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.

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 Retire — Archive stale and superseded learnings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 --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.
--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

Given /post-mortem --quick "insight text":

Quick Step 1: Generate Slug

Create a slug from the content: first meaningful words, lowercase, hyphens, max 50 chars.

Quick Step 2: Write Learning Directly

Write to: .agents/learnings/YYYY-MM-DD-quick-<slug>.md

---
type: learning
source: post-mortem-quick
date: YYYY-MM-DD
maturity: provisional
utility: 0.5
---

# Learning: <Short Title>

**Category**: <auto-classify: debugging|architecture|process|testing|security>
**Confidence**: medium

## What We Learned

<user's insight text>

## Source

Quick capture via `/post-mortem --quick`

This skips the full pipeline — writes directly to learnings, no council or backlog processing.

Quick Step 3: Confirm

Learned: <one-line summary>
Saved to: .agents/learnings/YYYY-MM-DD-quick-<slug>.md

For deeper reflection, use `/post-mortem` without --quick.

Done. Return immediately after confirmation.

---

Execution Steps

Pre-Flight Checks

Before proceeding, verify:

1. Git repo exists: git rev-parse --git-dir 2>/dev/null — if not, error: "Not in a git repository"
2. Work was done: git log --oneline -1 2>/dev/null — if empty, error: "No commits found. Run /implement first."
3. Epic context: If epic ID provided, verify it has closed children. If 0 closed children, error: "No completed work to review."

If --process-only: Skip Pre-Flight Checks through Step 3. Jump directly to Phase 3: Process Backlog.

Step 0.4: Load Reference Documents (MANDATORY)

Before Step 0.5 and Step 2.5, load required reference docs into context using the Read tool:

REQUIRED_REFS=(
  "skills/post-mortem/references/checkpoint-policy.md"
  "skills/post-mortem/references/metadata-verification.md"
  "skills/post-mortem/references/closure-integrity-audit.md"
  "skills/post-mortem/references/four-surface-closure.md"
)

For each reference file, use the Read tool to load its content and hold it in context for use in later steps. Do NOT just test file existence with [ -f ] -- actually read the content so it is available when Steps 0.5 and 2.5 need it.

If a reference file does not exist (Read returns an error), log a warning and add it as a checkpoint warning in the council context. Proceed only if the missing reference is intentionally deferred.

Step 0.5: Checkpoint-Policy Preflight (MANDATORY)

Read references/checkpoint-policy.md for the full checkpoint-policy preflight procedure. It validates the ratchet chain, checks artifact availability, and runs idempotency checks. BLOCK on prior FAIL verdicts; WARN on everything else.

Step 1: Identify Completed Work and Record Timing

Record the post-mortem start time for cycle-time tracking:

PM_START=$(date +%s)

If epic/issue ID provided: Use it directly.

If no ID: Find recently completed work:

# Check for closed beads
bd list --status closed --since "7 days ago" 2>/dev/null | head -5

# Or check recent git activity
git log --oneline --since="7 days ago" | head -10

Step 1.5: RPI Session Metrics

Read .agents/rpi/rpi-state.json and extract session ID, phase, verdicts, and streak data. If absent or unparseable, skip silently. Prepend a tweetable summary to the report: > RPI streak: N consecutive days | Sessions: N | Last verdict: PASS/WARN/FAIL. See references/streak-tracking.md for extraction logic and fallback behavior.

Step 2: Load the Original Plan/Spec

Before invoking council, load the original plan for comparison:

1. If epic/issue ID provided: bd show <id> to get the spec/description
2. Search for plan doc: ls .agents/plans/ | grep <target-keyword>
3. Check git log: git log --oneline | head -10 to find the relevant bead reference

If a plan is found, include it in the council packet's context.spec field:

{
  "spec": {
    "source": "bead na-0042",
    "content": "<the original plan/spec text>"
  }
}

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.

Step 2.2: Load Implementation Summary

Check for a crank-generated phase-2 summary:

PHASE2_SUMMARY=$(ls -t .agents/rpi/phase-2-summary-*-crank.md 2>/dev/null | head -1)
if [ -n "$PHASE2_SUMMARY" ]; then
    echo "Phase-2 summary found: $PHASE2_SUMMARY"
    # Read the summary with the Read tool for implementation context
fi

If available, use the phase-2 summary to understand what was implemented, how many waves ran, and which files were modified.

