review

name	review
model	claude-sonnet-5
effort	low
argument-hint	[--model <gpt-model>] [--effort <level>] [review focus]
description	Delegate a read-only code review to a Codex subagent (GPT 5.6 Sol by default) that inspects code and reports findings without editing anything. Use when the user wants Codex — or "GPT-5.5" — to review, look over, audit, or give a second opinion on a diff, branch, pull request, or commit, whether they say "have Codex review…", "get Codex's take on my changes", or type `/codex:review`. Do NOT use when the user wants Codex to fix or implement changes (use codex:implement); to drive desktop or browser UI (use codex:computer); or when they want Claude's own review rather than a Codex second opinion (use the code-review skill).

Codex Review

Delegate a read-only code review to a Codex subagent running GPT 5.6 Sol (or another GPT model the user names). The subagent reads code and returns findings; it does not write files.

Choosing the model

The -m flag selects the model. Default to gpt-5.6-sol, but honor any specific model the user asks for:

• Use gpt-5.6-sol unless the user names a different model.
• If the user specifies a model — e.g. "review with gpt-5.6-terra", "use gpt-5.5-codex", "with the <name> model" — pass that exact string to -m instead. Don't validate or second-guess the name; Codex will error if it's unknown.
• If they typed /codex:review --model <name> <task> (or -m <name>), strip that flag from the review instructions and use <name> as the model.

Choosing the effort

Reasoning effort is set with -c model_reasoning_effort="<level>". Default to high, but honor any level the user asks for:

• Use high unless the user names a different level.
• If the user asks in prose — e.g. "low effort", "medium effort" — substitute that level.
• If they typed /codex:review --effort <level> [focus], strip that flag from the review instructions and use <level> as the effort.

The command below shows -m gpt-5.6-sol and high effort; substitute the chosen model and effort.

How to invoke

Codex has a dedicated review subcommand. Pick the target based on what the user wants reviewed:

User asks to review	Flag
Working-tree changes (staged + unstaged + untracked)	--uncommitted
Everything on this branch vs main (or another branch)	--base main
A specific commit	--commit <SHA>
Nothing specified (defaults to current branch vs its merge base)	(no flag)

Command shape:

codex exec review \
  -m gpt-5.6-sol \
  -c model_reasoning_effort="high" \
  --skip-git-repo-check \
  -C "$PWD" \
  <target-flag> \
  "<REVIEW-PROMPT>"

• -m gpt-5.6-sol — the model; default gpt-5.6-sol, or the model the user named (see Choosing the model).
• -c model_reasoning_effort="high" — reasoning effort; default high, or the level the user named (see Choosing the effort).
• codex exec review (not plain codex review) — the exec form is non-interactive and prints to stdout.
• No -s / -a needed: review mode is inherently read-only.
• Always set an explicit Bash timeout. Reviews are slow and the Bash default (120000 ms / 2 min) will cut Codex off before it finishes. Pass timeout: 600000 (10 min — the maximum the Bash tool allows) on every codex exec review call. For a very large diff that may exceed 10 minutes, run the Bash call with run_in_background: true and poll instead, since a foreground call cannot exceed the 600000 ms cap.
• If the user did not specify a target, ask once — or default to --uncommitted if there are uncommitted changes, otherwise --base main (or the repo's default branch).

Passing the review prompt

• If the user typed /codex:review <text>, use <text> as the review instructions.
• If they said "have Codex review this", pass a short prompt describing what to focus on ("check for correctness bugs and race conditions", "focus on the auth changes", etc.). An empty prompt is legal — Codex will do a generic review — but a focused prompt gives better results.
• Codex has no memory of this Claude conversation. Include specifics (file paths, what the change is meant to do, known constraints) in the prompt.

After the review

Summarize Codex's findings in your own words, grouped by severity or file. Cite specific file:line references from Codex's output. Do not paste the entire transcript.

If Codex flags something Claude has already discussed with the user or already decided against, note that briefly — don't re-litigate.

When NOT to use this skill

• The user wants Codex to fix what it finds → use codex:implement instead (or run this first and then codex:implement with the findings as the prompt).
• The user wants Claude's own review, not Codex's second opinion → use the code-review skill.
• The target isn't a git repo → this skill needs one. Ask the user, or use codex:implement in read-only mode with -s read-only.