commit-sweep

name	commit-sweep
description	Periodically sweep recent commits for cross-cutting problems, gotchas, and regressions that per-commit review misses. Use to audit the last N commits from a higher level (e.g. weekly, or before a release).

Commit Sweep

Per-commit review sees one diff at a time. This sweep looks across recent commits for the problems that only show up in aggregate: a pattern applied inconsistently, a fix that regressed something two commits later, docs that drifted from code, tests that went quiet.

Procedure

1. Pick the window. Default last 15 commits: git log --oneline -15. Get the combined diff: git diff HEAD~15..HEAD.
2. Sweep for, specifically:
   • Inconsistency — same idea done differently across commits (naming, error handling, a helper reimplemented instead of reused). Grep to confirm it's a pattern, not a one-off.
   • Silent regressions — a later commit that partially reverted or broke an earlier fix. Check whether the earlier fix still has a test guarding it.
   • Guard/safety drift — any new bare milo-self run, MILO_RUN_UNGUARDED, raised mem/concurrency caps (CLAUDE.md). Run bun run scripts/lint.ts --all.
   • Doc drift — for each touched system, does its doc's doc-meta key-files still match reality? Is last-verified stale? (docs/doc-standards.md)
   • Test coverage that quietly shrank — new behavior with no fixture; a .skip that slipped in; a fixture whose @expect would pass even if the feature were deleted (docs/testing.md false-confidence check).
   • TODO/FIXME left behind without an owner or issue.
3. Verify before reporting. For each suspected issue, confirm it (grep, run the fixture, break-and-test). Don't report speculation — this is where a second-model check helps: scripts/agent_review.sh implementation --diff HEAD~15.
4. Report as a terse list: commit <sha> · file:line · problem · fix. Group by theme. File anything mechanical into scripts/lint.ts and anything judgmental into CONVENTIONS.md so it's caught automatically next time.
5. Log recurring findings in docs/feedback/README.md so the workflow tightens.

Output

A ranked list, most-severe first, each with a concrete fix. If clean, say so and note what you checked — a sweep that finds nothing should still show its work.