Retro

Propose improvements to the agent's environment based on where a session struggled. Propose only: change nothing until the user picks a candidate. Fix the environment, not the code the session produced (that is code-review).

Steps

1. Create a task per step below, each with its completion criterion.
2. Get the session record:
   • No session named: use the current conversation.
   • Session named, or the current one is too long to recall reliably: run python3 scripts/session_digest.py <id-or-path> (--list N --project <repo> to find it). Digest each subagent transcript it lists when the struggle happened there.
3. Mark the struggle moments: errors, retries, hook blocks, long searches, oversized results, compactions, user corrections, wrong turns. Cite each by digest call number (#12) or by quoting the user.
4. Read the repo's existing checks (Makefile, package.json scripts, pre-commit config, CI workflows, .claude/settings.json hooks). A check that exists but is unwired or broken is the finding, in place of a new one.
5. Map each struggle moment to a candidate using the categories below. Drop any candidate you can't trace to a struggle moment.
6. Present candidates ranked by severity: cost of the struggle times how often it will recur. Each candidate states the moment, the fix, and the file it lands in.

Categories