Claude Code Plugins

Community-maintained marketplace

Feedback

Executes OpenAI Codex CLI for code analysis, refactoring, and automated editing. Activates when users mention codex commands, code review requests, or automated code transformations requiring advanced reasoning models.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name codex
description Executes OpenAI Codex CLI for code analysis, refactoring, and automated editing. Activates when users mention codex commands, code review requests, or automated code transformations requiring advanced reasoning models.

Codex Execution Skill

Prerequisites

  • Codex CLI installed and configured (~/.codex/config.toml)
  • Verify availability: codex --version on first use per session

Workflow Checklist

For every Codex task, follow this sequence:

  1. Detect HPC/Slurm environment:

    • Check if running on HPC cluster (look for /home/woody/, /home/hpc/, Slurm env vars)
    • If HPC detected: Always use --yolo flag to bypass Landlock sandbox restrictions
  2. Ask user for execution parameters via AskUserQuestion (single prompt):

    • Model: gpt-5, gpt-5-codex, or default
    • Reasoning effort: minimal, low, medium, high
  3. Determine sandbox mode based on task:

    • read-only: Code review, analysis, documentation
    • workspace-write: Code modifications, file creation
    • danger-full-access: System operations, network access
    • HPC override: Always add --yolo flag (bypasses Landlock restrictions)
  4. Build command with required flags:

    codex exec [OPTIONS] "PROMPT"
    

    Essential flags:

    • -m <MODEL> (if overriding default)
    • -c model_reasoning_effort="<LEVEL>"
    • -s <SANDBOX_MODE> (skip on HPC)
    • --skip-git-repo-check (if outside git repo)
    • -C <DIRECTORY> (if changing workspace)
    • --full-auto (for non-interactive execution, cannot be used with --yolo)

    HPC command pattern (with --yolo to bypass Landlock):

    codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check \
      "Analyze this code: $(cat /path/to/file.py)" 2>/dev/null
    

    Note: --yolo is an alias for --dangerously-bypass-approvals-and-sandbox and is REQUIRED on HPC clusters to avoid Landlock sandbox errors. Do not use --full-auto with --yolo as they are incompatible.

  5. Execute with stderr suppression:

    • Append 2>/dev/null to hide thinking tokens
    • Remove only if user requests verbose output or debugging
  6. Validate execution:

    • Check exit code (0 = success)
    • Summarize output for user
    • Report errors with actionable solutions
    • If Landlock/sandbox errors on HPC: verify --yolo flag was used, retry if missing
  7. Inform about resume capability:

    • "Resume this session anytime: codex resume"

Command Patterns

🔥 HPC QUICK TIP: On HPC clusters (e.g., /home/woody/, /home/hpc/), ALWAYS add --yolo flag to avoid Landlock sandbox errors. Example: codex exec --yolo -m gpt-5 ...

Read-Only Analysis

codex exec -m gpt-5 -c model_reasoning_effort="medium" -s read-only \
  --skip-git-repo-check --full-auto "review @file.py for security issues" 2>/dev/null

Stdin Input (bypasses sandbox file restrictions)

cat file.py | codex exec -m gpt-5 -c model_reasoning_effort="low" \
  --skip-git-repo-check --full-auto - 2>/dev/null

Note: Stdin with - flag may not be supported in all Codex CLI versions.

HPC/Slurm Environment (YOLO Mode - Bypass Landlock)

When running on HPC clusters with Landlock security restrictions, use the --yolo flag:

# Primary solution: --yolo flag bypasses Landlock sandbox
codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check \
  "Analyze this code: $(cat /path/to/file.py)" 2>/dev/null

Alternative: Manual Code Injection (if --yolo is unavailable):

# Capture code content and pass directly in prompt
codex exec -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check --full-auto \
  "Analyze this Python code: $(cat file.py)" 2>/dev/null

Or for large files, use heredoc:

codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check "$(cat <<'ENDCODE'
Analyze the following code comprehensively:

$(cat file.py)

Focus on: architecture, algorithms, multi-GPU optimization, potential bugs, code quality.
ENDCODE
)" 2>/dev/null

Note: --yolo is short for --dangerously-bypass-approvals-and-sandbox and is safe on HPC login nodes where you have limited permissions anyway. Do not combine --yolo with --full-auto as they are incompatible.

