Review Changes

Code Review

Review and report. Skip the brainstorming; the user wants findings, not a conversation.

1. Detect the review target

Gathering commands (no args)

# Check what changed — staged first, then unstaged, then last commit
git diff --cached --stat
git diff --stat
git diff HEAD~1 --stat

Pick the first non-empty scope. If all three are empty, say so and stop.

For PR reviews, also gather context:

gh pr view <n> --json title,body,labels,additions,deletions,changedFiles
gh pr diff <n>

2. Read full files, not just diffs

A diff tells you what moved; surrounding code tells you whether it still makes sense. Without that context, you miss broken call sites, stale comments, and silent convention violations.

• For every changed file, read at least the modified function/class plus its immediate callers and callees.
• If the diff is > 800 lines, read the full files for the highest-risk changes. Skip mechanical changes (renames, formatting) unless they mask something.

3. Learn conventions before flagging them

Skim CLAUDE.md, AGENTS.md, or similar project instructions. Read sibling files in the same directory and the last few commits touching these files. "This doesn't match the codebase style" only lands as a finding when you can point to where the real style lives.

4. What counts as a finding

Flag something only if all of the following are true:

1. It meaningfully impacts correctness, security, performance, or maintainability.
2. It was introduced by this change — pre-existing issues are out of scope unless the change makes them worse.
3. It is discrete and actionable — the author can fix it in one focused pass.
4. It does not require assuming unstated intent or speculative downstream effects.
5. It does not demand a level of rigor not present in the rest of the codebase.
6. The author would likely fix it if they were made aware of it.
7. It is not clearly an intentional change by the original author.

Do not flag:

• Trivial style unless it obscures meaning or violates documented standards.
• Pre-existing bugs unrelated to this change.
• Things that "might" break something unless you can identify the specific code path that provably breaks.
• Hypothetical issues without a concrete scenario.
• Speculative downstream effects — you must be able to identify the code that is provably affected.

How many findings to return: Output all findings that the original author would fix if they knew about it. Do not stop at the first qualifying finding. Continue until you've listed every qualifying issue. If there is nothing a person would definitely want to see and fix, prefer outputting no findings.

5. Review checklist

Check for:

• Logic bugs — wrong condition, off-by-one, missing null/undefined check, race condition.
• Security — injection, authz bypass, secret exposure (OWASP top 10 as a lens, not a checklist).
• Convention violations — only when you can cite the convention.
• Missing or weak tests — new behavior without test coverage.
• Dead or unreachable code introduced by this change.
• Breaking changes — removed exports, changed signatures, config format changes.

6. Finding format

Use this format for every finding:

### {n}. [P0–P3] Short imperative title

**File:** `path/to/file.ts:42-48`

What's wrong and why it matters — one paragraph, matter-of-fact tone.

If suggesting a fix, use an inline code block (≤ 3 lines). No commentary inside the block.

Priority levels:

Tone & content rules:

• Matter-of-fact, not accusatory or overly positive.
• No "Great job" or "Thanks for" — the user asked for findings, not compliments.
• Brief. One paragraph per finding. Do not introduce line breaks within the natural language flow unless necessary for a code fragment.
• The comment should clearly and explicitly communicate the scenarios, environments, or inputs necessary for the bug to arise. Immediately indicate if severity depends on these factors.
• The author should grasp the idea without close reading.
• Use ```suggestion blocks only for concrete replacement code (minimal lines; no commentary inside the block). Preserve the exact leading whitespace of replaced lines.
• Any code chunks should be ≤ 3 lines and wrapped in inline code tags or a code block.
• Avoid providing unnecessary location details in the comment body — rely on the file/line reference instead.

7. Overall verdict

After all findings, give a one-line verdict:

**Verdict:** PATCH IS CORRECT / PATCH HAS ISSUES — <1-sentence why>

Use PATCH HAS ISSUES if any P0 or P1 finding exists. Use PATCH IS CORRECT otherwise. "Correct" means existing code and tests will not break, and the patch is free of bugs and other blocking issues. Ignore non-blocking issues such as style, formatting, typos, documentation, and other nits.

8. Large diffs

When the diff spans many files or > 800 lines of real changes (not renames/formatting):

1. Read the diff stat to identify high-risk files (core logic, auth, data handling).
2. Prioritize P0/P1 coverage on those files.
3. Note in the report which files you reviewed deeply vs. skimmed.
4. If sub-agents are available, split by file group — one agent per group.

For purely mechanical changes (renames, formatting, generated code), say so and skip detailed review.

9. Document review

When the input is a .md file, review it as target architecture — the desired future state, not current code. Flagging "this doesn't match what's built yet" defeats the purpose; the author already knows.

Look for:

• Completeness — every stated requirement has a plausible implementation path.
• Internal consistency — no section contradicts another.
• Feasibility — realistic given the codebase, team, and constraints you can see.
• Gaps — error handling, migration/rollback, edge cases, observability, failure modes.

Report document findings using the same P0–P3 format, replacing file/line with section/heading.

10. Follow-up

After the report, ask what the user wants next:

• Fix the critical/high items
• Discuss a specific finding
• Expand scope to other files or branches