retro

name	retro
description	Session retrospective that turns this session's transcript into at most 10 principle-level knowledge proposals, each behind explicit user approval. Run at the end of a substantial session, or when the user asks for a retro.
invocation	user
disable-model-invocation	true

retro — session retrospective

Evidence-based shape (autodidact methodology/learning-loop.md): deterministic pre-pass → parallel fresh-context analysts → single curator merge → per-proposal approval. Never re-review the same context repeatedly — one fresh pass beats looping.

Procedure

1. Pre-pass (deterministic). Export or locate the session transcript, then:

   python3 "$(dirname "$SKILL_PATH")/scripts/signals.py" scan \
     --transcript <transcript-file>
   

   The JSON lists corrections (dominant signal), errors, and repeated-failure clusters, secret-redacted.

2. Parallel lens analysts (fresh context). Spawn three subagents in one batch — each gets ONLY the transcript path + the signals JSON, none of this session's conversation:

   • corrections lens: what did the user correct, and what rule would have prevented each correction?
   • errors lens: which failures repeated, and what check/knowledge would have avoided the retry cluster?
   • successes lens: what worked well enough to become a reusable pattern? Each analyst returns findings abstracted to principle level (instance-level learnings decay; principles compound).

3. Curator merge (you). Merge the three lists: dedupe, drop anything that is really codebase-derivable, keep at most 10 proposals (CONTRACT §Tunables). Before finalizing, read .autodidact/retro-acceptance.md if present and deprioritize finding types the user has repeatedly rejected.

4. Per-proposal approval. Present each proposal separately (topic, the principle, suggested outcome mapping). Nothing is written silently.

5. Capture accepted proposals through the capture CLI (--source retro; outcome: approve for patterns that worked, correct for corrected behavior, decline for things to avoid) — see the capture skill; never hand-write entries.

6. Log acceptance. Append one line per proposal to .autodidact/retro-acceptance.md (create the dir if needed): YYYY-MM-DD accepted|rejected <lens> — <topic>. This log is what makes step 3's self-tuning possible.