with-codex

Original：🇺🇸 English

Translated

2 scripts

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).

1installs

Sourceseika139/dotfiles

Added on2026-02-19

NPX Install

npx skill4agent add seika139/dotfiles with-codex

SKILL.md Content

View Translation Comparison →

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:

bash

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

bash

~/.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

bash

~/.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

bash

~/.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

bash

~/.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:

markdown

## 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)

bash

~/.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 claude
```
then
```
claude
```

Alternative: Non-Interactive Mode

Only use when tmux is unavailable or user explicitly requests:

bash

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

with-codex

NPX Install

Tags

SKILL.md Content

With Codex - Claude and Codex Collaboration

CRITICAL REQUIREMENT

Environment Requirements

Standard Workflow (MUST FOLLOW)

Step 1: Setup - Split pane and start Codex

Step 2: Claude performs its own analysis first

Step 3: Send the same prompt to Codex

Step 4: Wait for Codex response

Step 5: Capture Codex output

Step 6: Present combined results

Step 7: Cleanup (when conversation ends or user requests)

Available Commands

Error Handling

Alternative: Non-Interactive Mode

Best Practices