This was generated by AI during triage.
Split out from #462 (item 4), the foggiest of the four. The pipeline's Routines fire as fresh sessions with no memory; recurring review feedback from the maintainer on agent PRs currently evaporates instead of improving future firings.
To settle in a grill-with-docs session (this one most needs it — expect an ADR):
- Where learned lessons live: the playbook, the Routine prompts,
CONTEXT.md, a dedicated lessons file, or the steward skill (see the steward-skill issue)
- Who folds feedback back: the maintainer manually, the fix worker as part of finishing a PR, or a periodic dedicated Routine that mines merged/closed agent PRs for recurring comments
- What counts as a lesson: one-off nits vs. patterns seen across N PRs; how to avoid prompt bloat
- When: at PR close, on a schedule, or on demand
Marked ready-for-human — this is pure design; nothing is actionable until the grill session lands decisions.
Split out from #462 (item 4), the foggiest of the four. The pipeline's Routines fire as fresh sessions with no memory; recurring review feedback from the maintainer on agent PRs currently evaporates instead of improving future firings.
To settle in a grill-with-docs session (this one most needs it — expect an ADR):
CONTEXT.md, a dedicated lessons file, or the steward skill (see the steward-skill issue)Marked
ready-for-human— this is pure design; nothing is actionable until the grill session lands decisions.