Auto Review

---
description: "Review-fix loop: automatically review code, fix issues, and repeat until clean"
argument-hint: "[base ref for merge-base, e.g. 'origin/develop', or empty for auto-detect]"
---

# Auto Review Loop

Automatically review code, fix found issues, and re-review until clean. Maximum **3 iterations**.

## Configuration

- **Review agent**: Use the `code-reviewer` agent (from pr-review-toolkit plugin, Opus model)
- **Diff scope**: $ARGUMENTS (see below for default behavior)
- **Max iterations**: 3
- **Commit**: Do NOT commit during the loop. Only report results at the end.

## Determining Diff Scope

If `$ARGUMENTS` is provided, use it directly as the diff scope.

If `$ARGUMENTS` is empty, auto-detect the PR base branch:
1. Run `gh pr view --json baseRefName -q '.baseRefName'` to get the PR's base branch
2. If a PR exists, use that as `<base-branch>`
3. If no PR exists, try to detect the default base branch:
   - Check for `develop` branch: `git rev-parse --verify origin/develop`
   - If not found, check for `main`: `git rev-parse --verify origin/main`
   - If not found, check for `master`: `git rev-parse --verify origin/master`
4. If none of the above work, fall back to `git diff` (unstaged changes)

Once `<base-branch>` is determined, compute the diff scope using **merge-base**:
```bash
git diff $(git merge-base origin/<base-branch> HEAD)

Why merge-base: This compares the branch divergence point against the working tree, which:

• Includes uncommitted fixes from previous iterations (unlike ...HEAD which only compares commits)
• Is unaffected by origin/<base> advancing after a fetch (unlike git diff origin/<base> which would show upstream changes as inverse diffs)

Workflow

Execute the following loop up to 3 times:

Step 1: Review

Launch a code-reviewer agent (via Task tool, subagent_type: general-purpose) with the following prompt. Make sure to pass the determined diff scope.

The agent should:

• Run git diff <diff-scope> to get the changes to review
• Review the changes following the code-reviewer guidelines (CLAUDE.md compliance, bug detection, code quality)
• Rate each issue with a confidence score (0-100)
• Only report issues with confidence >= 80
• Return a structured list of issues with file paths, line numbers, descriptions, and suggested fixes
• If no issues found, explicitly state "NO ISSUES FOUND"

Step 2: Check Results

Parse the reviewer's response:

• If the reviewer found no issues (or stated "NO ISSUES FOUND"), exit the loop and go to the Summary step.
• If issues were found, proceed to Step 3.

Step 3: Fix Issues

For each issue reported by the reviewer:

1. Read the relevant file
2. Apply the suggested fix (or implement a better fix if the suggestion is insufficient)
3. Verify the fix makes sense in context

Important: Do NOT fix issues you disagree with. Use your judgment. If a reported issue seems like a false positive, skip it.

Step 4: Re-review

Go back to Step 1 for the next iteration. The reviewer will see the updated diff including your fixes.

Important for re-review: When launching the reviewer agent for iteration 2+, include the previous iteration's issues in the prompt context so the reviewer can verify they were addressed and focus on finding genuinely new issues rather than re-reporting variations of already-fixed problems.

Summary

After the loop ends (either clean review or max iterations reached), report:

1. Iterations completed: How many review-fix cycles ran
2. Issues found and fixed: Each issue について以下を明記すること
   • 課題: レビュワーが指摘した問題の内容（ファイルパス・行番号を含む）
   • 修正方針: どのように修正したか、なぜその方針を選んだか
   • スキップした場合: 修正しなかった理由
3. Remaining issues: 最終レビューで残った未解決の指摘（理由付き）
4. Status: コードがクリーンか、未解決の指摘が残っているか

出力例:

### Iteration 1 (3 issues found)

1. [Critical] SQL injection vulnerability in `src/api/users.ts:42`
   - 課題: ユーザー入力がそのままSQLクエリに埋め込まれている
   - 修正方針: パラメータ化クエリに変更。ORMのプレースホルダー機能を使用

2. [Important] Missing null check in `src/utils/parse.ts:15`
   - 課題: parseResult が null の場合に例外が発生する
   - 修正方針: early return で null を処理し、呼び出し元の型定義も修正

3. [Important] Unused import in `src/components/App.tsx:3`
   - スキップ: リンターが検出する範囲のため対応不要と判断

### Iteration 2 (0 issues found)
Clean. No issues detected.

Notes

• Do not commit or push during this process
• If the reviewer keeps finding new unrelated issues each iteration (not re-reports of the same issue), note this in the summary
• If a fix introduces a new issue caught in the next review, prioritize fixing the regression
• Keep fixes minimal and focused - do not refactor or improve code beyond what the reviewer flagged