Codex CLI

ultrathink - Take a deep breath. We're not here to write code. We're here to make a dent in the universe.

The Vision

Codex orchestration should feel inevitable: minimal risk, maximum clarity.

Your Work, Step by Step

1. Select strategy: Model, sandbox, and reasoning effort.
2. Prepare context: Inject the smallest, sharpest prompt.
3. Execute: Run Codex with clear constraints.
4. Validate output: Check for correctness and scope compliance.
5. Summarize: Report findings and next steps.

Ultrathink Principles in Practice

• Think Different: Choose the safest path to insight.
• Obsess Over Details: Respect sandbox boundaries.
• Plan Like Da Vinci: Shape the prompt before execution.
• Craft, Don't Code: Keep commands precise.
• Iterate Relentlessly: Re-run with refined prompts.
• Simplify Ruthlessly: Reduce noise and scope.

Codex CLI Integration Skill (v2.37)

This skill enables Claude to orchestrate OpenAI's Codex CLI (v0.79+) with the gpt-5.2-codex model for code generation, review, analysis, and automated editing. Includes Context7 MCP integration for documentation access.

When to Use This Skill

Ideal Use Cases:

• Complex code analysis requiring deep understanding
• Large-scale refactoring across multiple files
• Automated code generation with safety controls
• Second opinion / cross-validation on code implementations
• Parallel processing of independent code tasks
• Session-based iterative development workflows

Quick Start

Prerequisites

Verify Codex CLI installation:

codex --version  # Should show v0.50.0+

Authentication (first time):

codex  # Interactive login via ChatGPT account
# Or: export CODEX_API_KEY=sk-...

Model Selection

Default model: gpt-5.2-codex - Optimized for software engineering tasks

Sandbox Modes

Core Commands

Basic Execution

# Read-only analysis (default)
codex exec -m gpt-5.2-codex "analyze src/auth for security issues"

# Code editing (workspace-write)
codex exec -m gpt-5.2-codex --full-auto "fix bug in login.py"

# With reasoning effort
codex exec -m gpt-5.2-codex --config model_reasoning_effort=high "complex analysis"

# Skip git check (non-git directories)
codex exec --skip-git-repo-check "analyze code"

Suppress Thinking Tokens

Add 2>/dev/null to suppress stderr (thinking tokens):

codex exec -m gpt-5.2-codex "review code" 2>/dev/null

Session Resume

# Resume last session (stdin for prompt - required due to CLI bug)
echo "continue with fixes" | codex exec resume --last 2>/dev/null

# Resume with full-auto
echo "apply fixes" | codex exec resume --last --full-auto 2>/dev/null

# Resume specific session
echo "follow-up" | codex exec resume SESSION_ID

Important: Resume inherits model, reasoning, and sandbox from original session.

JSON Output

# JSON Lines output
codex exec --json -m gpt-5.2-codex "analyze code" > output.jsonl

# Extract session ID
SID=$(grep -o '"thread_id":"[^"]*"' output.jsonl | head -1 | cut -d'"' -f4)

# Extract agent message
grep '"type":"agent_message"' output.jsonl | jq -r '.item.text'

Orchestration Patterns

Pattern 1: Context Pre-injection

Claude collects information first, injects into prompt for faster execution:

# Collect errors
ERRORS=$(npm run lint 2>&1 | grep error)

# Inject context
codex exec -m gpt-5.2-codex --full-auto "Fix these errors:
$ERRORS

Files: src/auth/login.ts, src/utils/token.ts
Constraint: Only modify listed files."

Pattern 2: Session Reuse

Related tasks reuse sessions for context preservation:

# First: analyze
codex exec -m gpt-5.2-codex "analyze src/auth for issues"

# Continue: fix (reuses context)
echo "fix the issues you found" | codex exec resume --last --full-auto

When to reuse:

• Analyze → Fix (knows findings)
• Implement → Test (knows implementation)
• Test → Fix (knows failures)

Pattern 3: Parallel Execution

Independent tasks run simultaneously:

# Parallel analysis
codex exec --json -m gpt-5.2-codex "analyze auth" > auth.jsonl 2>&1 &
codex exec --json -m gpt-5.2-codex "analyze api" > api.jsonl 2>&1 &
wait

# Parallel fixes with resume
AUTH_SID=$(grep -o '"thread_id":"[^"]*"' auth.jsonl | head -1 | cut -d'"' -f4)
echo "fix issues" | codex exec resume $AUTH_SID --full-auto &
# ...
wait

Parallelizable:

• Different directories/modules
• Different analysis dimensions (security/performance/quality)
• Read-only operations

Must serialize:

• Writing same files
• Dependent on prior results

Interactive Workflow

Before running Codex tasks, confirm with user:

1. Model selection: gpt-5.2-codex or gpt-5.2?
2. Reasoning effort: low, medium, or high?
3. Sandbox mode: Based on task requirements

Decision Matrix

Code Review Workflow

Independent Review

Use Codex as second opinion on Claude's work:

codex exec -m gpt-5.2-codex --sandbox read-only "Review src/payment/processor.py for:
1. Race conditions in transaction processing
2. Proper error handling and rollback
3. Security issues with payment data
4. Edge cases that could cause data loss
Provide specific line numbers and severity ratings."

