ask-codex - SKILL.md Agent Skill

name: ask-codex description: >- Get an independent opinion from OpenAI Codex CLI (GPT 5.4 model) on a design decision, code review, or debugging hypothesis. Use when you want a second opinion from a different model, need to challenge your own approach, or want to validate a fix before committing. Triggers: "ask codex", "get codex opinion", "what does codex think", "second opinion", or when /ask-council invokes it.

Ask Codex

Get an independent opinion from the Codex CLI (codex exec) which runs OpenAI's GPT 5.4 model.

When to use

Validating a fix or design decision before committing
Getting a contrarian view from a different model
When /ask-council needs an external model's perspective

How to invoke

cd <relevant-repo-root> && codex exec --skip-git-repo-check \
  -c 'sandbox_permissions=["disk-full-read-access"]' \
  "<prompt>"

Key flags

Flag	Purpose
`--skip-git-repo-check`	Run without requiring a git repo (always include)
`-c 'sandbox_permissions=["disk-full-read-access"]'`	Let Codex read the codebase (required)
`-m <model>`	Override model (default: gpt-5.4)

Prompt guidelines

Tell Codex to read specific files — it has disk access and will read them itself
Be opinionated in framing: "Which would you ship and why?" not "What do you think?"
Include labeled options (A, B, C...) with brief pros/cons
Ask for a ranked recommendation
Keep prompt under ~500 words

Example

codex exec --skip-git-repo-check -c 'sandbox_permissions=["disk-full-read-access"]' \
  "Read src/lib/keychain.ts and src/lib/lazy-import.ts. I'm deciding between:
   A) bare import() — simple but loses install prompt
   B) { optional: true } flag on lazyImport
   C) split into lazyImport + tryImport
   Which would you ship and why?"

Resuming sessions

Each codex exec run prints a session id: <uuid> line in its output header. To capture it and resume later:

# Capture output and extract session ID
OUTPUT=$(codex exec --skip-git-repo-check -c 'sandbox_permissions=["disk-full-read-access"]' "your prompt" 2>&1)
SESSION_ID=$(echo "$OUTPUT" | grep 'session id:' | awk '{print $NF}')
echo "$OUTPUT"

# Resume that session with a follow-up
codex exec resume "$SESSION_ID" --skip-git-repo-check \
  -c 'sandbox_permissions=["disk-full-read-access"]' \
  "Follow-up prompt here"

# Or just resume the most recent session (no ID needed)
codex exec resume --last --skip-git-repo-check \
  -c 'sandbox_permissions=["disk-full-read-access"]' \
  "Follow-up prompt here"

In practice, --last is the simplest option since Claude typically resumes the session it just ran.

Caveats

Codex runs in a sandbox — it can read files but won't modify them
Allow 2-3 minutes timeout (timeout: 180000)
Use 2>&1 | tail -120 to capture output — Codex prints tool calls before its final opinion
If not installed: npm install -g @openai/codex