codex-collab

name

Codex Collab

description

This skill should be used when the user asks to "collaborate with Codex", "use Codex for planning", "get Codex review", "delegate to Codex", "code review", "PR review", "review this code", "review this PR", "Codexと協調", "Codexにレビュー", "Codexに計画を作成させたい", "Codexに任せる", "Codexに委任", "Codexと連携", "Codexに相談", "Codexの意見", "Claude plans", "Claude-led", "Claudeが計画", "Claude主導", "Codexに実装させる", "Codexに実装を任せる", "Claudeがレビュー", "コードレビューして", "PRをレビューして", "コードをレビューして", "このコードをレビュー", or mentions coordinating tasks between Claude Code and Codex CLI. NOTE: This is the default skill for generic code/PR review requests. Use devils-advocate only when the user explicitly asks for adversarial/red-team critique of a design or proposal.

Codex Collaboration Skill

Coordinate tasks between Claude Code and OpenAI Codex CLI using adaptive workflow selection based on model strengths.

Overview

This skill enables effective collaboration between two AI systems with two workflow modes:

Codex-Leads (従来):

• Codex: Planning, code review, architectural decisions
• Claude Code: Implementation, file operations, testing

Claude-Leads (新規):

• Claude Code: Deep analysis, planning, code review
• Codex: Fast implementation with workspace-write sandbox

デフォルト（auto）は常に Codex-Leads を選択。claude-leads は workflow: claude-leads を明示指定した場合のみ有効。

通信方式: MCP primary + Bash fallback のデュアルモード。

• MCP mode (mcp__codex__codex/codex-reply): ステートフルな会話（threadId で文脈保持）。ANSI 除去不要、ファイル I/O 不要。
• Bash mode (codex exec/codex review): ステートレス実行。codex_run_exec() が入出力、ANSI 除去、exit code ハンドリングを統合処理。MCP 未設定時の自動フォールバック。

Prerequisites

Before starting collaboration:

1. MCP tools check (primary): Try mcp__codex__codex with a lightweight probe. If available → MCP mode.
2. CLI fallback: If MCP unavailable, verify codex CLI is available: which codex or codex --version
3. Verify codex exec works (Bash mode): codex exec -s read-only - <<< "test"
4. Check for project settings in .claude/codex-collab.local.md
5. If neither MCP nor CLI available, inform user and proceed with Claude-only mode

Workflow: Review Type (Default)

Phase 1: Task Analysis

When receiving a task for collaboration:

1. Parse the task description to identify:

   • Core objective
   • Affected files/components
   • Complexity level
   • Required context

2. Gather relevant context:

   • Read related files
   • Check existing tests
   • Review recent changes

Phase 2: Request Plan from Codex

MCP mode (primary):

mcp__codex__codex(prompt: "[planning prompt]", sandbox: "read-only")
→ Returns response + threadId (saved for Phase 4)

Bash mode (fallback):

source scripts/codex-helpers.sh
PROMPT_FILE=$(codex_write_prompt "$PLANNING_PROMPT" "plan")
OUTPUT_FILE="$(codex_tmp_path 'codex-plan-output.md')"
codex_run_exec "$PROMPT_FILE" "$OUTPUT_FILE" "read-only"

Read results from tool response (MCP) or output file (Bash).

Phase 3: Implement Based on Plan

After receiving Codex's plan:

1. Validate the plan is reasonable
2. Present plan to user for confirmation
3. Execute implementation step by step
4. Track changes made

Phase 4: Request Review from Codex

After implementation:

1. Stage changes for Codex visibility (important!):

git add -A
git reset -- tmp/ 2>/dev/null || true

Why? Staging ensures all changes are visible to Codex regardless of its file discovery method.

2. Run review (mode-dependent):

MCP mode (primary):

# Get diff, embed in prompt, continue same thread from Phase 2
mcp__codex__codex-reply(threadId: "[from Phase 2]", prompt: "[review prompt with diff]")
→ Parse verdict directly from response (no codex_infer_verdict needed)

Bash mode (fallback):

source scripts/codex-helpers.sh
REVIEW_OUTPUT="$(codex_tmp_path 'codex-review-output.md')"
codex_run_review "$REVIEW_OUTPUT" "$MODEL" || REVIEW_EXIT=$?

if [ "${REVIEW_EXIT:-0}" -ne 0 ]; then
  REVIEW_PROMPT_FILE=$(codex_write_prompt "$REVIEW_PROMPT" "review")
  codex_run_exec "$REVIEW_PROMPT_FILE" "$REVIEW_OUTPUT" "read-only"
fi

3. Parse verdict and findings (Bash mode):

RESPONSE=$(cat "$REVIEW_OUTPUT")
VERDICT=$(codex_infer_verdict "$RESPONSE") || true
FINDINGS=$(codex_extract_review_findings "$RESPONSE")

