brain-scanlisted

# Brain Scan Find the recurring tasks this person's agents actually perform, then measure whether relevant Skills improve the real outputs when explicitly loaded. ## Choose a mode - **Fast (default):** identify the highest-signal recurring task from the session history and run 3 valid controlled pairs. Use for a quick directional answer and first Skill recommendation. - **Deep:** cover at least 3 recurring task families and complete at least 10 valid controlled pairs in total. Use when the user asks for a comprehensive scan, a broader capability map, or stronger evidence before changing the agent setup. Ask which mode only when the user has not indicated depth and the extra Deep runtime or model usage would be material. Otherwise default to Fast. Invalid, contaminated, or timed-out pairs do not count toward either minimum. ## Default assessment Use exactly two arms: - `base`: the selected Skill is absent. - `loaded`: the same model, task, fixtures, tools, limits, and ordinary context, with the selected Skill's frozen `SKILL.md` content supplied before work begins. Do not add a discovery arm. Do not replace this comparison with current setup versus an incremental prompt patch. Never promise uplift. ## Run 1. Confirm which local session roots the user authorizes. Raw logs stay local. 2. Extract correction episodes without a model call: ```bash python3 <skill-dir>/vendor/skill-miner/scripts/corrections.py \ --sources claude,opencode,codex --no-llm --out <pri