Bridge OpenAI Codex CLI into Claude Code for cross-model review, verification, task delegation and collaborative generation.
- Codex CLI installed:
npm install -g @openai/codex - OpenAI API key configured:
codex login - Working inside a git repository
Send a prompt to Codex CLI in read-only sandbox mode and return the result.
/codex-bridge:codex "Analyze the error handling in src/api/handler.go"
Run Codex code review on current changes or a specific branch/commit.
/codex-bridge:codex-review # Review uncommitted changes
/codex-bridge:codex-review main # Diff against main branch
/codex-bridge:codex-review abc1234 # Review specific commit
/codex-bridge:codex-review "Focus on security" # Review with custom focus
The Codex Bridge skill triggers when the user explicitly mentions "codex" — e.g., "ask codex", "let codex review", "用 codex 分析", etc. The skill routes to the appropriate command, which automatically collects conversation context and project state to compose a self-contained prompt for Codex — so vague references like "that function" or "the file we discussed" are resolved before reaching Codex.
| Pattern | Description | Sandbox |
|---|---|---|
| Code Review | Run codex review on changes |
read-only |
| Second Opinion | Get Codex's take on a design decision | read-only |
| Task Delegation | Delegate implementation to Codex | workspace-write (requires approval) |
| Collaborative Generation | Claude designs, Codex implements | workspace-write (requires approval) |
| Document Review | Have Codex review specs/docs | read-only |
See usage-patterns.md for detailed prompt templates.
- Default sandbox mode is
read-only— no file writes workspace-writerequires manual user approval- Dangerous flags (
--dangerously-bypass-approvals-and-sandbox,--full-auto,--sandbox danger-full-access) are explicitly forbidden - Users should avoid passing secrets (API keys, tokens, passwords) in Codex prompts
MIT