with-codex
Enables Claude Code to collaborate with OpenAI Codex CLI on Windows. Use this skill when the user wants to get a second opinion from Codex, compare approaches between Claude and Codex, or leverage both AI assistants for collaborative problem-solving. This skill supports both non-interactive mode (automatic response retrieval) and interactive mode (visual pane splitting with tmux).
SKILL.md
| Name | with-codex |
| Description | Enables Claude Code to collaborate with OpenAI Codex CLI on Windows. Use this skill when the user wants to get a second opinion from Codex, compare approaches between Claude and Codex, or leverage both AI assistants for collaborative problem-solving. This skill supports both non-interactive mode (automatic response retrieval) and interactive mode (visual pane splitting with tmux). |
name: with-codex description: Enables Claude Code to collaborate with OpenAI Codex CLI on Windows. Use this skill when the user wants to get a second opinion from Codex, compare approaches between Claude and Codex, or leverage both AI assistants for collaborative problem-solving. This skill supports both non-interactive mode (automatic response retrieval) and interactive mode (visual pane splitting with tmux).
With Codex - Claude and Codex Collaboration
This skill enables collaboration between Claude Code and OpenAI Codex CLI for second opinions, validation, and collaborative problem-solving.
CRITICAL REQUIREMENT
Claude Code MUST be running inside a tmux session for this skill to work.
If not in tmux, inform the user: "このスキルを使用するには、tmuxセッション内でClaude Codeを起動する必要があります。以下のコマンドを実行してください:
tmux new-session -s claude
claude
"
Environment Requirements
- WSL (Ubuntu) with tmux installed
- OpenAI Codex CLI installed and authenticated in WSL
- Claude Code running inside a tmux session
- Skill scripts at:
~/.claude/skills/with-codex/scripts/
Standard Workflow (MUST FOLLOW)
When this skill is triggered, ALWAYS execute these steps in order:
Step 1: Setup - Split pane and start Codex
~/.claude/skills/with-codex/scripts/codex-manager.sh setup
This splits the current tmux pane:
- Left pane: Claude Code (current)
- Right pane: Codex CLI (newly created, with dark background)
Step 2: Claude performs its own analysis first
Analyze the user's request independently before querying Codex.
Step 3: Send the same prompt to Codex
~/.claude/skills/with-codex/scripts/codex-manager.sh send "YOUR_PROMPT_HERE"
Replace YOUR_PROMPT_HERE with the actual question/task from the user.
Step 4: Wait for Codex response
~/.claude/skills/with-codex/scripts/codex-manager.sh wait 60
Wait up to 60 seconds for Codex to complete its response.
Step 5: Capture Codex output
~/.claude/skills/with-codex/scripts/codex-manager.sh capture 200
Capture the last 200 lines of Codex's output.
Step 6: Present combined results
Present results in this format:
## Claude's Analysis
[Your independent analysis]
## Codex's Analysis
[Captured response from Codex]
## Synthesis
- **Agreement**: [Points where both AIs agree]
- **Differences**: [Alternative perspectives from Codex]
- **Recommendation**: [Best combined approach]
Step 7: Cleanup (when conversation ends or user requests)
~/.claude/skills/with-codex/scripts/codex-manager.sh cleanup
Available Commands
| Command | Description |
|---|---|
setup | Split pane and start Codex on the right |
send "prompt" | Send prompt to Codex pane |
capture [lines] | Capture Codex output (default: 100 lines) |
wait [timeout] | Wait for response to stabilize (default: 60s) |
cleanup | Close the Codex pane |
status | Check pane status |
focus | Switch focus to Codex pane |
Error Handling
If setup fails with "Not running inside tmux":
- Inform user they need to start Claude Code inside tmux
- Provide the commands:
tmux new-session -s claudethenclaude
Alternative: Non-Interactive Mode
Only use when tmux is unavailable or user explicitly requests:
codex exec "your prompt" 2>/dev/null
Best Practices
- ALWAYS use tmux interactive mode by default
- Let user see both AIs working side-by-side
- Wait adequate time (30-60s) for complex Codex queries
- Present both perspectives without bias
- Acknowledge both AIs can be wrong - user makes final decision