Codex — Second Opinion

Run OpenAI's Codex CLI non-interactively from Claude Code. Different model family = different blind spots, different strengths.

When to Use

• Code review or architecture critique (standalone or as part of /roast)
• Second opinion on a plan or approach
• Debugging when stuck — different model may see different patterns
• User explicitly asks for a second perspective

Command

_id=$(date +%s%N); codex exec -s read-only -c 'sandbox_permissions=["disk-full-read-access"]' -c 'model_reasoning_effort="xhigh"' -o "/tmp/codex-${_id}.md" "<prompt>" > "/tmp/codex-${_id}-log.md" 2>&1

• -s read-only + disk-full-read-access — can read any file on disk, not write (no codebase conflicts)
• -o — final answer to file. _id = epoch-nanosecond timestamp
• -c 'model_reasoning_effort="xhigh"' — always xhigh
• -C <dir> — working directory (defaults to cwd, set when reviewing a different project)

Run via Bash with run_in_background: true. You already have _id from the command — read /tmp/codex-${_id}.md directly. Log at /tmp/codex-${_id}-log.md if debugging.

How to Prompt Codex

Codex has zero context from your session. Everything it needs must be in the prompt or readable from the filesystem.

Give it orientation first:

• Tell it to read ./CLAUDE.md (project root) for project context, knowledge map, and conventions
• Point it to agent/knowledge/ and agent/tasks/ when relevant — written records is more efficient than re-explaining what's already documented
• Name the specific files and directories to review — it can read them, but won't know which ones matter unless told

Add session context it can't get from files:

• Describe the current approach, decisions made, constraints discovered — things only in your conversation
• State the question clearly: what you want reviewed, what kind of answer you want

Request structured output:

• "List the top 5 issues, each with: what's wrong, where, and how you'd fix it"
• "Pros and cons of this approach as a numbered list"
• Not: "What do you think?"

GPT-5 is sensitive to contradictory instructions — more so than other models. Keep prompts clean and unambiguous.