postmortem

name	postmortem
description	Post-mortem a context compaction — surface which main-thread tasks should have been subagents. Use when the user asks for /postmortem, or when reviewing why a session compacted.

/postmortem — Dispatch-Floor Post-Mortem

Post-mortem the most recent (or an older) context compaction. Surface which main-thread writes should have been dispatched to subagents.

Data source: ~/.claude/metrics/compactions.jsonl (written by the PreCompact hook). Each row references the transcript that led up to the compaction.

Dispatch-floor doctrine: any task that touches more than one file or produces more than 40 lines should be dispatched to a subagent. Compaction is the failure signal.

Step 1: Run the analyzer

python3 ~/.claude/skills/postmortem/analyze.py

Optional integer arg: how many compactions back. 0 (default) = most recent. 1 = second-most-recent. Etc.

python3 ~/.claude/skills/postmortem/analyze.py 2   # three compactions ago

The script prints a JSON digest to stdout. If the metrics file is missing or empty, it prints a friendly message to stderr and exits 1 — report that to the user and stop.

Step 2: Read the digest

The digest shape:

{
  "compaction": { "ts": "...", "trigger": "auto|manual", "session_id": "...", "cwd": "..." },
  "transcript": { "path": "...", "message_count": N },
  "totals": {
    "dispatches": 5,
    "inline_writes": 12,
    "inline_bash": 3,
    "inline_bash_heavy": 1,
    "inline_write_lines_est": 487
  },
  "inline_writes_over_floor": [ { "tool": "Write", "file_path": "...", "lines": 62 }, ... ],
  "dispatches": [ { "description": "...", "subagent_type": "..." }, ... ]
}

Step 3: Write the post-mortem

150 words or less. Deliver the signal, stop.

Include:

• When the compaction fired (ts), message count for the session, trigger (auto vs manual).
• Dispatch ratio: X inline writes vs Y dispatches.
• Top 3 inline writes over the 40-line floor, by file path and line count.
• One concrete suggestion for next session — something specific the transcript shows (e.g., "dispatch wiki-page rewrites next time", "batch the Edit storm on foo.rs into a coder brief").

Don't moralize. Don't explain the doctrine. Don't pad. Assume the reader knows what dispatching is — this is diagnostic output, not a lecture.

Notes

• If totals.dispatches is already high and inline_writes_over_floor is short, say so — "dispatch discipline was solid, the compaction came from volume not structure." Don't manufacture failure.
• If the transcript can't be read, the script exits 1 — surface that error to the user verbatim, don't silently continue.