distill-feedback

# distill-feedback — close the learn-from-corrections loop The Stop hook `session-feedback-capture.py` queues finished sessions into `~/.claude/feedback/queue.jsonl`. This skill processes that queue: it finds the user turns that were **durable corrections** of the agent's work and turns them into rules — so the same correction never has to be given twice. **Why LLM-semantic, not keywords:** we independently tested a keyword detector. It scored F1 **0.42** on held-out cases and missed ~60% of real corrections (every keyword-free one, e.g. "в следующий раз лучше через python"). An LLM applying the rubric below scored F1 **0.97** on the same set. So detection is semantic. Evidence: `knowledge/agent-systems/self-learning-agents/effectiveness-test/RESULTS.md` (in the private hub). **Why human-gated:** ACE (arXiv 2510.04618) and TRACE (arXiv 2606.13174) both warn that a noisy extractor poisons the rule set. Altering durable rules is also above the auto-act line (`autonomy-risk-tiers.md`). So this skill **proposes**; the user approves before anything is written. ## Procedure ### 1. Extract the queue (deterministic) ```bash python ~/.claude/skills/distill-feedback/scripts/extract_feedback_queue.py --limit 8 ``` Returns JSON: `{pending, sessions:[{session_id, cwd, ts, user_turns:[...]}]}`. `--limit` bounds the LLM pass (billing: distillation is opt-in, not every-session). If `pending` is 0, stop — nothing to do. ### 2. Detect durable corrections (LLM-semantic, prefer a fresh sub...