@rules/routing.md
Codex Skill
Defaults
| Parameter | Default |
|---|---|
| Model | gpt-5.3-codex |
| Reasoning effort | high |
| Sandbox | read-only |
Do NOT ask the user for model or reasoning effort unless explicitly requested.
Routing
Use this skill when the request actually needs Codex CLI or a Codex session.
- Read rules/routing.md before building a command when the request might be out of scope.
- Route away to another skill or direct editing when the user wants generic writing, runbook cleanup, or skill creation without needing Codex CLI itself.
Running a Task
Read references/recipes.md for concrete command recipes before switching sandboxes or resuming a session.
Sandbox Mode Selection
| Mode | When to use |
|---|---|
read-only (default) | Analysis, review, no file changes |
workspace-write | User explicitly requests file edits |
danger-full-access | Network or broad access needed |
Command
codex exec --skip-git-repo-check \
-m gpt-5.3-codex \
--config model_reasoning_effort="high" \
--sandbox read-only \
"your prompt here" 2>/dev/nullAlways:
- Use
--skip-git-repo-check - Append
2>/dev/nullto suppress thinking tokens (show stderr only when debugging) - Confirm with user before using
--full-autoor--sandbox danger-full-access
Resuming a Session
echo "your prompt here" | codex exec --skip-git-repo-check resume --last 2>/dev/nullAll flags go between exec and resume. Omit configuration flags when resuming unless the user requests them. A resumed session inherits the original model, reasoning effort, and sandbox unless you intentionally change them.
After Completion
- Inform the user: "You can resume this Codex session by saying 'codex resume'."
- Use
AskUserQuestionto confirm next steps or whether to resume. - When output includes warnings or partial results, summarize and ask how to adjust.
Critical Evaluation
Treat Codex as a colleague, not an authority.
- Trust your knowledge when confident. Push back on incorrect claims directly.
- Research disagreements using WebSearch or docs before accepting Codex's claims.
- Remember knowledge cutoffs — Codex may not know about recent releases or API changes.
When Codex is Wrong
- State your disagreement clearly to the user
- Provide evidence (your knowledge, web search, docs)
- Optionally resume to discuss:
echo "This is Claude following up. I disagree with [X] because [evidence]. What's your take?" | codex exec --skip-git-repo-check resume --last 2>/dev/null - Frame disagreements as discussions — either AI could be wrong
- Let the user decide when there's genuine ambiguity
Error Handling
- Report failures whenever
codex --versionorcodex execexits non-zero. - When output includes warnings or partial results, summarize and ask how to adjust.