session-postmortem

Cross-cutting (WORKFLOW.md §2), applied to a session rather than a product. It reports what happened, with evidence; it does not fix, export or file anything.

Core rule: every finding cites path:line. No citation, no finding. Every number comes from the transcript or from a command that was run, never from memory.

When

After a session that repeated work, ignored its plan, stalled, skipped an expected skill, or cost more than expected, and someone wants the cause.

Steps

1. Intake, one question at a time, until there is a statement naming the session, the turn range if known, what was expected, what happened, and the observable that matters (wall-clock, tokens, a repeated action). "It took too long" is a complaint, not a statement. Nothing below starts before the operator has answered; if they are away, write the questions and stop.
2. Locate the transcripts. Resolve each session to absolute paths. Confirm a past session by quoting its first prompt and timestamp. List every candidate rejected, with the reason, or write "none". Enumerate subagent transcripts too; in those, the "user" is the parent agent.
3. Context safety, every file, every time. One record can be a megabyte.
   • Measure first (wc -lc, list lines over a size threshold).
   • Never print a transcript to read it. Get line numbers and counts first, then small fields from named lines, cut to a few hundred characters.
   • Anything longer than that for one record means the slice was too wide.
4. Analyse in seven dimensions, in parallel where the transcript is long, splitting by turn range:
   1. timeline of skills fired and not fired;
   2. adherence to the plan, compaction points first;
   3. repeated work;
   4. stumbles (errors, retries, dead ends);
   5. quality evidence (was success shown or asserted);
   6. conflicts between requests;
   7. cost and time. Discard any returned finding without path:line.
5. Report: the verdict first, then findings in the order of the operator's complaint, each with its citation, then what could not be determined.

Output / checkpoint

No pipeline checkpoint. A report with cited findings and an explicit coverage note (what was and was not read).

Guardrails

• Read-only. Never modify, move or delete a session file.
• Operator words only. Hook output, system reminders and tool results are not the operator's prompts.
• Report, do not diagnose the framework. State what the transcript shows; whoever triages decides whether a block changes. No advice unless asked.
• No bundle export, no issue drafting, no scrubbing: out of scope here by choice.

Origin

Mechanism rewritten from the diagnosing-superpowers skill of the superpowers collection (MIT), read 2026-10-02: the intake-before-analysis gate, rejected-candidate listing, the seven analysis dimensions, the path:line rule and the transcript context-safety procedure. Left out: bundle export, scrubbing, issue writing and similar-session search. Wording and structure are ours. Not run on real work.