The review uses [P1]-[P4] priority markers as the primary source for verdict inference:

• [P1]/[P2] → fail
• [P3]/[P4] → conditional
• No findings + sufficient output → pass
• Metadata block (verdict: pass/conditional/fail) is also supported as an alternative

Phase 5: Handle Review Result

Based on review verdict:

Pass: Report completion to user

Conditional:

1. Apply suggested improvements
2. Re-request review if significant changes

Fail:

1. Analyze failure reasons
2. Either fix issues or escalate to user

Settings and Configuration

Reading Project Settings

Check for .claude/codex-collab.local.md in project root:

---
sandbox: read-only
language: ja
---

# Project-specific instructions

Parse YAML frontmatter for:

• model: Codex model to use (omit to use the Codex default from ~/.codex/config.toml)
• sandbox: read-only | workspace-write | danger-full-access
• workflow: Workflow mode (auto | codex-leads | claude-leads, default: auto; auto は常に codex-leads を選択)
• exchange.enabled: Enable planning exchange (default: true, codex-leads only)
• exchange.max_iterations: Maximum rounds for multi-turn exchange (default: 3)
• exchange.user_confirm: When to ask user confirmation (never | always | on_important)
• exchange.history_mode: How to handle history (full | summarize; Bash fallback only — MCP threads preserve history)
• review.enabled: Enable review iteration (default: true, codex-leads only)
• review.max_iterations: Maximum rounds for review iteration (default: 5)
• review.max_verdict_retries: Retries when verdict is missing/unclear (default: 3)
• review.user_confirm: When to ask user confirmation for reviews (default: never)
• claude_leads.sandbox: Sandbox for Codex implementation (default: workspace-write)
• claude_leads.consult_codex: Enable plan consultation phase (default: true)
• claude_leads.safety_checkpoint: Pre-implementation checkpoint (stash | wip-commit | none, default: stash)
• claude_leads.review.max_iterations: Max review-fix iterations (default: 3)
• codex.wait_timeout: Max execution time for codex exec in seconds (default: 180, max: 600; Bash fallback only)

Settings Priority

Apply settings in this order (later overrides earlier):

1. Safe defaults: sandbox=read-only
2. Global settings: ~/.claude/codex-collab.local.md
3. Project settings: .claude/codex-collab.local.md
4. Command arguments: Explicit user request

Safe Defaults

Always start with secure defaults:

• workflow: auto - 常に codex-leads を選択（claude-leads は明示指定時のみ）
• sandbox: read-only - Codex cannot modify files (codex-leads)
• exchange.enabled: true - Planning exchange enabled by default
• exchange.max_iterations: 3 - Prevent runaway exchanges
• exchange.user_confirm: on_important - Ask user for major decisions
• exchange.history_mode: summarize - Efficient token usage
• review.enabled: true - Review iteration enabled by default
• review.max_iterations: 5 - More iterations allowed (goal is clear, diff is small)
• review.user_confirm: never - Auto-iterate without confirmation
• claude_leads.sandbox: workspace-write - Codex can modify project files (claude-leads)
• claude_leads.consult_codex: true - Plan consultation enabled
• claude_leads.safety_checkpoint: stash - Git stash before implementation
• claude_leads.review.max_iterations: 3 - Claude review iterations

Quality Gates

Plan Quality Criteria

A valid plan from Codex must include:

• Clear list of files to modify
• Specific changes for each file
• Rationale for approach
• Identified risks or concerns
• Test coverage considerations

If plan is incomplete, request clarification from Codex.

Review Acceptance Criteria

Accept review as "Pass" only when:

• All changed files reviewed
• No critical bugs identified
• Security concerns addressed
• Design aligns with original plan
• Test coverage adequate

Running Codex

MCP パターン（推奨）

MCP ツールが利用可能な場合、ステートフルな通信を使用:

# 新規セッション開始
mcp__codex__codex(prompt: "...", sandbox: "read-only")
→ Returns response + threadId

# 同一スレッドで継続
mcp__codex__codex-reply(threadId: "...", prompt: "...")
→ Returns response (conversation context preserved)

codex exec パターン（フォールバック）

MCP が利用できない場合、codex exec（ステートレス実行）を使用:

# ヘルパー関数を使用（推奨）
source scripts/codex-helpers.sh
PROMPT_FILE=$(codex_write_prompt "$PROMPT_CONTENT" "plan")
OUTPUT_FILE="$(codex_tmp_path 'codex-output.md')"
codex_run_exec "$PROMPT_FILE" "$OUTPUT_FILE" "read-only"

