Review

Unified Review

Review code using Codex CLI (always) and optionally Cursor Agent CLI. Cross-reference findings, debate disagreements, filter by severity, and apply fixes.

Commands

Codex (always):

codex exec --full-auto --sandbox read-only --cd <project_directory> "<request>"

Cursor (when enabled):

cursor agent --print --model composer-2 --trust --workspace <project_directory> --mode ask "<request>"

Prompt Rules

Every request sent to Codex or Cursor MUST include both sentences verbatim:

1. "No questions or confirmations needed. Proactively output specific proposals, fixes, and code examples."
2. "Filter findings by: (1) Critical issues (bugs, security, design flaws), (2) Issues worth fixing that are easy to address. Omit minor nitpicks and style preferences."

Workflow

Phase 0: Delegate to a Background Subagent

External review tools take minutes to respond. Never run them inline in the main conversation — the user's session would block with no visible progress.

If you are the main conversation agent: do not execute Phases 1–6 yourself. Instead:

1. Launch a background subagent that inherits the conversation context (Claude Code: Agent tool with subagent_type: "fork"). Prompt it with: "You are the review subagent. Read <absolute path to this SKILL.md> and execute its workflow from Phase 1 onward. Scope: <scope from the user's request>. Report findings only — do not modify any files."
2. Tell the user in one line that the review is running in the background, then return control immediately. Continue other pending work; do not wait for the review.
3. When the subagent's completion notification arrives, relay its report to the user as-is (do not re-run the review), then handle fix application per Phase 6.

If you are the review subagent: skip this phase and execute Phases 1–6 directly. Never apply fixes yourself — produce the report only; the main agent owns file changes.

Phase 1: Determine Scope

If $ARGUMENTS specifies files or a scope, use that. Otherwise default to changed files since the last commit:

git diff HEAD --name-only

If the diff is empty, inform the user and stop.

Phase 2: Determine Review Type

Examine the scope and classify:

• Plan — implementation plans, task lists, TODO documents
• Design — architecture docs, design decisions
• Code — source code files (default)

Phase 3: Execute Reviews

1. Read .claude/settings.json in the project root. If it contains "cursorEnabled": true, Cursor will also be used.
2. Compose the review prompt including both mandatory prompt rules.
3. Execute:
   • Codex only (default): single Bash call
   • Codex + Cursor: two independent Bash calls in parallel

Phase 4: Scrutinize and Debate

Review every finding yourself.

When you doubt or disagree with a finding, start a discussion autonomously:

1. Quote the specific point from the tool's response
2. State your concern or counterargument concretely
3. Send a follow-up prompt to that tool, including: "Be frank and direct. Don't hold back — push back candidly." (plus both standard prompt rules)
4. Continue until convergence (no round limit)
5. Record the outcome

When tools disagree with each other (Codex + Cursor mode):

1. Present both positions
2. State your own analysis
3. Follow up with the tool whose position you find weaker
4. Resolve and record

Phase 5: Classify and Filter

Assign severity to each surviving finding:

Severity	Action	Criteria
CRITICAL	Must fix	Bugs, security vulnerabilities, data loss, crashes
HIGH	Must fix	Design flaws, race conditions, significant logic errors
MEDIUM	Consider	Performance, maintainability, minor logic gaps
LOW	Skip	Style preferences, minor nitpicks, cosmetic issues

Drop all LOW findings.

Phase 6: Report and Fix

Report findings to the user:

## レビュー結果

### スコープ
<files reviewed>

### 使用ツール
Codex / Codex + Cursor

---

### CRITICAL
- [ ] <finding with file:line and explanation>

### HIGH
- [ ] <finding with file:line and explanation>

### MEDIUM（検討）
- <finding with brief explanation>

---

### クロスリファレンス（両ツール使用時のみ）
一致: N件 / Codexのみ: N件 / Cursorのみ: N件 / 意見相違（解決済）: N件

### 議論結果（議論発生時のみ）
**論点**: <topic>
**対象ツール**: <Codex or Cursor>
**結論**: <resolution and reasoning>

---

### 適用した修正
- <what was changed and why>

Apply fixes:

• CRITICAL and HIGH: apply immediately without asking
• MEDIUM: list for user decision, do not apply automatically