Step 2.3: Reconcile Plan vs Delivered Scope

Compare the original plan scope against what was actually delivered:

1. Read the plan from .agents/plans/ (most recent)
2. Compare planned issues against closed issues (bd children <epic-id>)
3. Note any scope additions, removals, or modifications
4. Include scope delta in the post-mortem findings

Step 2.4: Closure Integrity Audit (MANDATORY)

Read references/closure-integrity-audit.md for the full procedure. Mechanically verifies:

1. Evidence precedence per child — every closed child resolves on the strongest available evidence in this order: commit, then staged, then worktree
2. Phantom bead detection — flags children with generic titles ("task") or empty descriptions
3. Orphaned children — beads in bd list but not linked to parent in bd show
4. Multi-wave regression detection — for crank epics, checks if a later wave removed code added by an earlier wave
5. Stretch goal audit — verifies deferred stretch goals have documented rationale

Include results in the council packet as context.closure_integrity. WARN on 1-2 findings, FAIL on 3+.

If a closure is evidence-only, emit a proof artifact with bash skills/post-mortem/scripts/write-evidence-only-closure.sh and cite at .agents/releases/evidence-only-closures/<target-id>.json. Record evidence_mode plus repo-state detail for replayability. A valid durable packet is acceptable audit evidence even when the child intentionally has no scoped-file section.

Step 2.5: Pre-Council Metadata Verification (MANDATORY)

Read references/metadata-verification.md for the full verification procedure. Mechanically checks: plan vs actual files, file existence in commits, cross-references in docs, and ASCII diagram integrity. Failures are included in the council packet as context.metadata_failures.

Step 2.6: Pre-Council Deep Audit Sweep

Skip if --quick or --skip-sweep.

Before council runs, dispatch a deep audit sweep to systematically discover issues across all changed files. This uses the same protocol as /vibe --deep — see the deep audit protocol in the vibe skill (skills/vibe/) for the full specification.

In summary:

1. Identify all files in scope (from epic commits or recent changes)
2. Chunk files into batches of 3-5 by line count (<=100 lines -> batch of 5, 101-300 -> batch of 3, >300 -> solo)
3. Dispatch up to 8 Explore agents in parallel, each with a mandatory 8-category checklist per file (resource leaks, string safety, dead code, hardcoded values, edge cases, concurrency, error handling, HTTP/web security)
4. Merge all explorer findings into a sweep manifest at .agents/council/sweep-manifest.md
5. Include sweep manifest in council packet — judges shift to adjudication mode (confirm/reject/reclassify sweep findings + add cross-cutting findings)

Why: Post-mortem council judges exhibit satisfaction bias when reviewing monolithic file sets — they stop at ~10 findings regardless of actual issue count. Per-file explorers with category checklists find 3x more issues, and the sweep manifest gives judges structured input to adjudicate rather than discover from scratch.

Skip conditions:

• --quick flag -> skip (fast inline path)
• --skip-sweep flag -> skip (old behavior: judges do pure discovery)
• No source files in scope -> skip (nothing to audit)

Step 3: Council Validates the Work

Council Verdict:

Run /council with the retrospective preset and always 3 judges:

/council --deep --preset=retrospective validate <epic-or-recent>

Default (3 judges with retrospective perspectives):

• plan-compliance: What was planned vs what was delivered? What's missing? What was added?
• tech-debt: What shortcuts were taken? What will bite us later? What needs cleanup?
• learnings: What patterns emerged? What should be extracted as reusable knowledge?

Post-mortem always uses 3 judges (--deep) because completed work deserves thorough review.

Four-Surface Closure: Validate all four surfaces -- Code, Documentation, Examples, and Proof. A PASS verdict requires all four surfaces addressed, not just code correctness. Read skills/post-mortem/references/four-surface-closure.md for the closure checklist and common gaps.

Timeout: Post-mortem inherits council timeout settings. If judges time out, the council report will note partial results. Post-mortem treats a partial council report the same as a full report — the verdict stands with available judges.

The plan/spec content is injected into the council packet context so the plan-compliance judge can compare planned vs delivered.

With --quick (inline, no spawning):

/council --quick validate <epic-or-recent>

Single-agent structured review. Fast wrap-up without spawning.

With debate mode:

/post-mortem --debate epic-123

Enables adversarial two-round review for post-implementation validation. Use for high-stakes shipped work where missed findings have production consequences. See /council docs for full --debate details.

Advanced options (passed through to council):

• --mixed — Cross-vendor (Claude + Codex) with retrospective perspectives
• --preset=<name> — Override with different personas (e.g., --preset=ops for production readiness)
• --explorers=N — Each judge spawns N explorers to investigate the implementation deeply before judging
• --debate — Two-round adversarial review (judges critique each other's findings before final verdict)

Step 3.5: Prediction Accuracy (Pre-Mortem Correlation)

When a pre-mortem report exists for the current epic (ls -t .agents/council/*pre-mortem*.md), cross-reference its prediction IDs against actual vibe/implementation findings. Score each as HIT (prediction confirmed), MISS (did not materialize), or SURPRISE (unpredicted issue). Write a ## Prediction Accuracy table in the report. Skip silently if no pre-mortem exists. See references/prediction-tracking.md for the full table format and scoring rules.

