| name | post-mortem |
| description | Introspective end-of-session retrospective — the agent reflects on its OWN live context (method, decisions, dead-ends, token spend, lessons) and writes an honest, root-caused report ending in a reusable playbook. Use at the end of a long or complex session (30min+, multi-phase, sub-agents spawned, or a hard problem solved) when the user says "post-mortem", "retrospective", "session retro", "debrief this session", "what did we learn", "write up this session". NOT a log parser — it digs the current context, not JSONL files. |
| argument-hint | [output-dir] |
| allowed-tools | ["Bash","Read","Write"] |
| model | opus |
| effort | high |
| context | main |
| user-invocable | true |
Post-Mortem
Produce a rich retrospective of the current live session by introspecting your own working memory — the reasoning, decisions, dead-ends, and token spend that only exist in this context, not in any file. The gold-standard output is honest, root-caused, quantified where possible, and ends in a reusable playbook.
This is introspection, NOT log-parsing. 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 you re-writing the same decoder. Do NOT read session JSONL transcripts. Read your own memory of what happened.
Output: a timestamped file post-mortem-YYYYMMDDHHmm.md (see step 1) + a 5-line spoken summary in chat. Always full — every core section, every qualifying session.
Distinct from /save-context (forward-looking resume state) and /instruct-compact (ephemeral compaction steering). This is a backward-looking debrief for learning, meant to be read by a human later.
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
These are non-obvious and counter your defaults; violating either makes the report worthless.
- You cannot measure your own main-context tokens. The naive assumption is that a "token report" means exact numbers. From inside a live session you can't see your own accumulation, cache ratios, or compaction pre/post figures — so those are
[estimate ±] with a one-line basis. What you can measure: sub-agent tokens (they arrive in the completion notifications you received) and tool-call counts you remember. Tag every number [measured] or [estimate] — an honest "~300-500k [estimate], not measurable from inside" beats a fabricated exact number.
- You will under-report your own mistakes (self-serving bias — you're grading your own homework). Counter it structurally: §Échecs must be at least as substantial as §Wins, and every dead-end gets
cause → lesson. No "I erred" without "because X, therefore next time Y".
Other guards (elaborated in reference.md): walk oldest→newest so recency bias doesn't erase the early hour; prefer "I don't recall exactly" over a tidy fabrication (per project rule: never hallucinate); force ONE idée maîtresse + a generalizable playbook, not a bullet dump.
Instructions
- Get the timestamp (you cannot generate it yourself): run
date +%Y%m%d%H%M. Build the filename post-mortem-<that>.md. Default location = project root, or the dir passed as argument.
- Detect compaction. If this session was continued from a compacted summary (a resume-summary sits in your context, or you recall a compaction fired), add the honesty notice from
reference.md §Compaction and mark pre-compaction phases lower-fidelity.
- Reconstruct from memory, oldest→newest, per Gotcha 2's ordering.
- Fill every core section from
reference.md; add optional sections when the session warrants them (menu there); never drop a core one.
- Write the file, then print the 5-line chat summary (mission · headline result · biggest win · biggest avoidable cost · one playbook line). Never auto-commit.
The sections (full detail + template in reference.md)
Core (always, all nine): Mission · Idée maîtresse (the ONE most-leveraged decision) · Chronologie · Bilan chiffré · Ce qui a marché ✅ · Échecs & fausses pistes ❌ (root-caused) · Où économiser 💰 · Playbook réutilisable · Rôle de l'humain.
reference.md holds the per-section directives, the number-tagging rules, the compaction notice, the optional-sections menu, and the exact output template. Follow it exactly.
