Codex System — Deep Reasoning Partner

Codex CLI (gpt-5.3-codex) is your highly capable supporter for deep reasoning tasks.

詳細ルール: .claude/rules/codex-delegation.md

Context Management (Opus 4.6)

Claude の 1M コンテキストにより、直接呼び出しの許容範囲が拡大した。ただし大きな出力はサブエージェント経由を推奨。

状況	方法
短い質問（〜50行回答）	直接呼び出しOK
詳細な設計相談	サブエージェント経由（推奨）
デバッグ分析	サブエージェント経由（推奨）
Agent Teams 内での相談	Teammate が直接呼び出し

When to Consult (MUST)

Situation	Trigger Examples
Design decisions	「どう設計？」「アーキテクチャ」 / "How to design?"
Debugging	「なぜ動かない？」「エラー」 / "Debug" "Error"
Trade-off analysis	「どちらがいい？」「比較して」 / "Compare" "Which?"
Complex implementation	「実装方法」「どう作る？」 / "How to implement?"
Refactoring	「リファクタ」「シンプルに」 / "Refactor" "Simplify"
Code review	「レビューして」「確認して」 / "Review" "Check"

When NOT to Consult

•Simple file edits, typo fixes
•Following explicit user instructions
•git commit, running tests, linting
•Tasks with obvious single solutions
•Codebase analysis → Claude does this directly (1M context)

How to Consult

In Agent Teams (Preferred for /startproject)

Architect Teammate が Codex を直接呼び出し、Researcher Teammate と双方向通信する。

code

/startproject 内の Phase 2 で、Architect Teammate として Codex を活用:
- 設計検討 → Codex に相談
- Researcher からの調査結果を受けて設計を修正
- 設計決定を .claude/docs/DESIGN.md に記録

Subagent Pattern (Standalone consultation)

Use Task tool with subagent_type='general-purpose' for larger outputs.

code

Task tool parameters:
- subagent_type: "general-purpose"
- run_in_background: true (optional, for parallel work)
- prompt: |
    Consult Codex about: {topic}

    codex exec --model gpt-5.3-codex --sandbox read-only --full-auto "
    {question for Codex}
    " 2>/dev/null

    Return CONCISE summary (key recommendation + rationale).

Direct Call (Up to ~50 lines response)

bash

codex exec --model gpt-5.3-codex --sandbox read-only --full-auto "Brief question" 2>/dev/null

Sandbox Modes

Mode	Use Case
`read-only`	Analysis, review, debugging advice
`workspace-write`	Implementation, refactoring, fixes

Language Protocol

•Ask Codex in English
•Receive response in English
•Execute based on advice (or let Codex execute)
•Report to user in Japanese

Task Templates

Design Review

bash

codex exec --model gpt-5.3-codex --sandbox read-only --full-auto "
Review this design approach for: {feature}

Context:
{relevant code or architecture}

Evaluate:
1. Is this approach sound?
2. Alternative approaches?
3. Potential issues?
4. Recommendations?
" 2>/dev/null

Debug Analysis

bash

codex exec --model gpt-5.3-codex --sandbox read-only --full-auto "
Debug this issue:

Error: {error message}
Code: {relevant code}
Context: {what was happening}

Analyze root cause and suggest fixes.
" 2>/dev/null

Code Review

See: references/code-review-task.md

Refactoring

See: references/refactoring-task.md

Integration with Gemini

Task	Use
Need external research first	Gemini → then Codex
Design decision	Codex directly
Library comparison	Gemini research → Codex decision
/startproject	Agent Teams: Researcher (Gemini) ↔ Architect (Codex)

Why Codex?

•Deep reasoning: Complex analysis and problem-solving
•Code expertise: Implementation strategies and patterns
•Consistency: Same project context via context-loader skill
•Parallel work: Background execution or Agent Teams teammate