Post-Mortem

Retrospective of a working session — reasoning, decisions, dead-ends, spend — honest, root-caused, quantified, priced, ending in a reusable playbook.

Introspection for the narrative, aggregation for the cost. The payload — "I trusted the fork's summary, it had dropped 2 lines, I lost 15 min, lesson: grep the source first" — lives only in context. A log parser sees Bash ran 110×; it cannot see that 4 of those were the same decoder rewritten. Never mine logs for the why.

Carve-out — cost only. Summing message.usage counters from session JSONL and sub-agent .output files is authorized. It is the only route to a [measured] cost: no total_cost_usd field exists, cost must be computed. Numbers may be summed, reasoning may not be mined. Recipe, sources and boundary → reference.md §Cost.

Output: post-mortem-YYYYMMDDHHmm.md + a 5-line summary in chat. Always full — every core section, every qualifying session.

Distinct from /save-context (forward-looking resume state) and /instruct-compact (compaction steering). This is a backward-looking debrief for learning.

When to Use

• End of a long (30min+) or multi-phase session, a hard problem solved, or sub-agents/forks were spawned.
• User says: "post-mortem", "retrospective", "session retro", "debrief", "what did we learn", "write up this session".
• Skip the ritual for trivial sessions: if it was a greeting or a one-file fix, say "Session is light — a post-mortem adds little" and write a 3-line note instead of the full ceremony.

Gotchas — the two rules that make or break the report