Codex
npx machina-cli add skill veithly/codex-skill/codex --openclawCodex CLI Integration
Delegate coding tasks to OpenAI's Codex CLI with configurable execution modes including YOLO mode for fully autonomous operation.
Usage
/codex <task_description> [options]
Options
--mode=yolo- YOLO mode: No approvals, no sandbox (use in hardened environments only)--mode=auto- Full auto mode: Workspace write access, minimal interaction (default)--mode=safe- Safe mode: Requires approval for each action--timeout=<seconds>- Execution timeout (default: 1800)--add-dir=<path>- Additional directory for Codex to access
Examples
# Fix a bug with default auto mode
/codex Fix the null pointer exception in src/api/handlers.ts
# Implement feature with YOLO mode
/codex --mode=yolo Implement user authentication with JWT in the express app
# Refactor with safe mode
/codex --mode=safe Refactor the database connection pool to use async/await
# Run tests and fix failures
/codex Run all tests and fix any failing ones
Context
- Task description: $ARGUMENTS
- Current working directory will be used as Codex workspace
- Git repository state should be clean before execution
Your Role
You are the Codex Orchestration Coordinator managing automated coding tasks through OpenAI's Codex CLI. You handle task translation, mode selection, execution monitoring, and result integration.
Workflow
Phase 1: Environment Verification
Before executing any Codex command, verify the environment:
# Check if Codex CLI is installed
codex --version 2>/dev/null || echo "CODEX_NOT_INSTALLED"
# Check for API key (just verify it exists, don't expose)
[ -n "$OPENAI_API_KEY" ] && echo "API_KEY_SET" || echo "API_KEY_NOT_SET"
# Check git status (ensure we can rollback)
git status --porcelain
If Codex is not installed, provide installation instructions:
npm install -g @openai/codex
# or
pnpm add -g @openai/codex
Phase 2: Task Analysis
Parse the user's arguments to determine:
- Task Description - The coding task to delegate to Codex
- Execution Mode - yolo, auto, or safe (default: auto)
- Timeout - Maximum execution time (default: 1800 seconds)
- Additional Options - Extra directories, model preferences, etc.
Argument Parsing:
- Extract
--mode=<value>if present, default toauto - Extract
--timeout=<value>if present, default to1800 - Extract
--add-dir=<path>if present - Remaining text is the task description
Phase 3: Mode Selection & Command Construction
IMPORTANT: Use codex exec for non-interactive execution (not just codex).
Based on parsed arguments, construct the appropriate Codex command:
YOLO Mode (--mode=yolo)
codex exec --dangerously-bypass-approvals-and-sandbox "<task>"
WARNING: Only use in hardened environments (Docker, VM, CI/CD)
Auto Mode (--mode=auto) - Default
codex exec --full-auto "<task>"
Equivalent to: --sandbox workspace-write --ask-for-approval on-request
Safe Mode (--mode=safe)
codex exec -s workspace-write "<task>"
Uses workspace-write sandbox with default approval policy.
Phase 4: Execution
Execute the constructed command:
# Example execution pattern
codex exec --full-auto "Your task description here" 2>&1
Key Points:
- Always use
codex execfor non-interactive (automated) execution - Capture both stdout and stderr
- Monitor for timeout conditions
- The command may take several seconds to minutes depending on task complexity
Phase 5: Result Collection
After execution, gather results:
# Check what files changed
git diff --name-only
# Show summary of changes
git diff --stat
# Show detailed changes (if manageable size)
git diff
Phase 6: Report Generation
Generate a comprehensive report including:
-
Execution Summary
- Mode used
- Task sent to Codex
- Execution status (success/failure)
- Tokens used (if available from output)
-
Changes Made
- List of modified files
- Summary of changes per file
- New files created
- Files deleted
-
Codex Output
- Relevant messages from Codex
- Any warnings or errors encountered
- Model used
-
Verification Steps
- How to test the changes
- Recommended manual review areas
- Suggested follow-up actions
-
Rollback Instructions
# If changes are problematic, rollback with: git checkout -- . # or reset to last commit: git reset --hard HEAD
Error Handling
Codex Not Installed
Codex CLI is not installed.
To install, run:
npm install -g @openai/codex
Or with pnpm:
pnpm add -g @openai/codex
API Key Not Configured
OpenAI API key not found.
Set your API key:
# Windows PowerShell
$env:OPENAI_API_KEY = "your-api-key"
# Bash/Zsh
export OPENAI_API_KEY="your-api-key"
Execution Timeout
Codex execution timed out after {timeout} seconds.
Consider:
1. Breaking the task into smaller steps
2. Increasing timeout: /codex --timeout=600 <task>
3. Using a more specific task description
stdin is not a terminal
This error occurs when using `codex` directly instead of `codex exec`.
Solution: The skill uses `codex exec` for non-interactive execution.
Best Practices
- Use Git - Always ensure your workspace is a git repository
- Clean State - Commit or stash changes before running Codex
- Start Safe - Use safe mode for unfamiliar tasks, escalate as needed
- Be Specific - Provide detailed task descriptions with file paths
- Verify Results - Always review Codex's changes before committing
- Backup First - For critical code, create a branch before execution
Security Considerations
- YOLO mode bypasses all safety measures - use only in isolated environments
- API keys should never be committed to version control
- Code review is essential even for AI-generated changes
- Access scope - Prefer minimal sandbox levels when possible
Integration with Claude Workflow
This skill works seamlessly with other Claude skills:
# Use Codex for implementation, then review with Claude
/codex Implement the payment processing module
/review Check the new payment module for security issues
# Debug with Claude, fix with Codex
/debug Investigate the authentication failure
/codex --mode=safe Fix the authentication token validation bug
# Plan with Claude, implement with Codex
/code Plan the new API structure
/codex Implement the planned API endpoints
Source
git clone https://github.com/veithly/codex-skill/blob/master/skills/codex/SKILL.mdView on GitHub Overview
Automates coding tasks by delegating them to OpenAI Codex via the Codex CLI. It supports YOLO mode for fully autonomous execution, auto mode for workspace write with minimal interaction, and safe mode for action-by-action approvals. Includes timeout control and optional access to extra directories.
How This Skill Works
The skill parses the user task and options, then constructs a codex exec command based on the chosen mode (--mode=yolo|auto|safe) and flags (--timeout, --add-dir). It runs the command non-interactively, capturing stdout and stderr, and finally reports the repository changes by inspecting git status and diffs.
When to Use It
- Automatically fix a bug in a repository using default auto mode
- Implement a feature with YOLO mode in hardened environments
- Refactor code with safe mode to require per-action approvals
- Run tests and automatically fix failing tests
- Extend the workspace by providing an extra directory via --add-dir for Codex access
Quick Start
- Step 1: Verify Codex CLI installation and API key presence
- Step 2: Parse the task description and select mode/timeout/add-dir
- Step 3: Run codex exec with the appropriate flags and review git changes
Best Practices
- Verify Codex CLI installation and API key presence before first run
- Prefer auto mode for routine tasks; switch to yolo only in trusted environments
- Set a sensible timeout with --timeout to prevent long-running tasks
- Provide necessary directories with --add-dir to grant Codex access to the workspace
- Review git diffs and test outcomes after execution and rollback if needed
Example Use Cases
- Fix a bug with default auto mode
- Implement user authentication with JWT in the express app
- Refactor the database connection pool to use async/await
- Run all tests and fix any failing ones
- Add an additional directory for Codex access using --add-dir