captain-codex

captain-codex

Claude Code plugin that puts Codex in charge. Codex plans, Claude implements, Codex reviews; loop until Codex is satisfied.

What It Does

One command. You describe what you want; include ad-hoc instructions for any phase in natural language.

/captain-codex refactor mac app to enable ios app with code sharing

/captain-codex refactor auth module. for planning, focus on backwards compat. when implementing, don't touch the database layer. reviewer should be strict about test coverage.

Ad-hoc instructions are merged with your configured defaults for each phase.

Why

Claude is a strong implementor; fast, creative, good across large codebases. But it reward-hacks. It takes shortcuts to look done: skips edge cases, writes tests that pass without verifying behavior, deviates from plans when compliance is hard, declares victory early. You need a separate verifier.

Codex is better at architectural reasoning; cleaner module boundaries, more principled dependency graphs. The verifier should be a different model than the implementor. Same-model review has anchoring bias; the reviewer shares the implementor's blind spots. Cross-model review catches things neither catches alone.

I've been doing this manually for a long time; copying plans from Codex to Claude, pasting output back for review, feeding feedback in, repeating. It produces noticeably better code than either model alone. But you have to babysit the whole loop. This plugin automates it.

Dependencies

• codex-plugin-cc; OpenAI's official plugin (provides the review gate hook infrastructure)
• Codex CLI installed and authenticated (codex login)
• Claude Code v2.1.34+
• jq

Installation

This plugin is available through the jul-sh Claude Code plugin marketplace.

Add the marketplace:

/plugin marketplace add jul-sh/claude-plugins

Install the plugin:

/plugin install captain-codex@jul-sh

Commands

Command	Description
/captain-codex <task>	Full pipeline: plan, implement, review loop
/captain-codex:status	Current phase, round, review history
/captain-codex:instructions	View/edit plan, implementation, and review instructions
/captain-codex:config	View/edit plugin config

Flags: --skip-plan <path>, --max-rounds <n>, --supervised.

How It Works

Planning. Codex reads the codebase and writes an implementation plan. Saved to tasks/<slug>.md. Reviews happen in the same Codex session, so Codex retains full context of the plan it wrote.

Implementation. Claude receives the plan and implements autonomously, maintaining a worklog in the plan file.

Review loop. When Claude finishes, a Stop hook resumes the Codex planning session for review. Rejected; Claude gets feedback and continues. Approved; done. Max rounds exceeded; you decide.

Supervised mode. --supervised pauses after planning and after each review round for human approval.

Configuration

Three instruction sets control what each phase does:

Config key	Controls
plan_instructions	What Codex should focus on when planning
implementation_instructions	How Claude should implement
review_instructions	What Codex should check during review

Edit via /captain-codex:instructions or directly in config files.

User-level: ~/.claude-architect/config.json Project-level override: .claude-architect/config.json

/captain-codex:config                           # view all
/captain-codex:config codex.model gpt-5.4      # set a value
/captain-codex:config max_rounds 15             # set a value

See templates/default-config.json for all options.

License

MIT