Codex Integration

Sibyl uses OpenAI's Codex CLI (GPT-5.4) as an independent third-party reviewer, providing a different AI perspective alongside Claude.

When Codex Is Used

Pipeline Stage	Usage
idea_debate	Independent review of generated ideas
result_debate	Independent analysis of experiment results
review	Parallel review alongside Critic and Supervisor
writing_sections	Optional: entire paper written by GPT-5.4 (when writing_mode: codex)

Setup

1. Install Codex CLI

npm install -g @openai/codex

2. Configure ~/.codex/config.toml

model = "gpt-5.4"
model_reasoning_effort = "high"

3. Set API Key

export OPENAI_API_KEY="your-openai-api-key"

4. Register MCP Server

Preferred:

claude mcp add --scope local codex -e OPENAI_API_KEY=your-key-here -- codex mcp-server

Manual JSON fallback:

{
  "mcpServers": {
    "codex": {
      "command": "codex",
      "args": ["mcp-server"],
      "env": {
        "OPENAI_API_KEY": "your-key-here"
      }
    }
  }
}

5. Enable in Config

codex_enabled: true    # Default: false (opt-in after Codex MCP is configured)

For your own local machine, once Codex MCP and OPENAI_API_KEY are working, it is recommended to enable codex_enabled: true in your local config.yaml. That file lives in your working tree, but Git does not track or commit it, so your machine can default to Codex without changing the repo-wide default.

How It Works

1. Sibyl calls mcp__codex__codex with a review prompt
2. Codex CLI routes to GPT-5.4 (configured in config.toml)
3. The review is written to the workspace alongside Claude's review
4. The Synthesizer agent integrates both perspectives

Important: MCP calls do NOT pass a model parameter — the model configured in ~/.codex/config.toml is used. This is because ChatGPT accounts may not support API-level model override.

The approval-policy should be set to "never" for fully autonomous operation.

Disabling Codex

codex_enabled: false

When disabled:

• Idea debate and result debate run without Codex review
• Review stage uses only Critic + Supervisor
• Writing cannot use codex mode (falls back to parallel)

Server-Side Codex

When using experiment_mode: "server_codex", Codex CLI must also be installed on the GPU server:

# On GPU server
npm install -g @openai/codex

And set OPENAI_API_KEY in the server environment. Configure the path if non-standard:

server_codex_path: "/usr/local/bin/codex"