now → RPC compact → after-compact prompt.
| Check | Result |
|---|---|
Unit regression (test/reinject*.ts, test/reinject-deferred-consume.test.ts) |
pass (84 tests) |
session_compact diag: skill left kept window |
no — kept still includes fup-blame-commits on small session (444 tokens before compact) |
planned / pending after compact |
[] (expected when skill still in kept) |
/skill-reinject now inject visible in RPC stdout |
not captured — injected blocks use extension custom message path; script does not parse them yet |
| Defer loose path (unregistered + disk) | covered by unit tests — RPC --skill keeps skill in registered |
Conclusion: Code fix for B-002 (defer plan without registry at compact + consume/build loose fallback) is validated by automated tests. Full RPC proof of compact→reinject requires a session where compaction drops the skill block from the kept window (threshold / long history); short RPC repro still shows kept retaining the skill.
Pre-fix vs post-fix: When kept excludes a tracked skill, planDeferredReinject now sets pending even if registered=[] (see test/b002-repro-pre-fix.test.ts case 3 + test/reinject.test.ts).