post-mortem

Post-Mortem Skill

Purpose: Wrap up completed work — validate it shipped correctly, extract learnings, process the knowledge backlog, activate high-value insights, and retire stale knowledge.

Runtime note: Hook-driven closeout is runtime-dependent. Claude/OpenCode can wire Phase 2-5 maintenance through lifecycle hooks. Codex CLI v0.115.0+ supports native hooks (same behavior). For older Codex versions without hook surfaces, finish closeout with ao codex stop.

Loop position

Move 7 (capture evidence + learning, then ratchet) of the [operating loop](../../docs/architecture/operating-loop.md). Two outputs per loop turn: evidence (test names, snapshot keys, council verdicts, citation events) recorded against the bead and .agents/ratchet/; learnings promoted only under the [ratchet rules](../../docs/architecture/operating-loop.md#the-promotion-ratchet) — noticed once stays in the handoff, repeats twice goes to .agents/learnings/, changes future behavior updates a SKILL.md or template, must-never-regress becomes a gate, core doctrine promotes into PRODUCT.md/GOALS.md/docs/cdlc.md. Most observations die at handoff. That is correct.

Six phases:

1. Council — Did we implement it correctly?
2. Extract — What did we learn?
3. Process Backlog — Score, deduplicate, and flag stale learnings
4. Activate — Promote high-value learnings to MEMORY.md and constraints
5. Retire — Archive stale and superseded learnings
6. Harvest — Surface next work for the flywheel

Quick Start

/post-mortem                    # wraps up recent work
/post-mortem epic-123           # wraps up specific epic
/post-mortem --quick "insight"  # quick-capture single learning (no council)
/post-mortem --scope=pr <num>   # learn from a merged/rejected PR outcome (absorbed /pr-retro)
/post-mortem --process-only     # skip council+extraction, run Phase 3-5 on backlog
/post-mortem --skip-activate    # extract + process but don't write MEMORY.md
/post-mortem --deep recent      # thorough council review
/post-mortem --mixed epic-123   # cross-vendor (Claude + Codex)
/post-mortem --skip-checkpoint-policy epic-123  # skip ratchet chain validation

Codex Closeout

Codex CLI v0.115.0+ has native hooks and handles closeout automatically (no extra steps needed). For older Codex versions (hookless fallback), run these after the post-mortem workflow writes learnings and next work:

ao codex stop
ao codex status

ao codex stop uses the latest transcript or history fallback to queue/persist learnings and run close-loop maintenance without runtime hooks.

Flags

| Flag | Default | Description | |------|---------|-------------| | --quick "text" | off | Quick-capture a single learning directly to .agents/learnings/ without running a full post-mortem. Formerly handled by /retro --quick. | | --scope=pr [num] | off | Read a merged/rejected PR outcome as the wrap-up signal instead of a closed bead/epic — mine reviewer feedback into a PR learning. Absorbed the former pr-retro skill. See [references/pr-scope.md](references/pr-scope.md). | | --process-only | off | Skip council and extraction (Phase 1-2). Run Phase 3-5 on the existing backlog only. | | --skip-activate | off | Extract and process learnings but do not write to MEMORY.md (skip Phase 4 promotions). | | --deep | off | 3 judges (default for post-mortem) | | --mixed | off | Cross-vendor (Claude + Codex) judges | | --explorers=N | off | Each judge spawns N explorers before judging | | --debate | off | Two-round adversarial review | | --skip-checkpoint-policy | off | Skip ratchet chain validation | | --skip-sweep | off | Skip pre-council deep audit sweep |

Quick Mode

Read [references/quick-mode.md](references/quick-mode.md) when you need the --quick flag procedure (slug generation, direct learning write, confirmation).

PR-Outcome Scope (--scope=pr)

Read [references/pr-scope.md](references/pr-scope.md) when invoked with --scope=pr. It swaps the bead/epic wrap-up signal for a PR's merge/reject/changes-requested outcome: discover the PR, analyze the outcome, mine reviewer feedback via gh, extract success/failure patterns, and write a dated PR learning to .agents/learnings/. After the PR learning lands, the standard maintenance phases (process → activate → retire → harvest) run as usual. The trigger phrases "PR retro", "learn from this PR", and the legacy pr-retro command all route here.

Execution Steps

Read [references/execution-steps.md](references/execution-steps.md) when you need the full Phase 1 procedure: pre-flight checks, reference loading (Step 0.4), checkpoint-policy preflight (0.5), plan/spec loading (Steps 1-2.3), closure integrity audit (2.4), metadata verification (2.5), deep audit sweep (2.6), council invocation (Step 3), and prediction accuracy (3.5).

Step 2.1: Load Compiled Prevention Context

Before council and retro synthesis, load compiled prevention outputs when they exist:

• .agents/planning-rules/*.md
• .agents/pre-mortem-checks/*.md

Use these compiled artifacts first, then fall back to .agents/findings/registry.jsonl only when compiled outputs are missing or incomplete. Carry matched finding IDs into the retro as Applied findings / Known risks applied context so post-mortem can judge whether the flywheel actually prevented rediscovery.

Phase 2: Extract Learnings

Read [references/phase-2-extract.md](references/phase-2-extract.md) when you need the inline learning extraction procedure: gather context (EX.1), classify (EX.2), write learnings (EX.3), test pyramid gap analysis (EX.3.5), scope classification (EX.4), findings registry (EX.5-6).

Before backlog processing, normalize reusable council findings into .agents/findings/registry.jsonl.

Use the tracked contract in docs/contracts/finding-registry.md:

• persist only reusable findings that should change future planning or review behavior
• require dedup_key, provenance, pattern, detection_question, checklist_item, applicable_when, and confidence
• applicable_when must use the controlled vocabulary from the contract
• append or merge by dedup_key
• use the contract's temp-file-plus-rename atomic write rule

After the registry mutation, refresh compiled outputs immediately so the same session can benefit from the updated prevention set. If hooks/finding-compiler.sh exists, run:

bash hooks/finding-compiler.sh --quiet 2>/dev/null || true

Step ACT.3: Feed Next-Work

Actionable improvements identified during processing -> append one schema v1.4 batch entry to .agents/rpi/next-work.jsonl using the tracked contract in [../../docs/contracts/next-work.schema.md](../../docs/contracts/next-work.schema.md) and the write procedure in [references/harvest-next-work.md](references/harvest-next-work.md). Follow the claim/finalize lifecycle documented in references/harvest-next-work.md.

The machine-checkable typed contract is the committed JSON Schema pair [../../schemas/next-work-batch.v1.schema.json](../../schemas/next-work-batch.v1.schema.json) (one JSONL line = one batch entry) and [../../schemas/next-work-item.v1.schema.json](../../schemas/next-work-item.v1.schema.json) (each items[] element). Validate written rows with [../../scripts/validate-next-work.sh](../../scripts/validate-next-work.sh) (--strict to reject malformed output naming the offending field; advisory by default). The inline bash in references/harvest-next-work.md remains a dependency-free fallback when the script is unavailable.

mkdir -p .agents/rpi
# Build VALID_ITEMS via the schema-validation flow in references/harvest-next-work.md
# Then append one entry per post-mortem / epic.
# If a harvested item already maps to a known proof surface, preserve it on the
# item as "proof_ref" instead of burying target IDs in free text. Example item:
# [{"title":"Verify the parity gate after proof propagat