---
name: pr-review-pass
description: Structured review of small feature PRs in three separate passes - correctness, then tests, then style - with each pass reported on its own. Use when reviewing a pull request or a diff under roughly 400 changed lines, especially when the review should end in actionable comments rather than a general vibe.
---

# PR Review Pass

## When to use

- Reviewing a feature PR or a focused diff (under ~400 changed lines)
- The review output must be concrete comments, not a summary

Do NOT use for giant refactors, generated code, or merges of other
people's long-lived branches. Say the PR is too big for this pass and
stop.

## The three passes - always in this order

### Pass 1: Correctness

Read the diff line by line. Hunt only for things that are WRONG:

- Off-by-one and boundary conditions
- Null/undefined paths the new code introduces
- Error handling that swallows or misroutes failures
- Race conditions in handlers the diff touches
- Behaviour the PR description promises but the code does not do

### Pass 2: Tests

- Is every new behaviour in Pass 1 covered by a test that would fail if
  the behaviour broke?
- Are the assertions checking behaviour, not implementation detail?
- Name the exact missing test cases - "add a test where X is null" -
  never "needs more tests".

### Pass 3: Style

Last, and clearly separated. Naming, structure, idioms. Cap this pass
at the five comments that matter most. If everything is fine, say so in
one line; do not manufacture style comments.

## Output format

Three headed sections, one per pass, in order. Each finding: file and
line, what is wrong, why it matters, a suggested change. End with a
verdict line: how many correctness issues, how many missing tests, how
many style notes.

## Hard boundaries

- Never approve or merge. The review is advice; the human decides.
- Never mix passes. A correctness bug reported in the style section
  gets ignored; keep the lanes clean.
- If you are unsure whether something is a bug, say "verify:" and ask
  the question instead of asserting it.
