Run postmortem

A finished agent run reports what it intended; this skill establishes what it did. Read-only reconstruction, no universes, no replays — the deliverable is a report the user can land/trim/redo from. A completion claim in a session log is not evidence; every claim in the report cites a SHA, a plan path, a session turn, or a command you ran yourself.

Sibling of catch-up (which summarizes for re-entry) — run-postmortem audits, with receipts.

Freeze

Pin what you're auditing before reading anything: repo, branch, HEAD SHA, dirty state, and a base. Prefer a base the user names, else the earliest relevant plan's commit, else merge-base with trunk. If no base is defensible, stop and say so — a postmortem without a base audits nothing.

Done when the report header names branch, HEAD, base SHA, and dirty state.

Plan lineage

Inventory every plan the run worked from: exported plan files (prompt-exports/, docs/plans/, wherever this repo keeps them), the original user request, and superseding revisions. For each: task, gates, and the alternatives it considered but discarded — discards are where the run's judgment shows.

When plan exports are missing, degrade explicitly: reconstruct intent from session history and commit messages, and mark the lineage reconstructed so the reader knows it's inference.

Done when every discovered plan is classified relevant / superseded / abandoned, or the lineage is marked reconstructed.

Session recovery

RepoPrompt: agent_manage op=list_sessions + get_log for workers; history op=search for older sessions. Gotcha (from workflows/README.md): history indexes tool summaries only — a zero-match search reads like absence and lies. Before believing any negative ("the review never ran"), grep the session JSON under ~/Library/Application Support/RepoPrompt CE/Workspaces/*/AgentSessions/.

Keep summaries, not transcripts — dispatch an explore agent for bulk log reduction and hold only its findings.

Done when every relevant session has an ID, its contribution, and a terminal state — or the gap is named in the report.

Commit ownership

Walk base..HEAD plus the dirty snapshot. Map every changed region to a plan item, a justified enabling change, or unowned. Note reverts, and work that landed then got disconnected (wired in, later orphaned).

Done when every in-scope commit and dirty path has an owner or is listed unowned.

Rot

Run rot-hunt over the generation: patterns the run planted first and then copied. Its seed table lands in the report verbatim.

Working / not working — with receipts

The core of the report. For each planned outcome, one row:

• Claim — the behavior the run says it delivered
• Receipt — what proves it on this branch: a command you ran now, a runtime log from the run showing the new code path fired, or not_run + one reason (blocked / not_applicable / instrumentation_absent)
• Status — working / not_working / unverified

A green test suite alone caps a row at unverified — tests prove the code can pass tests; a receipt proves the feature ran. Prefer running the named gates and the feature itself; where you can't, say not_run and why rather than promoting the run's own claim.

Done when every planned outcome has a row and no row's receipt is another agent's assertion.

Verdict and rollup

Close with:

1. Verdict — DONE / DONE_WITH_LIMITATIONS / PARTIAL / NOT_DONE / INDETERMINATE, one sentence of justification
2. The working/not-working table
3. Rot seed table with remove/tolerate calls
4. Unowned changes and disconnected work
5. Recommendation — land as-is, trim first (name the cuts), fix the seeds first (ordered list from rot-hunt), or redo with the lessons (name the lessons: what the next run's plan must say differently)

Write the report to docs/investigations/postmortem-<branch>-<date>.md when that folder exists, else present inline. Landing, trimming, and redoing all wait for the user.

When this skill is the wrong fit

• Re-entry summary, no audit needed → catch-up
• One recurring pattern, run already understood → rot-hunt directly
• Reviewing a human-authored branch → aa-second-opinion / taste-review