Phase 2: Extract Learnings

Inline extraction of learnings from the completed work (formerly delegated to the retro skill).

Step EX.1: Gather Context

# Recent commits
git log --oneline -20 --since="7 days ago"

# Epic children (if epic ID provided)
bd children <epic-id> 2>/dev/null | head -20

# Recent plans and research
ls -lt .agents/plans/ .agents/research/ 2>/dev/null | head -10

Read relevant artifacts: research documents, plan documents, commit messages, code changes. Use the Read tool and git commands to understand what was done.

If retrospecting an epic: Run the closure integrity quick-check from references/context-gathering.md (Phantom Bead Detection + Multi-Wave Regression Scan). Include any warnings in findings.

Step EX.2: Classify Learnings

Ask these questions:

What went well?

• What approaches worked?
• What was faster than expected?
• What should we do again?

What went wrong?

• What failed?
• What took longer than expected?
• What would we do differently?

What did we discover?

• New patterns found
• Codebase quirks learned
• Tool tips discovered
• Debugging insights
• Test pyramid gaps found during implementation or review

For each learning, capture:

• ID: L1, L2, L3...
• Category: debugging, architecture, process, testing, security
• What: The specific insight
• Why it matters: Impact on future work
• Confidence: high, medium, low

Step EX.3: Write Learnings

Write to: .agents/learnings/YYYY-MM-DD-<topic>.md

---
id: learning-YYYY-MM-DD-<slug>
type: learning
date: YYYY-MM-DD
category: <category>
confidence: <high|medium|low>
maturity: provisional
utility: 0.5
---

# Learning: <Short Title>

## What We Learned

<1-2 sentences describing the insight>

## Why It Matters

<1 sentence on impact/value>

## Source

<What work this came from>

---

# Learning: <Next Title>

**ID**: L2
...

Step EX.3.5: Test Pyramid Gap Analysis

Compare planned vs actual test levels per the test pyramid standard (test-pyramid.md in the standards skill). For each closed issue: check planned test_levels metadata against actual test files. Write a ## Test Pyramid Assessment table (Issue | Planned | Actual | Gaps | Action). Gaps with severity >= moderate become next-work.jsonl items with type tech-debt.

Step EX.4: Classify Learning Scope

For each learning extracted in Step EX.3, classify:

Question: "Does this learning reference specific files, packages, or architecture in THIS repo? Or is it a transferable pattern that helps any project?"

• Repo-specific -> Write to .agents/learnings/ (existing behavior from Step EX.3). Use git rev-parse --show-toplevel to resolve repo root — never write relative to cwd.
• Cross-cutting/transferable -> Rewrite to remove repo-specific context (file paths, function names, package names), then:
  1. Write abstracted version to ~/.agents/learnings/YYYY-MM-DD-<slug>.md (NOT local — one copy only)
  2. Run abstraction lint check:

     file="<path-to-written-global-file>"
     grep -iEn '(internal/|cmd/|\.go:|/pkg/|/src/|AGENTS\.md|CLAUDE\.md)' "$file" 2>/dev/null
     grep -En '[A-Z][a-z]+[A-Z][a-z]+\.(go|py|ts|rs)' "$file" 2>/dev/null
     grep -En '\./[a-z]+/' "$file" 2>/dev/null
     

     If matches: WARN user with matched lines, ask to proceed or revise. Never block the write.

Note: Each learning goes to ONE location (local or global). No promoted_to needed — there's no local copy to mark when writing directly to global.

Example abstraction:

• Local: "Compile's validate package needs O_CREATE|O_EXCL for atomic claims because Zeus spawns concurrent workers"
• Global: "Use O_CREATE|O_EXCL for atomic file creation when multiple processes may race on the same path"

Step EX.5: Write Structured Findings to Registry

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

This registry is the v1 advisory prevention surface. It complements learnings and next-work; it does not replace them.

Step EX.6: Refresh Compiled Prevention Outputs

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

This promotes registry rows into .agents/findings/*.md, refreshes .agents/planning-rules/*.md and .agents/pre-mortem-checks/*.md, and rewrites draft constraint metadata under .agents/constraints/. Active enforcement still depends on the constraint index lifecycle and runtime hook support, but compilation itself is no longer deferred.

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 and the write procedure in references/harvest-next-work.md. Follow the claim/finalize lifecycle documented in references/harvest-next-work.md.

