Codex CLI — AI Coding Tool Guide

OpenAI's open-source CLI coding agent. Reasoning strategies, task decomposition, sandboxed execution, and configuration for multi-model workflows.

codex-cliopenaiclisandboxreasoning

Codex CLI

Codex CLI is OpenAI's open-source command-line coding agent. It runs in the terminal, uses OpenAI models, and supports sandboxed execution with configurable reasoning strategies for different task types.

What Makes Codex CLI Different

  • Open-source — Apache 2.0 license, inspect and modify the source
  • Reasoning strategies — Configurable approaches for planning, coding, debugging, and reviewing
  • Sandboxed execution — Run untrusted code safely in Docker or VM sandboxes
  • OpenAI ecosystem — First-class support for GPT-5 and o-series reasoning models
  • Task decomposition — Breaks complex tasks into subtasks with dependency tracking

Getting Started

npm install -g @openai/codex-cli
export OPENAI_API_KEY="sk-..."
cd my-project
codex

Prompt Patterns

Basic Prompts

codex "Add TypeScript types to src/services/api.ts and handle error responses"

Reasoning Strategy

For complex problems requiring step-by-step analysis:

codex --strategy reasoning \
  "Design a caching layer for the API. Consider:
   - Cache invalidation strategy (TTL vs event-driven)
   - What to cache (queries, computed results, rate-limited endpoints)
   - Where to cache (in-memory, Redis, edge)
   - How to handle cache misses under load"

Planning Strategy

For tasks that need architecture before implementation:

codex --strategy plan \
  "Add OAuth2 login with GitHub and Google. Plan the implementation:
   - Which library to use (NextAuth vs Clerk)
   - Database schema changes needed
   - Middleware updates for protected routes
   - Session management strategy
  Present the plan, then implement step by step."

Code Review Strategy

codex --strategy review \
  "Review src/services/payment.ts for:
   - Security: input validation, injection risks
   - Reliability: error handling, retry logic
   - Performance: N+1 queries, unnecessary recomputation
   - Compliance: PCI-DSS requirements for payment data"

Debugging Strategy

codex --strategy debug \
  "The image upload endpoint returns 500 on PNG files > 10MB.
   Trace the pipeline: multer → sharp → S3 upload.
   Find where the failure occurs and why only PNGs fail."

Configuration

Codex CLI reads ~/.codex/config.json:

{
  "model": "gpt-5",
  "reasoning": {
    "default_strategy": "auto",
    "max_steps": 20,
    "checkpoint_frequency": 5
  },
  "sandbox": {
    "enabled": true,
    "type": "docker",
    "image": "codex-sandbox:latest"
  },
  "permissions": {
    "network": "allowed",
    "filesystem_write": "allowed",
    "command_execution": "ask"
  },
  "git": {
    "auto_commit": false,
    "branch_prefix": "codex/"
  }
}

Task Dependencies

Codex CLI tracks task completion and dependencies:

codex \
  "Task 1: Create the User model in prisma/schema.prisma
   Task 2: Create the API endpoint (depends on Task 1)
   Task 3: Write integration tests (depends on Task 2)
   Task 4: Generate API docs (depends on Task 2)"

Tasks execute in dependency order. If Task 1 fails, Tasks 2-4 are skipped.

Strategy Selection

StrategyBest ForFlags
autoGeneral tasks(default)
reasoningComplex analysis, architectureStep-by-step thinking
planMulti-step featuresPlan first, implement after approval
reviewCode review, security auditAnalysis only, no changes
debugBug investigationTrace paths, identify root cause