codex

codex — external second opinion

Summon a fresh Codex CLI (OpenAI) as an independent outside model to challenge an idea, design, plan, or diff before acting on it. Use when the user writes /codex "...", asks to involve codex, or whenever you want a second opinion that isn't anchored to your own reasoning.

Prerequisites

• codex CLI installed and authenticated (codex --version).
• The default model is whatever the user's Codex config selects (e.g. gpt-5.5, xhigh reasoning) — strong for review. Do not override the model unless asked.

Default command

Run from the relevant workspace root. Capture the clean final answer with -o; stdout also carries hook noise and a token-usage line, so never parse stdout for the answer:

codex exec -s read-only -o /tmp/codex-<tag>.md "<prompt>" > /tmp/codex-<tag>.log 2>&1

• Pick a unique <tag> per call (a short topic slug).
• -s read-only lets codex read the repo to ground its opinion but never modifies files — the right default for a review / second-opinion pass.
• --skip-git-repo-check if the workspace is not a git repo.
• -C <dir> to root codex somewhere other than the current directory.
• -m <model> to override the model — only on request.

Execution (background + wait)

Codex with high reasoning takes minutes on a real review, so run it in the background and wait for it to finish — don't block the turn:

• Claude Code: launch via the Bash tool with run_in_background: true; you're re-invoked on exit. Then read /tmp/codex-<tag>.md for the verdict, and /tmp/codex-<tag>.log if it failed (non-zero exit or empty answer file).
• Other agents / plain shell: background it (... &) and wait for the process to exit, or run in the foreground with a generous timeout (codex is non-interactive and won't prompt).

Never treat streamed/partial stdout as the final answer — wait for exit and read the -o file.

Prompt shape

Give codex enough to form an independent view:

• the workspace path and task context;
• the idea / design / plan / diff you want challenged, stated plainly;
• the exact question and focus areas (correctness, regressions, hidden assumptions, simpler alternatives, infra/production risk, tests);
• the output shape you want: verdict, where it agrees/disagrees, blocking vs non-blocking issues, concrete fixes.

Ask codex to challenge the idea and surface flaws and alternatives — the goal is an outside perspective, not validation. Default to read-only; use -s workspace-write only if the user explicitly wants codex to implement changes.

After it answers

Reconcile codex's response against the actual code and facts before acting — it's an external opinion, not ground truth. Summarize the verdict and state plainly which points you agree with, which you reject, and why.