Problem
The same recommendation can recur across many learning files and sessions yet never get applied, because review is a separate manual step. (In one real project an identical recommendation appeared in ~20 learning files across 4 epics with zero applied — the seeded skill files were untouched since init.) Capture works; the feedback loop doesn't close.
Proposed change
- In
review, rank proposals by recurrence ("seen N times across M sessions") so the strongest rise to the top.
- At session end, surface the top recurring learnings and prompt (via the runner skill flow) to run
review / apply before the next run.
- Emit canonical repo-root skill/agent paths in learning "suggested updates" (not ephemeral
.worktrees/issue-N/.agents/... or run-prefixed paths) so auto-apply resolves correctly.
Files
src/commands/review.ts, src/lib/learning.ts.
Problem
The same recommendation can recur across many learning files and sessions yet never get applied, because
reviewis a separate manual step. (In one real project an identical recommendation appeared in ~20 learning files across 4 epics with zero applied — the seeded skill files were untouched sinceinit.) Capture works; the feedback loop doesn't close.Proposed change
review, rank proposals by recurrence ("seen N times across M sessions") so the strongest rise to the top.review/ apply before the next run..worktrees/issue-N/.agents/...or run-prefixed paths) so auto-apply resolves correctly.Files
src/commands/review.ts,src/lib/learning.ts.