# 直接実行（モデルは通常 Codex デフォルトを使用。指定する場合のみ -m を付ける）
codex exec -s read-only - < prompt.txt 2>&1 | tee output.md

Codex CLI Options

• -m, --model <model> - Specify model (e.g., gpt-5.6-sol; omit to use the Codex default)
• -s, --sandbox <mode> - read-only | workspace-write | danger-full-access
• -C, --cd <dir> - Working directory
• --full-auto - Automatic execution mode
• - - Read prompt from stdin

Important Notes

• MCP mode: Stateful sessions via threadId. Clean text output. No file I/O for prompts. Auto-detected in Step 0a.
• Bash mode: Each codex exec call is stateless (no conversation history between calls). Include all necessary context in each prompt.
• Use -s read-only for planning/review tasks (Codex won't modify files)
• Use -s workspace-write for implementation tasks (claude-leads workflow)
• Output may contain ANSI escape codes (Bash mode only); use codex_strip_ansi() or codex_run_exec() to clean
• Stdin input (Bash mode): Use redirect format (codex exec - < file) for reliable input
• Timeout: Bash tool has max 600s (10 minutes) timeout. MCP mode timeout is managed by MCP framework.
• Background agents: Background subagents (run_in_background: true) require pre-approved Bash permissions in ~/.claude/settings.json or .claude/settings.json. Without pre-approval, Bash tool calls are auto-denied because permission prompts are unavailable in background mode.

Error Handling

CLI Unavailable

If codex command is not found:

1. Inform user: "Codex CLI is not installed or not in PATH"
2. Offer to proceed with Claude-only mode
3. Continue with standard Claude Code workflow

Codex Timeout or Error

If codex exec returns non-zero exit code or times out:

1. Check error message in output file
2. Retry once with simplified prompt
3. If still failing, proceed manually and inform user

Bash Tool Timeout

Bash tool has max 600s (10 minutes) timeout. For long-running tasks:

1. Set appropriate codex.wait_timeout setting
2. Consider breaking tasks into smaller parts
3. Use run_in_background: true for background execution

Structured Communication Protocol

This plugin uses a minimal protocol header to enable structured communication between Claude Code and Codex CLI.

Protocol Header

Every prompt to Codex includes a ~15-line protocol header:

## Protocol (codex-collab/v1)
format: yaml
rules:
  - respond with exactly one top-level YAML mapping
  - include required fields: type, id, status, body
  - if unsure or blocked, use type=action_request with clarifying questions
types:
  task_card: {body: title, context, requirements, acceptance_criteria, proposed_steps, risks, test_considerations}
  result_report: {body: summary, changes, tests, risks, checks}
  action_request: {body: question, options, expected_response}
  review: {body: verdict, summary, findings, suggestions}
status: [ok, partial, blocked]
verdict: [pass, conditional, fail]
severity: [low, medium, high]
next_action: [continue, stop]

Message Types

Type	Purpose	Used By
task_card	Task definition with acceptance criteria	Codex (planning)
result_report	Execution results with check status	Claude (reporting)
action_request	Request for information or decision	Both
review	Review verdict and findings	Codex (review)

Parsing Strategy

• Lenient: Require only top-level envelope and core keys
• Tolerant: Accept extra fields and minor formatting differences
• Fallback: If YAML parsing fails, fall back to unstructured parsing

Multi-turn Exchange

The protocol supports two independent iteration modes:

Planning Exchange (exchange.*)

Iterative discussion during planning phase:

Flow Control:

• next_action: continue - Request further exchange
• next_action: stop - Exchange complete
• type: action_request - Implies next_action: continue

Settings:

• exchange.enabled: true - Global kill-switch
• exchange.max_iterations: 3 - Max rounds
• exchange.user_confirm: on_important - User confirmation timing
• exchange.history_mode: summarize - History management

Termination Conditions:

1. next_action: stop received
2. exchange.max_iterations reached
3. Repeated same question detected

Review Iteration (review.*)

Auto-iterate on review findings:

Flow:

1. Codex reviews → CONDITIONAL/FAIL
2. Claude fixes issues
3. Re-request review
4. Repeat until PASS or max reached

Settings:

• review.enabled: true - Enable auto-iteration
• review.max_iterations: 5 - Higher than exchange (goal is clear, diff is small)
• review.user_confirm: never - Auto-iterate without confirmation

Note: exchange.* and review.* are completely independent (no inheritance).

References

Detailed documentation in references/:

• protocol-cheatsheet.yaml - Minimal protocol header for prompts
• protocol-schema.yaml - Full protocol schema with examples
• planning-prompt.md - Template for requesting plans
• review-prompt.md - Template for requesting reviews
• codex-options.md - Codex CLI configuration options
• workflow-patterns.md - Alternative workflow patterns