Code Modification

codex exec -m gpt-5 -c model_reasoning_effort="high" -s workspace-write \
  --skip-git-repo-check --full-auto "refactor @module.py to async/await" 2>/dev/null

Resume Session

echo "fix the remaining issues" | codex exec --skip-git-repo-check resume --last 2>/dev/null

Cross-Directory Execution

codex exec -C /path/to/project -m gpt-5 -c model_reasoning_effort="medium" \
  -s read-only --skip-git-repo-check --full-auto "analyze architecture" 2>/dev/null

Using Profiles

codex exec --profile production -c model_reasoning_effort="high" \
  --full-auto "optimize performance in @app.py" 2>/dev/null

CLI Reference

Core Flags

Flag Values When to Use
-m, --model gpt-5, gpt-5-codex Override default model
-c, --config key=value Runtime config override (repeatable)
-s, --sandbox read-only, workspace-write, danger-full-access Set execution permissions
--yolo flag REQUIRED on HPC - Bypasses all sandbox restrictions (alias for --dangerously-bypass-approvals-and-sandbox). Cannot be used with --full-auto
-C, --cd path Change workspace directory
--skip-git-repo-check flag Allow execution outside git repos
--full-auto flag Non-interactive mode (workspace-write + approvals on failure). Cannot be used with --yolo
-p, --profile string Load configuration profile from config.toml
--json flag JSON event output (CI/CD pipelines)
-o, --output-last-message path Write final message to file
-i, --image path[,path...] Attach images (repeatable or comma-separated)
--oss flag Use local open-source model (requires Ollama)

Configuration Options

Model Reasoning Effort (-c model_reasoning_effort="<LEVEL>"):

  • minimal: Quick tasks, simple queries
  • low: Standard operations, routine refactoring
  • medium: Complex analysis, architectural decisions (default)
  • high: Critical code, security audits, complex algorithms

Model Verbosity (-c model_verbosity="<LEVEL>"):

  • low: Minimal output
  • medium: Balanced detail (default)
  • high: Verbose explanations

Approval Prompts (-c approvals="<WHEN>"):

  • on-request: Before any tool use
  • on-failure: Only on errors (default for --full-auto)
  • untrusted: Minimal prompts
  • never: No interruptions (use with caution)

Configuration Management

Config File Location

~/.codex/config.toml

Runtime Overrides

# Override single setting
codex exec -c model="gpt-5" "task"

# Override multiple settings
codex exec -c model="gpt-5" -c model_reasoning_effort="high" "task"

Using Profiles

Define in config.toml:

[profiles.research]
model = "gpt-5"
model_reasoning_effort = "high"
sandbox = "read-only"

[profiles.development]
model = "gpt-5-codex"
sandbox = "workspace-write"

Use with:

codex exec --profile research "analyze codebase"

Resume Behavior

Automatic inheritance:

  • Model selection
  • Reasoning effort
  • Sandbox mode
  • Configuration overrides

Resume syntax:

# Resume last session
codex exec resume --last

# Resume with new prompt
codex exec resume --last "continue with next steps"

# Resume via stdin
echo "new instructions" | codex exec resume --last 2>/dev/null

# Resume specific session
codex exec resume <SESSION_ID> "follow-up task"

Flag injection (between exec and resume):

# Change reasoning effort for resumed session
codex exec -c model_reasoning_effort="high" resume --last

Error Handling

Validation Loop

  1. Execute command
  2. Check exit code (non-zero = failure)
  3. Report error with context
  4. Ask user for direction via AskUserQuestion
  5. Retry with adjustments or escalate

Permission Requests

Before using high-impact flags, request user approval via AskUserQuestion:

  • --full-auto: Automated execution
  • -s danger-full-access: System-wide access
  • --yolo / --dangerously-bypass-approvals-and-sandbox:
    • HPC clusters: No approval needed (required for operation)
    • Personal machines: Request approval (full system access)

Partial Success Handling

When output contains warnings:

