bug-autopsy

Ask the agent to autopsy the resolved failure. It scopes the blast radius, reconstructs a timestamped timeline from git, logs, and CI evidence, runs a 5-Whys chain to a systemic cause, and writes a post-mortem with checkbox prevention items that a linter verifies.

Start from the user-visible failure and ask why down the causal chain, evidencing each level with a commit, file:line, or log reference. Go at least three levels deep and stop at the first systemic cause — a missing guardrail or test — never at a person.

The linter enforces required sections (Summary, Impact, Timeline, Root cause, Contributing factors, Fix, Prevention, Links), timestamps and evidence citations on every timeline entry and why, a whys chain at least three levels deep, and checkbox prevention items with owners. Blame-y phrasing produces warnings, not failures.

No. The skill explicitly excludes live debugging and on-call triage of a still-burning incident — stabilize first, then autopsy after. It also does not cover understanding healthy systems or auditing codebase claims.

No. The linter is stdlib-only Python 3, runs offline, and is deterministic. Git history, CI logs, and issue-tracker access improve evidence quality but are optional; the workflow degrades to whatever records exist.

Post-mortems are persisted as Postmortem concepts in an Open Knowledge Format bundle, typically under docs/knowledge/, linked from the subject's index.md with a dated log entry and re-pinned source fingerprints, so failure knowledge travels with the repo.