review-gate

Independent multi-agent review that decides whether a frename change (code PR or design doc) may merge. Use before opening or merging any agent-made PR, after every fix round, and when asked to "revie

SKILL.md

GitHub

Review gate

The author of a change never judges it. Reviewers are subagents started with fresh context: they get the issue, the design doc (if any), the commit SHA under review, and the diff against main — not the author's reasoning. Run them in parallel. A verdict applies to that SHA only.

Reviewers

Code PR — all three, every round:

  1. Correctness — tries to break the change. Reads the diff and the code around it, looks for logic errors, panics/unwrap in production paths, data loss (renames, XMP writes, comment sidecars, undo), Windows path issues, race conditions in async tasks, regressions in untouched callers. Runs the local gate. Where it suspects a bug it writes a failing test to prove it. Checks git diff origin/main... -- '*.rs' '*.toml' 'rust-toolchain*' '.cargo/**' '.github/**' for weakened gates: removed or loosened asserts, new #[ignore], deleted tests, cfg that hides a test on the CI platforms, CI steps or flags removed or relaxed. Each is a blocker unless the issue requires it. Any change to a guarded file (list in nightly → Trust) that the issue does not explicitly ask for is a blocker.
  2. Design and quality — could this be simpler, smaller, more in line with the codebase? Checks Iced Elm architecture rules (.cursor/skills/iced-elm-architecture), core/UI separation, naming, duplication, dead code, scope creep beyond the issue, missing tests for new core behaviour.
  3. Product — does it do what the issue and design ask, from the editor's point of view? Keyboard flow, discoverability, consistency with existing UI, README and version.md text (short, user language, per readme skill). For UI changes it looks at the screenshots and checks that the PR body shows every screen the change touches (a missing or wrong screenshot is a major finding). New UI text has keys in every shipped language.

Design doc (advisory only, one round, never blocks; see nightly step 3) — reviewers 2 and 3, judging the design: missing flows, edge cases, simpler alternatives, feasibility (is every format/API claim sourced?).

Verdict format

Each reviewer returns:

VERDICT: APPROVE | CHANGES_REQUIRED
FINDINGS:
- [blocker|major|minor] file:line — problem — concrete failure scenario — suggested fix

APPROVE is allowed with only minor findings. Any blocker or major means CHANGES_REQUIRED. A finding without a concrete failure scenario or a concrete improvement is dropped.

Loop

  1. Fix every blocker/major (and minors that are cheap and clearly right).
  2. Re-run the local gate.
  3. Start a new round with fresh reviewers (never reuse a reviewer that saw an earlier round — it anchors on its old findings). Give them the full current diff.
  4. Repeat until all three approve in the same round. Four rounds without that → the change is finished as far as possible and left unmerged for the owner (nightly → "Owner review").

Record every round in the PR description: SHA, each reviewer's verdict, findings count, what was fixed.