  1. Summarize successful operations
  2. Detail failures with context
  3. Use AskUserQuestion to determine next steps
  4. Propose specific adjustments

Troubleshooting

File Access Blocked

Symptom: "shell is blocked by the sandbox" or permission errors

Root cause: Sandbox read-only mode restricts file system

Solutions (priority order):

  1. Stdin piping (recommended):

    cat target.py | codex exec -m gpt-5 -c model_reasoning_effort="medium" \
      --skip-git-repo-check --full-auto - 2>/dev/null
    
  2. Explicit permissions:

    codex exec -m gpt-5 -s read-only \
      -c 'sandbox_permissions=["disk-full-read-access"]' \
      --skip-git-repo-check --full-auto "@file.py" 2>/dev/null
    
  3. Upgrade sandbox:

    codex exec -m gpt-5 -s workspace-write \
      --skip-git-repo-check --full-auto "review @file.py" 2>/dev/null
    

Invalid Flag Errors

Symptom: "unexpected argument '--add-dir' found"

Cause: Flag does not exist in Codex CLI

Solution: Use -C <DIR> to change directory:

codex exec -C /target/dir -m gpt-5 --skip-git-repo-check \
  --full-auto "task" 2>/dev/null

Exit Code Failures

Symptom: Non-zero exit without clear message

Diagnostic steps:

  1. Remove 2>/dev/null to see full stderr
  2. Verify installation: codex --version
  3. Check configuration: cat ~/.codex/config.toml
  4. Test minimal command: codex exec -m gpt-5 "hello world"
  5. Verify model access: codex exec --model gpt-5 "test"

Model Unavailable

Symptom: "model not found" or authentication errors

Solutions:

  1. Check configured model: grep model ~/.codex/config.toml
  2. Verify API access: Ensure valid credentials
  3. Try alternative model: -m gpt-5-codex
  4. Use OSS fallback: --oss (requires Ollama)

Session Resume Fails

Symptom: Cannot resume previous session

Diagnostic steps:

  1. List recent sessions: codex history
  2. Verify session ID format
  3. Try --last flag instead of specific ID
  4. Check if session expired or was cleaned up

HPC/Slurm Sandbox Failures

Symptom: "Landlock sandbox error", "LandlockRestrict", or all file operations fail

Root Cause: HPC clusters use Landlock/seccomp kernel security modules that block Codex's default sandbox

✅ SOLUTION: Use the --yolo flag (priority order):

  1. YOLO Flag (PRIMARY SOLUTION - WORKS ON HPC):

    # Bypasses Landlock restrictions completely
    codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check \
      "Analyze this code: $(cat /full/path/to/file.py)" 2>/dev/null
    

    Why this works: --yolo (alias for --dangerously-bypass-approvals-and-sandbox) disables the Codex sandbox entirely, allowing direct file access on HPC systems. Note: Do not use --full-auto with --yolo as they are incompatible.

  2. Manual Code Injection (fallback if --yolo unavailable):

    # Pass code directly in prompt via command substitution
    codex exec -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check --full-auto \
      "Analyze this code comprehensively: $(cat /full/path/to/file.py)" 2>/dev/null
    
  3. Heredoc for Long Code:

    codex exec --yolo -m gpt-5 -c model_reasoning_effort="high" --skip-git-repo-check "$(cat <<'EOF'
    Analyze the following Python code for architecture, bugs, and optimization opportunities:
    
    $(cat /home/user/script.py)
    
    Provide technical depth with actionable insights.
    EOF
    )" 2>/dev/null
    
  4. Run on Login Node (if compute node blocks outbound):

    # SSH to login node first, then run codex there (not in Slurm job)
    ssh login.cluster.edu
    codex exec --yolo -m gpt-5 --skip-git-repo-check "analyze @file.py" 2>/dev/null
    
  5. Use Apptainer/Singularity (if cluster supports):

    # Build image with Codex installed, then run via Slurm
    singularity exec codex.sif codex exec --yolo -m gpt-5 "task"
    

Best Practice for HPC:

  • Always use --yolo flag on HPC clusters - it's safe on login nodes where you already have limited permissions
  • Run analysis on login nodes, submit only heavy compute jobs to Slurm
  • Keep code files on shared filesystem readable from login nodes
  • Combine --yolo with $(cat file.py) for maximum compatibility

