ai-session-watch

# ai-session-watch Project-local instinct learning for `ai-engineering`. Two modes: passive observation during a session, and active consolidation on demand. No daemons, no background workers -- the LLM itself is the observer. ## Quick start ``` /ai-session-watch # passive observation mode (silent until --review) /ai-session-watch --review # consolidate observations into observations.yml ``` ## Workflow 1. Observation is **always-on** via the `instinct-observe.py` hooks (PreToolUse + PostToolUse) — no manual activation, and `/ai-start` does NOT start it. Invoking bare `/ai-session-watch` is optional and only emits a one-line acknowledgement. 2. As work happens, the model passively notes corrections, error recoveries, and skill-invocation sequences. 3. Before a commit or PR, invoke `/ai-session-watch --review` to run the 5-step consolidation: extract → enrich → write → evaluate → create work items. 4. High-confidence proposals (>= 0.7, evidence >= 3) auto-generate work items via `gh issue create` or `az boards work-item create`. ## Artifact Set | Artifact | Purpose | | ---------------------------------------------------- | --------------------------------------------------------------- | | `.ai-engineering/state/observation-events.ndjson` | Append-only observation stream from hooks. Retain last 30 days. | | `.ai-engineering/observations/observations.yml` ...