Claude Code: Delegate tasks to OpenAI Codex agent

codex-delegate.md

description
Delegate a coding task to OpenAI Codex agent for autonomous execution

You are delegating a coding task to a headless Codex agent. Codex will do the actual implementation work.

Your Task

1. Identify the task from conversation context
2. Read CLAUDE.md and inject context directly into the prompt
3. Execute Codex and report results

Step 1: Identify the Task from Context

Review the conversation history and identify:

• What coding task does the user want accomplished?
• What files/areas of the codebase are involved?
• What does "done" look like?

Restate the task clearly. If unclear, ask the user to clarify.

Step 2: Gather Context

1. Read CLAUDE.md if it exists - extract relevant project context
2. Use Read/Grep/Glob to find relevant code patterns
3. Note any conventions or styles to follow

Step 3: Formulate Task for Codex

Write a detailed task specification that INCLUDES the CLAUDE.md context directly:

# Project Context
[Paste relevant sections from CLAUDE.md here]

# Task
[Clear description based on conversation context]

# Relevant Files
[List key files Codex should know about]

# Requirements
- [specific requirement 1]
- [specific requirement 2]

# Success Criteria
- [how to know it's done correctly]

Do NOT:

• Tell Codex how to implement (let it decide)
• Over-constrain the solution

DO:

• Include CLAUDE.md context directly in the prompt
• Be specific about requirements
• State clear success criteria

Step 4: Execute Codex

Run Codex with your formulated task:

codex exec --full-auto "YOUR_TASK_WITH_CONTEXT_HERE" 2>&1

The --full-auto flag enables:

• --sandbox workspace-write - Can modify files in the workspace
• -a on-request - Model decides when to ask for approval

Step 5: Report Results

Delegation Report

Task Identified

[The task you identified from context]

Context Provided

[Summary of CLAUDE.md context you included]

What Codex Did

• Files created/modified
• Summary of changes

Verification

• Did it meet the success criteria?
• Any issues encountered?

Files Changed

[List with brief descriptions]

Follow-up Needed

• Any remaining tasks
• Things to review

---

Important

• YOU identify the task from conversation context
• Inject CLAUDE.md content DIRECTLY into the prompt (don't create AGENTS.md)
• If Codex fails, report what happened and suggest fixes
• You can run Codex multiple times with refined prompts