Best Practices

Reasoning Effort Selection

  • minimal: Syntax fixes, simple renaming
  • low: Standard refactoring, basic analysis
  • medium: Complex refactoring, architecture review
  • high: Security audits, algorithm optimization, critical bugs

Sandbox Mode Selection

  • read-only: Default for any analysis or review
  • workspace-write: File modifications only
  • danger-full-access: Network operations, system commands (rare)

Stderr Suppression

  • Always use 2>/dev/null unless:
    • User explicitly requests thinking tokens
    • Debugging failed commands
    • Troubleshooting configuration issues

Profile Usage

Create profiles for common workflows:

  • review: High reasoning, read-only
  • refactor: Medium reasoning, workspace-write
  • quick: Low reasoning, read-only
  • security: High reasoning, workspace-write

Stdin vs File Reference

  • Stdin: Single file analysis, avoids permissions
  • File reference: Multi-file context, codebase-wide changes

Safety Guidelines

HPC Clusters - --yolo is SAFE and REQUIRED:

  • HPC login nodes already have strict permissions (no root access, no network modification)
  • --yolo bypasses Codex sandbox but you still operate within HPC user restrictions
  • Always use --yolo on HPC to avoid Landlock errors

General Use - Exercise Caution:

  • Don't use --yolo on unrestricted systems (your laptop, cloud VMs with full sudo)
  • Prefer --full-auto + -s workspace-write for normal development

Always verify before:

  • Using danger-full-access sandbox (outside HPC)
  • Disabling approval prompts on production systems
  • Running with --yolo on personal machines with sudo access

Ask user approval for:

  • First-time workspace-write usage
  • System-wide access requests
  • Destructive operations (deletions, migrations)

Advanced Usage

CI/CD Integration

codex exec --json -o result.txt -m gpt-5 \
  -c model_reasoning_effort="medium" \
  --skip-git-repo-check --full-auto \
  "run security audit on changed files" 2>/dev/null

Batch Processing

for file in *.py; do
  cat "$file" | codex exec -m gpt-5 -c model_reasoning_effort="low" \
    --skip-git-repo-check --full-auto "lint and format" - 2>/dev/null
done

Multi-Step Workflows

# Step 1: Analysis
codex exec -m gpt-5 -c model_reasoning_effort="high" -s read-only \
  --full-auto "analyze @codebase for architectural issues" 2>/dev/null

# Step 2: Resume with changes
echo "implement suggested refactoring" | \
  codex exec -s workspace-write resume --last 2>/dev/null

When to Escalate

If errors persist after troubleshooting:

  1. Check documentation:

    WebFetch https://developers.openai.com/codex/cli/reference
    
    WebFetch https://developers.openai.com/codex/local-config#cli
    
  2. Report to user:

    • Error message verbatim
    • Attempted solutions
    • Configuration details
    • Exit codes and stderr output
  3. Request guidance:

    • Alternative approaches
    • Configuration adjustments
    • Manual intervention points

Model Selection Guide

Task Type Recommended Model Reasoning Effort
Quick syntax fixes gpt-5 minimal
Code review gpt-5 medium
Refactoring gpt-5-codex medium
Architecture analysis gpt-5 high
Security audit gpt-5 high
Algorithm optimization gpt-5-codex high
Documentation generation gpt-5 low

Common Workflows

Code Review Workflow

  1. Ask user: model + reasoning effort
  2. Run read-only analysis
  3. Present findings
  4. If changes needed: resume with workspace-write
  5. Validate changes
  6. Inform about resume capability

Refactoring Workflow

  1. Ask user: model + reasoning effort
  2. Analyze current code (read-only)
  3. Propose changes
  4. Get user approval
  5. Apply changes (workspace-write)
  6. Run validation/tests
  7. Report results

Security Audit Workflow

  1. Use high reasoning effort
  2. Run comprehensive analysis (read-only)
  3. Document findings
  4. Propose fixes
  5. Apply fixes if approved (workspace-write)
  6. Re-audit to verify
  7. Generate report