self-improvelisted

# Self-Improvement Loop — Memory to Rules Pipeline Continuously evolve the toolkit by promoting proven patterns from session memory to permanent rules, and by forging new skills from recurring workflows. ## The Loop ``` Work → Capture (auto, multi-trigger) → Consolidate (auto, threshold-based) → Approve (batched, session-scoped) → Promote → Enforce ``` **H.4.1 update**: capture + consolidation + low-risk graduation now run **automatically** at multiple natural breakpoints. The user no longer has to invoke `/self-improve` to keep the loop alive — but the command still works for on-demand triage and explicit Memory→Rule promotion. ### 1. Capture (Automatic, multi-trigger) - **Stop hook** (every assistant turn): `auto-store-enrichment.js` bumps per-signal counters in `~/.claude/self-improve-counters.json` (file paths, slash commands, skill invocations). Cheap, deterministic, no LLM. - **Auto-memory** records patterns in `MEMORY.md` during sessions (existing). - **Library session-snapshots** capture verbatim session content via pre-compact hooks (v2.1.0+; see `docs/library.md`). - **Forged agents/skills** accumulate personality over time (existing). ### 2. Consolidate (Automatic, threshold-based) - **Every 30th turn** (Stop hook): triggers `self-improve-store.js scan` if turns since last scan ≥30. Catches sessions that never compact. - **PreCompact hook**: same scan triggers at compaction. Heavier review at the natural "session in retrospect" moment. - **Thresholds**: - S