codex-driven-development

Codex-Driven Development

A Claude Code plugin that uses Claude as an orchestrator to plan work and dispatch OpenAI Codex CLI agents to write the actual code.

Install

Step 1: Add the marketplace

/plugin marketplace add tarundevi/codex-driven-development

Step 2: Install the plugin

/plugin install codex-driven-development

Step 3: Reload plugins

/reload-plugins

Prerequisites

• Claude Code installed and configured
• Codex CLI installed and configured (npm install -g @openai/codex)

How It Works

Phase	Who	What
Planning	Claude	Writes full implementation plan with bite-sized tasks
Plan Review	Claude agent	Verifies plan is complete, codex-ready, and self-contained
Rules Injection	Claude	Reads CLAUDE.md files, extracts relevant coding rules for Codex prompts
Dispatch	Claude	Selects model/reasoning per task complexity, runs codex exec
Implement	Codex	Writes code and tests — no git commands, structured report output
Verify	Claude	Parses report, runs tests independently, checks git diff
Spec Review	Codex	Verifies implementation matches spec (PASS/FAIL)
Quality Review	Codex	Verifies code quality (PASS/FAIL)
Retry	Claude	Re-dispatches with feedback, escalates model if needed
Commit	User	All git commits are manual

Model Selection

Task Type	Model	Reasoning
Mechanical (1-2 files, clear spec)	gpt-5.4-mini	low
Standard (multi-file, moderate judgment)	gpt-5.4	medium
Complex (architectural, cross-cutting)	gpt-5.4	high
Review	gpt-5.4	medium

Retry & Escalation

Failed review → retry up to 3x at same model → escalate model → ask user.

Plan Review

Before any Codex agent is dispatched, a Claude subagent reviews the plan to verify:

• All tasks are self-contained (Codex agents can't ask questions mid-task)
• Code snippets are complete, not abbreviated
• File paths are exact, dependencies are ordered correctly
• Each task has enough context inline — no "see Task N" references

Up to 3 review iterations before escalating to the user.

Configuration

Each phase can be assigned to either Claude or Codex. The defaults above work well for most workflows — override by telling Claude at the start of a session:

Use codex-driven-development but have Claude do the code reviews instead of Codex

Phase	Default	Overridable?
Planning	Claude	No — Codex can't plan
Plan Review	Claude agent	No — requires judgment
Implementation	Codex	Yes — use Claude for MCP/web-dependent tasks
Spec Review	Codex	Yes — Claude has broader context access
Quality Review	Codex	Yes — Claude has code-reviewer skill
Retry/Escalation	Claude	No — orchestration logic

Usage

The skill shows up in your skills list once installed. You can invoke it with:

Use codex-driven-development to implement this feature

Or reference it when asking Claude to build something:

Plan and build X using Codex agents to write the code

Updating

/plugin marketplace update codex-driven-development