prjct-cli 2.19.9 → 2.20.1

package/CHANGELOG.md

@@ -1,5 +1,25 @@
 # Changelog
 
+## [2.20.1] - 2026-05-17
+
+### Bug Fixes
+
+- skill-miss-detector — crew-isolation guard (no false nags after crew runs) (v2.19.10) (#339)
+
+
+## [Unreleased]
+
+### Fixed
+
+- **skill-miss-detector no longer false-positives after a crew run (#16 follow-up).** Crew implementer/reviewer run as isolated subagents in the *shared* working tree, so at the leader's Stop hook `getModifiedFiles()` saw their edits and path-overlap relevance fired — but the leader transcript never carries the memory references the subagent made in its own isolated transcript, producing a false skill-miss nag for every crew-touched file. Fix: `detectSkillMisses` collects the `files_touched` of crew runs whose `ended_at` is within `CREW_RUN_RECENCY_MS` (6h) via `crewRunStorage.list` and excludes them from path-overlap relevance; token-overlap detection stays active so non-crew work in the same session is still covered. Crew itself is unchanged (it was architecturally correct). Best-effort — any failure degrades to prior behavior. Tests: `core/__tests__/services/skill-miss-detector.test.ts`.
+
+## [2.20.0] - 2026-05-17
+
+### Features
+
+- Skill Resolution Feedback (#16) — skill-miss detector + improvement-signals widening + prjct skill-adherence (v2.19.9) (#338)
+
+
 ## [2.19.9] - 2026-05-16
 
 ### Bug Fixes
@@ -9,6 +29,17 @@
 
 ## [Unreleased]
 
+## [2.19.9] - 2026-05-16
+
+### Added
+
+- **Skill Resolution Feedback (harness #16, spec `08344b7d`).** At Stop, `core/services/skill-miss-detector.ts` flags captured project knowledge (`decision` / `gotcha` / `anti-pattern`) that was relevant to the session's work but never referenced — a "skill-miss" — and persists it as an `improvement-signal` tagged `source:skill-miss-detector`. Transcript-primary; changed-file overlap is a best-effort booster via `getModifiedFiles()` (absent on a clean tree → degrades to token overlap, never errors). Precision-leaning: relevance needs a path hit OR ≥2 distinctive shared tokens; `inferred`-provenance and captured-this-session memories are excluded; hashed `(type,key)` dedup makes Stop re-runs idempotent (mirrors `friction-detector`).
+- **`prjct skill-adherence [window] [--md]`** — read-only QA surface: skill-misses vs resolutions over a window (7d default), with an addressed ratio. Same Tier-1 read-only contract as `prjct retro` / `prjct health`.
+
+### Changed
+
+- **`buildImprovementSignals` now carries both detectors under one block.** The recall no longer hard-filters `source:friction-detector`; friction and skill-miss signals render under the existing `# prjct: improvement signals` header with **per-source budgets** (friction ≤3, skill-miss ≤2) so a noisy friction session can't starve skill-miss out of the shared 24h window. No parallel block (session-start advisory density stays bounded). The `resolves:` prose is extended with `resolves:skill-miss`; the 24h age-out remains the actual drop-from-rotation mechanism (a real `resolves:` consumer stays owned by the memory close|forget spec).
+
 ### Fixed
 - **`getProjectId` no longer silently mints a random orphan project.** Root cause: `ConfigManager.getProjectId()` fell through to `pathManager.generateProjectId()` (`crypto.randomUUID()`) whenever `readConfig()` returned null, so any path-resolution miss (daemon resolving the wrong cwd, config transiently unreadable, case-variant path) forked a brand-new project and scattered specs/memory across ghost projects with no error surfaced. Now returns `''` — the falsy sentinel 31/32 call sites already guard with `if (!projectId)` → callers fail loud ("run prjct init") instead of writing into a random new project. Only explicit `prjct init` (`createConfig`) mints. Regression test: `core/__tests__/infrastructure/config-manager-getprojectid.test.ts`.