Comprehensive Review

# Security audit
codex exec -m gpt-5.2-codex --sandbox read-only --config model_reasoning_effort=high \
  "Perform security audit of src/auth. Check for:
  - Authentication/authorization issues
  - Input validation vulnerabilities
  - Cryptographic weaknesses
  - Sensitive data exposure"

# Performance review
codex exec -m gpt-5.2-codex --sandbox read-only \
  "Analyze src/database for performance:
  - N+1 query problems
  - Missing indexes
  - Blocking operations"

Pull Request Review

codex exec -m gpt-5.2-codex --sandbox read-only \
  "Run 'git diff main...HEAD' to see changes.
  Review for:
  1. Breaking changes
  2. Performance implications
  3. Test coverage
  4. Security concerns
  Provide feedback by file with severity levels."

Prompt Design

Structure Formula

[Verb] + [Scope] + [Requirements] + [Output Format] + [Constraints]

Verb Selection

Examples

Bad vs Good:

# Bad: vague
codex exec "review code"

# Good: specific
codex exec -m gpt-5.2-codex --sandbox read-only \
  "Review src/auth for SQL injection, XSS.
  Output: markdown with severity levels.
  Format: file:line, description, fix suggestion."

Parallel Prompt Consistency

# Consistent structure for aggregation
FORMAT="Output JSON: {category, items: [{file, line, description}]}"

codex exec -m gpt-5.2-codex "review security. $FORMAT" &
codex exec -m gpt-5.2-codex "review performance. $FORMAT" &
codex exec -m gpt-5.2-codex "review quality. $FORMAT" &
wait

Claude-Codex Engineering Loop

Dual-AI Workflow

1. Claude plans → Architecture, requirements
2. Codex validates plan → Check logic, edge cases
3. Claude implements → Write code with tools
4. Codex reviews → Bug detection, security
5. Claude fixes → Apply corrections
6. Codex re-validates → Confirm quality
7. Repeat until standards met

Implementation

# Phase 2: Codex validates Claude's plan
echo "Review this implementation plan for issues:
[Claude's plan here]

Check for:
- Logic errors
- Missing edge cases
- Architecture flaws
- Security concerns" | codex exec -m gpt-5.2-codex --sandbox read-only

# Phase 4: Codex reviews Claude's code
codex exec -m gpt-5.2-codex --sandbox read-only \
  "Review implementation in src/feature for:
  - Bugs
  - Performance issues
  - Best practices
  - Security vulnerabilities"

Error Handling

1. Non-zero exit: Stop and report, ask for direction
2. Warnings: Summarize and ask how to proceed
3. High-impact flags: Ask permission before --full-auto, --sandbox danger-full-access

Post-Task Follow-up

After every Codex command:

1. Summarize outcome
2. Confirm next steps with user
3. Offer: "Resume session with 'codex resume' for continued analysis"

Configuration

Profile Setup (~/.codex/config.toml)

model = "gpt-5.2-codex"
model_reasoning_effort = "medium"

[profiles.review]
model = "gpt-5.2-codex"
model_reasoning_effort = "high"
sandbox_mode = "read-only"

[profiles.implement]
model = "gpt-5.2-codex"
model_reasoning_effort = "high"
sandbox_mode = "workspace-write"

Usage:

codex exec --profile review "analyze code"

Quick Reference

Documentation Access via Context7 MCP

Both Claude and Codex have Context7 MCP configured. Use it to access OpenAI documentation:

Available Documentation Libraries

Query Documentation Before Execution

# Before running complex Codex commands, verify syntax
mcp__context7__query-docs:
  libraryId: "/websites/developers_openai_codex"
  query: "exec sandbox modes full-auto workspace-write"

# On errors, look up solutions
mcp__context7__query-docs:
  libraryId: "/websites/developers_openai_codex"
  query: "error troubleshooting session resume"

MCP Server Configuration Reference

# ~/.codex/config.toml - Codex MCP configuration

# STDIO server (local command)
[mcp_servers.context7]
command = "npx"
args = ["-y", "@upstash/context7-mcp@latest"]

# Remote HTTP server
[mcp_servers.remote]
url = "https://example.com/mcp"
bearer_token_env_var = "API_TOKEN"

# With environment variables
[mcp_servers.server.env]
API_KEY = "value"

Verify MCP Servers

# List configured MCP servers
codex mcp list

# Add new MCP server
codex mcp add context7 -- npx -y @upstash/context7-mcp

# Test MCP server
npx @modelcontextprotocol/inspector codex mcp-server

See Also

• /openai-docs - OpenAI documentation access skill
• references/cli_reference.md - Complete CLI arguments
• references/prompt_patterns.md - Advanced prompt design
• references/parallel_execution.md - Parallel orchestration details