Review

Complements /code-review, does not replace it: run that skill for its checks, then add the two passes below. Findings from both go into one list.

Two passes, one does not replace the other

1. Static pass over the diff, line by line. Read the diff as a reviewer who cannot run anything (this is what Copilot does well and what reproduction misses). For every changed hunk ask:

• who else writes the same state? list every writer, not only the one in the diff;
• what can run between an await and the write that follows it? a newer request, a newer effect, a cleanup;
• which other paths do the same job? a second effect, a second caller, a fallback branch - a fix applied to one of them is not a fix;
• what happens on the early return, the error branch, the aborted branch? which flags stay set;
• does a guard cover every exit, or only the happy one?

2. Reproduction pass in the running app. Reproduce the defect before the fix and re-check after it. Force the case if it is not reachable by hand (hold a promise, stub a value) instead of declaring it untestable.

The static pass catches what a scenario never hits; reproduction catches what the diff does not show. Skipping either one is how a regression ships.

Apply both to own changes too: before the words "готово" / "done", re-read own diff as if someone else wrote it.

Reporting

Say plainly which findings are proven by code, which are reproduced, and which are guesses.

When reviewing someone else's review (Copilot, a colleague), check each finding against the current code before agreeing: some are written against an older revision and are already fixed.

Never publish

Findings go into the chat only. No inline comments, no PR comment, no review, no --comment, no gh pr comment / gh pr review. The author decides what goes into the PR, finding by finding.