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 propagation lands","type":"task","severity":"medium","source":"council-finding","description":"Re-run the targeted validator after the follow-up lands.","target_repo":"agentops","proof_ref":{"kind":"execution_packet","run_id":"6f36a5640805","path":".agents/rpi/runs/6f36a5640805/execution-packet.json"}}]
ENTRY_TIMESTAMP="$(date -Iseconds)"
SOURCE_EPIC="${EPIC_ID:-recent}"
VALID_ITEMS_JSON="${VALID_ITEMS_JSON:-[]}"

printf '%s\n' "$(jq -cn \
  --arg source_epic "$SOURCE_EPIC" \
  --arg timestamp "$ENTRY_TIMESTAMP" \
  --argjson items "$VALID_ITEMS_JSON" \
  '{
    source_epic: $source_epic,
    timestamp: $timestamp,
    items: $items,
    consumed: false,
    claim_status: "available",
    claimed_by: null,
    claimed_at: null,
    consumed_by: null,
    consumed_at: null
  }'
)" >> .agents/rpi/next-work.jsonl

Step ACT.4: Update Marker

date -Iseconds > .agents/ao/last-processed

This must be the LAST action in Phase 4.

Phases 3-6 (Maintenance): Read references/maintenance-phases.md for backlog processing, activation, retirement, and harvesting phases. Load when --process-only flag is set or when running full post-mortem.

Step 7: Report to User

Tell the user:

1. Council verdict on implementation
2. Key learnings
3. Any follow-up items
4. Location of post-mortem report
5. Knowledge flywheel status
6. Suggested next /rpi command from the harvested ## Next Work section (ALWAYS — this is how the flywheel spins itself)
7. ALL proactive improvements, organized by priority (highlight one quick win)
8. Knowledge lifecycle summary (Phase 3-5 stats)

The next /rpi suggestion is MANDATORY, not opt-in. After every post-mortem, present the highest-severity harvested item as a ready-to-copy command:

## Flywheel: Next Cycle

Based on this post-mortem, the highest-priority follow-up is:

> **<title>** (<type>, <severity>)
> <1-line description>

Ready to run:

/rpi ""

Or see all N harvested items in `.agents/rpi/next-work.jsonl`.

If no items were harvested, write: "Flywheel stable — no follow-up items identified."

---

Integration with Workflow

/plan epic-123
    |
    v
/pre-mortem (council on plan)
    |
    v
/implement
    |
    v
/vibe (council on code)
    |
    v
Ship it
    |
    v
/post-mortem              <-- You are here
    |
    |-- Phase 1: Council validates implementation
    |-- Phase 2: Extract learnings (inline)
    |-- Phase 3: Process backlog (score, dedup, flag stale)
    |-- Phase 4: Activate (promote to MEMORY.md, compile constraints)
    |-- Phase 5: Retire stale learnings
    |-- Phase 6: Harvest next work
    |-- Suggest next /rpi --------------------+
                                              |
    +----------------------------------------+
    |  (flywheel: learnings become next work)
    v
/rpi "<highest-priority enhancement>"

---

Examples

Wrap Up Recent Work

User says: /post-mortem

What happens:

1. Agent scans recent commits.
2. Runs /council --deep validate recent.
3. Extracts learnings, processes backlog, and promotes items.
4. Harvests next-work to .agents/rpi/next-work.jsonl.

Result: Report with learnings, stats, and a suggested /rpi command.

Other Modes

• Epic-specific: /post-mortem ag-5k2 — review against the target plan
• Quick capture: /post-mortem --quick "insight" — write a learning without council
• Process-only: /post-mortem --process-only — run backlog processing only
• Cross-vendor: /post-mortem --mixed ag-3b7 — broaden judgment coverage

Troubleshooting

Problem	Cause	Solution
Council times out	Epic too large or too many files changed	Split post-mortem into smaller reviews or increase timeout
No next-work items harvested	Council found no tech debt or improvements	Flywheel stable — write entry with empty items array to next-work.jsonl
Checkpoint-policy preflight blocks	Prior FAIL verdict in ratchet chain without fix	Resolve prior failure (fix + re-vibe) or skip checkpoint-policy via --skip-checkpoint-policy
Metadata verification fails	Plan vs actual files mismatch or missing cross-references	Include failures in council packet as context.metadata_failures — judges assess severity

---

See Also

• skills/council/SKILL.md — Multi-model validation council
• skills/vibe/SKILL.md — Council validates code (/vibe after coding)
• skills/pre-mortem/SKILL.md — Council validates plans (before implementation)

Reference Documents

• references/harvest-next-work.md
• references/learning-templates.md
• references/plan-compliance-checklist.md
• references/closure-integrity-audit.md
• references/security-patterns.md
• references/checkpoint-policy.md
• references/metadata-verification.md
• references/context-gathering.md
• references/output-templates.md
• references/backlog-processing.md
• references/activation-policy.md
• references/prediction-tracking.md
• references/retro-history.md
• references/streak-tracking.md
• references/maintenance-phases.md
• references/four-surface-closure.md