skill-minerlisted

# skill-miner Two jobs: (1) find where the agent actually suffers in real session logs, (2) find the best existing or drafted skill for each pain point. Never adopt anything here - that is skill-eval-loop's job, with proof. ## Mine corrections first (where skills win) ```bash python3 scripts/corrections.py --memory ~/.claude/projects/<proj>/memory --out corrections.json ``` Correction episodes are the moments the user told the agent it got something wrong: (agent message, user reply). An LLM groups them into themes, marks each one `knowledge_gap` (fixing it needs THIS user's rules, voice, formats or facts) vs generic, and maps it to the user's saved rule files (`--memory`). Knowledge-gap themes sort first. They are what a skill can measurably fix: frontier models already avoid generic mistakes, so generic-failure evals tie, while a skill carrying the user's rules wins. - Sources: Claude Code, Codex and OpenCode logs (`--sources claude,opencode,codex`, default: all found). `--per-session` caps marathon sessions so one session cannot fill the budget. - `--no-llm` extracts episodes only (free, used by the nightly timer); `--episodes FILE` clusters a saved extraction later. - The `session_index` also records **which skills each session loaded**, collected in the same pass that reads the episodes. skill-eval-loop's recheck needs it to tell a skill that earned its improvement from one that was never once loaded while the rate moved anyway. Tool and argument spellings