Second Opinion

Second Opinion — OpenAI Codex

Get an independent second opinion from OpenAI's Codex CLI. It uses Codex's default model — a completely different architecture and training run from Claude, which makes it genuinely useful for catching blind spots, validating reasoning on edge cases, or surfacing trade-offs Claude would miss (Claude's own errors tend to be invisible to Claude).

Codex has full read access to the project — it can read files, grep code, and explore the codebase. Its project instructions (AGENTS.md at the repository root) tell it to prime itself on the key docs before answering. You provide the specific question and any focused context; Codex handles the rest.

This is the default second-opinion engine. For a Gemini second opinion specifically, the user invokes the separate second-opinion-gemini skill ("ask Gemini"). When the user says "ask both", run this skill and that one, then synthesize all three views (yours + Codex + Gemini).

When to Use

• Architecture decisions with real trade-offs
• Debugging where you've been going in circles
• Code review on tricky logic or subtle edge cases
• Validating your reasoning before the user acts on it
• When the user explicitly asks for a second opinion

Don't use this for routine tasks — every call takes 10-60+ seconds and consumes ChatGPT-plan quota. Reserve it for decisions where being wrong has real consequences.

Process

Step 1: Prepare the Prompt

Codex has NO access to your conversation history, but it CAN read project files. Structure your prompt as:

1. The question — what exactly you want Codex to weigh in on
2. Your current thinking (recommended) — share your position so Codex can challenge it
3. Relevant area (optional) — if the project has distinct modules or components, state which one this is about so Codex primes on the right docs. Default is the whole project; only narrow this when it helps.
4. Specific context (if needed) — pipe code snippets or diffs via stdin when the relevant code is scattered or you want to focus attention on specific sections

For questions about existing project code, you can simply reference file paths — Codex will read them itself.

Step 2: Invoke Codex

In the commands below, <repo-root> is the root of your repository — the directory containing AGENTS.md (and your .git directory). Use an absolute path so AGENTS.md and the project docs are discoverable regardless of the current working directory.

CRITICAL: Permission-free invocation pattern. The command MUST start with codex (matches the Bash(codex:*) allow rule) and MUST NOT use $() command substitution or /tmp/ file redirects — both trigger permission prompts.

Standard call (Codex reads project files itself — preferred):

codex exec -s read-only -C <repo-root> '<your question here>' < /dev/null

With context piped via stdin (for code snippets, diffs, or focused excerpts — Codex appends piped stdin as a <stdin> block):

codex exec -s read-only -C <repo-root> '<your question here>' <<'CONTEXT_EOF'
<relevant code, diff, or context here>
CONTEXT_EOF

Required flags / arguments — no exceptions:

DO NOT pass -m. Leave the model unset so Codex uses its own default. Pinning a slug goes stale, and the -codex model family (e.g. *-codex) is rejected on a ChatGPT-account login — only the chat models work, so leaving it unset is correct. If the default ever resolves to something that is NOT a real GPT chat model, treat the run as unavailable rather than accepting a degraded second opinion (see Important Rules).

Output shape. Codex prints a running transcript to stdout: a startup banner, its reasoning, any exec/tool calls it makes (e.g. rg, file reads), a tokens used line, and then the final answer as the trailing block. The Bash tool captures everything. The noise is easily identified and ignored — the final agent message is the last paragraph(s) after the tokens used line. If the output contains a 429, usage limit, quota, or rate limit error, treat as unavailable.

Set a 600-second timeout on the Bash tool call (Codex may need time to read files and reason through complex questions).

DO NOT use:

• $() command substitution (e.g., RESPONSE=$(codex ...)) — triggers a permission prompt
• /tmp/ file redirects (e.g., -o /tmp/out.txt or 2>/tmp/err.txt) — triggers an "allow access to tmp/" prompt
• Nested heredocs inside $() — same issue