Install to ~/.codex/skills/

collaborating-with-claude-code

中文 | English

A skill for Codex CLI: via a JSON bridge script, it delegates tasks such as code review / debugging / alternative implementation comparisons to Anthropic Claude Code CLI (default model: claude-opus-4-5-20251101), and returns results as structured JSON for multi-model collaboration.

The main entry points of this repository are SKILL.md (the Codex skill definition) and scripts/claude_code_bridge.py (the bridge script).

Features

Compared to other similar skills / collaboration workflows, this skill has the following key advantages:

1. It uses state-of-the-art context engineering and follows the principle of progressive disclosure, so Codex can learn how to use the skill script with a single tool call, and then invoke it correctly on the second tool call—without needing to search for or read the script again.
2. It is compatible with a wide range of Anthropic-compatible proxies that enforce strict message-structure validation when extended thinking is enabled.
   • For some Anthropic-compatible proxy APIs, when thinking is enabled, an assistant message that contains a tool-call chain must follow rules like: the assistant message must start with thinking/redacted_thinking, then tool_use, followed by the corresponding tool_result, etc. However, when using Claude Code in print mode, if such a tool-call chain is produced, the thinking content may be filtered out on the Claude Code side, causing the assistant message to start directly with tool_use. This can trigger a 400 error from the router, even though the same issue usually does not occur in the interactive Claude Code UI.
   • To address this, the bridge script in this skill adopts a strategy of splitting one long agentic loop into many short loops:
     • Each iteration allows Claude Code to perform only a single agentic turn (at most one tool call, then stop).
     • Then the bridge script automatically sends a short “continue” instruction using the same session_id to let it proceed to the next step.
   • This approach maximizes compatibility when running Claude Code via various Anthropic-compatible proxy APIs.

Install to ~/.codex/skills/

1. Choose an installation directory (create it if it doesn't exist):

mkdir -p ~/.codex/skills

2. Clone this repository into the skills directory (the folder name is the skill name):

cd ~/.codex/skills
git clone https://github.com/ZhenHuangLab/collaborating-with-claude-code.git collaborating-with-claude-code

3. Verify the file structure (it should include at least SKILL.md and scripts/):

ls -la ~/.codex/skills/collaborating-with-claude-code

4. Confirm the claude_code_bridge.py script path:

By default, it is ~/.codex/skills/collaborating-with-claude-code/scripts/claude_code_bridge.py. If it changes, update the path in SKILL.md.

Tests show that explicitly declaring the correct script path in SKILL.md makes Codex execute the bridge script more efficiently.

After that, Codex CLI can discover it when loading local skills; mention collaborating-with-claude-code (or $collaborating-with-claude-code, or a similar request in natural language) in a conversation to trigger it.

Dependencies

• Python 3 (to run the bridge script).
• Claude Code CLI installed and available (make sure claude --version works).
• Claude Code authenticated (e.g. via the environment variable ANTHROPIC_API_KEY, or any other authentication method required by your local Claude Code setup).

Note: this skill runs Claude Code in full access mode by default (non-interactive, bypassing confirmations). Only use it in directories / repositories you trust.

Run manually (without Codex CLI)

python <script_loc> --cd "/path/to/repo" --PROMPT "Review the auth flow for bypasses; propose fixes as a unified diff."

Read-only review (avoid editing files / running commands):

python <script_loc> --no-full-access --cd "/path/to/repo" --PROMPT "Review the auth flow and list issues (no code changes)."

For a more complete parameter reference and multi-turn session usage, see SKILL.md.

Runtime status output (stderr)

• While the bridge is running, it emits minimal status lines to stderr by default: working..., session_id=..., done.
• The final structured result stays on stdout only (JSON is not mixed with status text).
• Disable status lines with --quiet-status.
• Adjust heartbeat interval with --status-interval-s 30.

Compatibility

Tested on codex v0.87, claude code v2.1.11 and v2.1.12.

License

MIT License. See LICENSE.