Codex Integration Skill

Purpose

This skill enables a collaborative workflow between Claude and OpenAI's Codex CLI. Send prompts to Codex, capture its output, and have Claude process the results.

How to Use

Invoke with: /codex <your prompt here>

If no prompt is provided, ask the user what they want Codex to do.

Execution Flow

Step 1: Determine the prompt

If the user provided a prompt after /codex, use it. Otherwise, ask them what task they want Codex to perform.

Step 2: Ask for processing mode

Ask the user how they want Claude to process Codex's output:

1. •Review - Analyze Codex's solution for correctness, security, performance, and best practices
2. •Compare - Have Claude solve the same problem, then compare approaches
3. •Integrate - Take Codex's output and integrate it into the codebase with Claude's refinements
4. •Raw - Just show Codex's output without additional processing

Step 3: Execute Codex

Run Codex in non-interactive mode with the user's prompt.

First, create a unique output file to avoid collisions with parallel invocations:

CODEX_OUTPUT=$(mktemp /tmp/codex-output-XXXXXX.json)

Then execute Codex:

codex exec --full-auto --json "<prompt>" 2>&1 | tee "$CODEX_OUTPUT"

Or as a single command:

CODEX_OUTPUT=$(mktemp /tmp/codex-output-XXXXXX.json) && codex exec --full-auto --json "<prompt>" 2>&1 | tee "$CODEX_OUTPUT" && echo "Output saved to: $CODEX_OUTPUT"

Options to consider:

• •--full-auto: Enables automatic execution with workspace-write sandbox
• •--json: Outputs structured JSONL for easier parsing
• •--sandbox read-only: For safer execution when reviewing only
• •-m <model>: Specify a different model (e.g., o3, o4-mini)

Step 4: Process the output based on mode

Review Mode

1. •Parse the Codex output
2. •Identify any code changes or suggestions made
3. •Analyze for:
   • •Correctness: Does it solve the stated problem?
   • •Security: Any vulnerabilities introduced?
   • •Performance: Efficiency concerns?
   • •Best practices: Follows project conventions?
   • •Edge cases: Are they handled?
4. •Provide a structured review with recommendations

Compare Mode

1. •First, read and understand Codex's solution
2. •Independently solve the same problem as Claude
3. •Compare both approaches:
   • •Similarities and differences
   • •Trade-offs of each approach
   • •Which is more maintainable/readable
   • •Which handles edge cases better
4. •Recommend which approach (or hybrid) to use

Integrate Mode

1. •Parse Codex's output for code changes
2. •Review the changes for quality and correctness
3. •Make refinements based on:
   • •Project coding standards
   • •Security best practices
   • •Performance optimizations
   • •Better naming/structure
4. •Apply the refined changes to the codebase
5. •Summarize what was changed

Raw Mode

1. •Display the Codex output directly
2. •Offer to do further processing if requested

Output Parsing

Codex --json output is JSONL format. Key event types:

• •message_output_item_added: Contains agent messages and tool calls
• •message_output_item_done: Completed items
• •response_completed: Final response

Extract code changes from code_interpreter tool calls or direct file modifications.

Error Handling

• •If Codex fails to run, check that it's installed and authenticated
• •If execution times out, suggest using --sandbox read-only or a simpler prompt
• •If output is too large, summarize key points and offer to dive into specifics

Example Invocations

• •/codex fix the type error in auth.ts
• •/codex add input validation to the user form
• •/codex refactor the database queries for better performance