commit-sweep

name	commit-sweep
description	Audit a range of recent ottochain commits, one review-persona lens per pass, and file findings as GitHub issues + a worksheet. Never fixes inline. Use to retro-review merged/landed work for defects the original review missed.

commit-sweep

Retro-review landed commits through each review persona in turn. Output is findings (issues + a worksheet), NOT fixes — a fix rides its own PR with its own review.

Input

A since-ref or range: --since 2026-07-01, or origin/main~20..origin/main, or a PR number's merge range.

Sequence

1. bin/worksheet commit-sweep-<date> — open a worksheet to hold findings.
2. List the range: git log --oneline --since=<date> (or the explicit range). Note the diffs that touch blast-radius files (docs/agents/blast-radius.md) — those get the most scrutiny.
3. One persona lens per pass. For each of the five personas, sweep the whole range with ONLY that persona's checklist in mind:
   • bin/agent-review consensus-safety <range> --dry-run (or run it) — signed messages, validators, combiners, determinism, the C1–C3/H1–H2/M1–M3/L-class defects.
   • bin/agent-review wire-compat <range> — OpenAPI/SDK/e2e drift, version lockstep, path prefixes.
   • bin/agent-review state-growth-determinism <range> — unbounded state, 512KB cap, gas, SortedMap.
   • bin/agent-review asset-economics <range> — conservation, consent, custody, nonce linearity.
   • bin/agent-review ai-smells-test-integrity <range> — tautological/self-regenerating tests, claim-vs-reality. Record each pass's findings in the worksheet under a per-persona heading.
4. For each real finding, file a GitHub issue: gh issue create with a clear title (lowercase, ≤72), the persona label, the commit/file:line, the defect class, and why it matters. Add blast-radius or needs-senior labels where warranted.
5. Link the issues back into the worksheet. Append a docs/agents/feedback.md entry if the sweep revealed a systemic gap (e.g. a missing guard suite).

Rules

• Never fix inline. A sweep produces findings; fixes are separate, reviewed PRs.
• Prefer false-positive-tolerant reporting: note uncertain findings as questions, not assertions — a human triages. But do not manufacture findings to look thorough; "range is clean against persona X" is a valid result.
• The claim-vs-reality check (ai-smells) is the highest-value pass: verify that each commit message's claims are backed by the diff and a runnable command (the #162 lesson).