codex

name

codex

description

Invoke OpenAI Codex CLI (codex-cli) to delegate coding tasks, ask questions, or run code reviews using OpenAI models. Also handles Codex initial setup and configuration. Use when: (1) user says /codex, "codex한테 물어봐", "codex 실행", "codex로 해줘", (2) user wants a second AI opinion on code, (3) user wants to delegate a task to Codex agent, (4) user wants Codex to review code or changes, (5) user asks to set up or configure codex ("codex 세팅", "codex 설정", "codex init"). Supports model selection, sandbox modes, web search, image input, structured output, and Claude ↔ Codex bridge configuration.

Codex CLI Skill

Delegate tasks to OpenAI Codex CLI (codex exec) and capture results. Docs: https://developers.openai.com/codex/cli

For setup, configuration, and Claude↔Codex bridge details: See references/setup.md.

Pre-flight: Update & Auth Check

Before every codex invocation, run:

npm i -g @openai/codex@latest 2>&1 | tail -3 && codex login status 2>&1

Skip only if already checked in the current session.

Core Commands

Ask / Task (non-interactive)

Progress → stderr; final answer → stdout.

codex exec "<prompt>" -o /tmp/codex-out.md 2>&1

Full-auto (no approval prompts, workspace-write sandbox)

codex exec --full-auto "<prompt>" -o /tmp/codex-out.md 2>&1

Code review

codex review --uncommitted 2>&1          # all local changes
codex review --base main 2>&1            # diff against branch
codex review --commit <sha> 2>&1         # specific commit

Working directory

codex exec -C /path/to/project "<prompt>" -o /tmp/codex-out.md 2>&1

Key Flags

Flag	Description
-m, --model <name>	Override model
-o <path>	Write final message to file
-C <path>	Set working directory
-i <file,...>	Attach PNG/JPEG images
-s <mode>	read-only / workspace-write / danger-full-access
-a <policy>	untrusted / on-request / never
--full-auto	-a on-request -s workspace-write
--yolo	Skip all approvals + sandbox (dangerous)
--search	Live web search (default: cached)
--json	JSONL event stream to stdout
--ephemeral	Don't persist session
--output-schema <path>	Enforce JSON Schema on output
--skip-git-repo-check	Run outside git repos

Models

Model	Notes
gpt-5.3-codex	Default. Most capable
gpt-5.3-codex-spark	Near-instant (ChatGPT Pro only)
gpt-5.2-codex	Previous gen
gpt-5.2	General purpose

Omit -m to use config default. If "not supported" error, retry without -m.

Workflow

1. Run update + auth check (once per session).
2. Determine task type: question, code task, or review.
3. Select model if user specified; otherwise use default.
4. Run codex exec -o /tmp/codex-out.md or codex review.
5. Set Bash timeout 120000–600000ms depending on complexity.
6. Read output file and relay result to user.

Setup Tasks

When user asks to set up or configure codex, read references/setup.md and follow the relevant section:

User Request	Reference Section
"codex 설치" / "codex install"	Installation & Update
"codex 로그인" / "codex auth"	Authentication
"codex 설정" / "config 수정"	config.toml Structure
"CLAUDE.md 공유" / "context 연동"	Claude ↔ Codex Bridge Setup
"MCP 추가" / "서버 연결"	MCP Servers
"스킬 공유" / "skills 연동"	Codex Skills, Claude ↔ Codex Bridge
"프로젝트 신뢰" / "trust"	Trust & Project Config

Quick Bridge Setup (Claude ↔ Codex)

To share project context and skills between Claude and Codex:

# 1. Symlink CLAUDE.md → AGENTS.md (Codex reads AGENTS.md)
ln -s CLAUDE.md AGENTS.md

# 2. Symlink skills directory
mkdir -p .agents && ln -s ../.claude/skills .agents/skills

# 3. Add fallback in ~/.codex/config.toml
#    project_doc_fallback_filenames = ["CLAUDE.md"]

Resume / Fork Sessions

codex exec resume --last "follow-up"     # continue last
codex exec resume <SESSION_ID> "prompt"   # specific session
codex fork --last                         # fork from last

Important Notes

• Always use codex exec (non-interactive), never bare codex (launches TUI).
• Append 2>&1 to capture stderr, or 2>/dev/null for stdout only.
• For long-running tasks, use run_in_background: true on the Bash tool.
• Codex reads AGENTS.md in project root (not CLAUDE.md by default).
• CODEX_API_KEY env var for CI/headless auth.
• on-failure approval policy is deprecated → use on-request.