Codex Collab - Dual AI Collaboration (OpenAI)

Use OpenAI's Codex CLI as an independent sub-agent within Claude Code. Claude orchestrates, Codex provides independent analysis powered by OpenAI models, Claude synthesizes the final result. All output goes through files — zero context pollution.

Prerequisites

• Codex CLI installed: npm install -g @openai/codex
• Authenticated: codex login
• Verify: codex --version

Usage

/codex-collab review                # → Delegates to /codex:review (official plugin)
/codex-collab review --base main    # → Delegates to /codex:review --base main
/codex-collab understand            # Codex analyzes project structure
/codex-collab opinion <topic>       # Get Codex's take on an architecture decision
/codex-collab compare <topic>       # Both AIs answer independently, then synthesize
/codex-collab free <prompt>         # Free-form prompt to Codex

When the user doesn't use a subcommand but mentions Codex in a collaborative way, infer the most appropriate mode from context.

Review Mode — Delegated to Official Plugin

/codex-collab review now delegates to /codex:review (official OpenAI plugin). The official plugin handles Codex CLI invocation, output capture, and lifecycle management.

• /codex-collab review → invoke /codex:review
• /codex-collab review --base main → invoke /codex:review --base main
• For adversarial/security reviews → invoke /codex:adversarial-review

After receiving the official plugin's output, Claude adds its own review perspective and presents the dual-AI synthesis report (Consensus / Divergent Views) following the collaboration protocol.

codex exec — General Non-Interactive Mode

For understand/opinion/compare/free modes. Use -o (--output-last-message) to save clean output.

# Write mode (default)
codex exec "Your prompt here" --full-auto -o "${SESSION_PATH}/output.md" 2>"${SESSION_PATH}/error.log"

# Read-only analysis (no --full-auto, use -s read-only instead)
codex exec "Analyze this project" -s read-only -o "${SESSION_PATH}/output.md" 2>"${SESSION_PATH}/error.log"

Error Handling

• If codex not found: npm install -g @openai/codex
• If auth fails: codex login
• If timeout (>5min): check partial output in file
• If empty output: proceed with Claude-only analysis
• Never block the workflow on Codex failure

Reference Files

Read these when you need details beyond what's in this SKILL.md:

• references/codex-cli-reference.md — READ when you need advanced Codex CLI flags (model selection, sandbox modes, config overrides). Contains full flag reference for both codex review and codex exec.
• references/codex-prompts.md — READ when constructing Codex prompts. Contains CLI-ready prompt templates for each collaboration mode with correct command mapping (review vs exec).
• references/collaboration-modes.md — READ when you need the detailed step-by-step flow for a specific mode. Contains definitions for review/understand/opinion/compare/free modes.
• references/collab-protocol.md — READ when writing synthesis reports or managing sessions. Contains core principles, synthesis template, session management, and error handling.