codex

name	codex
description	Consult OpenAI Codex CLI for a second opinion on plans, tests, code review, and more. Use this skill automatically whenever a second perspective would strengthen your output — especially for architectural plans, code review, test strategy, and debugging.
argument-hint	[task] [details...]
allowed-tools	Bash(codex *), Bash(codex exec *)

Runtime Scope

• Use this skill in Claude Code workflows.
• If the active assistant is Codex and this skill is referenced, do not invoke Codex CLI directly; use a subagent for the second-opinion task instead.

Consult the OpenAI Codex CLI as a second-opinion advisor. Use its output to compare, validate, or enrich your own analysis.

User request: $ARGUMENTS

Instructions

Step 1: Determine the consultation type

Based on the user's request, determine which mode applies:

• plan — Ask Codex to propose an implementation plan or architecture
• review — Ask Codex to review specific code for bugs, style, and improvements
• test — Ask Codex to suggest test cases or generate tests for specific code
• explain — Ask Codex to explain how specific code works
• debug — Ask Codex to analyze a bug or error and suggest fixes
• general — Any other consultation

Step 2: Gather context

Before calling Codex, gather relevant context so you can craft a focused prompt:

• Read the files relevant to the user's request
• Check git diff --stat or git log --oneline -10 if the request involves recent changes
• Understand what specific question or area the user wants Codex's perspective on

Step 3: Call Codex

Run Codex via codex exec in non-interactive mode. Always pass -C with the project directory.

For read-only consultation (review, explain, debug):

codex exec -s read-only "YOUR_PROMPT" -C /path/to/project

For code generation (test, plan with output):

codex exec -s workspace-write "YOUR_PROMPT" -C /path/to/project

For code review specifically:

codex exec review -C /path/to/project

CLI Reference (codex-cli v0.114.0)

Flag	Description
-s read-only	Read-only sandbox — can read files but not modify
-s workspace-write	Can write to workspace
-C /path	Set working directory
-m MODEL	Override model (e.g. -m o3)
--json	Output events as JSONL
--ephemeral	Don't persist session files
-o FILE	Write last message to file
-c key=value	Override config (e.g. -c model="o3")
--full-auto	Convenience alias for sandboxed auto-execution

Do NOT use these (they don't exist):

• --quiet — not a valid flag
• --approval-mode — not a valid flag
• Top-level codex "prompt" for non-interactive — use codex exec instead

Prompt crafting rules

• Include the specific file paths and relevant code context
• Be explicit about what you want: "Review this code for bugs", "Propose a plan to implement X", "Write unit tests for Y"
• Include constraints: "This project uses [framework/language]. Follow existing patterns."
• Keep it focused — one clear ask per invocation

If the output is too large or Codex needs to examine many files, break it into multiple focused calls.

If codex is not installed, tell the user:

Codex CLI is not installed. Install it with: npm install -g @openai/codex Make sure your OPENAI_API_KEY environment variable is set.

Then stop.

Step 4: Synthesize

Do NOT just dump Codex's raw output. Instead:

1. Present Codex's perspective — summarize its key points clearly
2. Add your own analysis — where do you agree or disagree? What did it miss? What did it catch that you wouldn't have?
3. Highlight conflicts — if Codex's suggestion contradicts the project's existing patterns or the user's preferences, flag it
4. Give a recommendation — a clear, unified recommendation combining the best of both perspectives

Format the output as:

---

Codex Consultation

Query: (what was asked)

Codex's Take

(summarized key points from Codex output)

My Analysis

(where you agree, disagree, or would extend Codex's suggestions)

Conflicts & Caveats

(any contradictions with existing codebase patterns, CLAUDE.md rules, or project conventions)

Recommendation

(unified recommendation — the best path forward considering both perspectives)

---

Step 5: Offer next steps

Ask the user if they want to:

• Act on the recommendation
• Ask Codex a follow-up question
• Get deeper analysis on a specific point

Keeping This Skill Up To Date

This skill was last verified against codex-cli v0.114.0. If Codex CLI is updated and commands fail, run codex exec --help to check current flags and update this skill accordingly. Common things that change between versions:

• Flag names and aliases
• Sandbox mode values
• Subcommands (e.g. exec review was added later)
• Default model names