session-autopsylisted

# session-autopsy — failure → instruction fix A task went red → green (deploy failed, CI broke, several retries until success). This skill converts that failure into an edit that makes the wrong path unenterable for the next agent. The deliverable is an edited instruction, verified — never "a lesson learned". ## When to use - "Learn from this session / why did it fail / fix the skill so this can't happen again." - **The phrase that always works, said in the session that went wrong: "run session-autopsy"** — that session can grep its own transcript for the errors, nothing else needed. From ANOTHER session (or later): pass the session ID (it is read via `session_search`) or paste the key errors + what finally worked — evidence is required, the whole log is not. - Standing rule: any pipeline task that needed a retry — run the autopsy before closing it. - Not for: the app itself being broken (`systematic-debugging`), publishing mechanics (`skill-pack-publishing`), missing knowledge (the buildout update-loop). ## Token discipline (hard rules) - Never read a full transcript — grep for error strings and the fix commit (ref 10 §2–3). - Load only the ref the current phase names (10 → 20 → 30, one at a time — three max per autopsy); report ≤ 40 lines; interview = one batch of ≤ 6 questions. - Chat output = the report only (symptom → cause → fix → proof). No narration of the process. ## Phases — in order, no skipping 1. **Evidence** (ref 10): first symptom, verbatim errors (grep-