| name | codex-review |
| description | Independent second-pass code review from Codex (GPT-5.x). Use when the user wants a second opinion on uncommitted changes, a branch diff, a commit, or a PR, or when a change is broad enough that another agent's perspective is useful before merging. |
Codex as independent reviewer
Codex is a different model family with different blind spots — that independence is the whole value. Its report is evidence, not verdict: you verify before you relay.
Command mechanics (flags, sandbox, resume) live in the shared wrapper's docs: C:\Users\shawn\.claude\skills\codex\SKILL.md.
Steps
-
Identify the review target. One of: uncommitted changes, a branch diff against a named base, a specific commit, or a PR. Done when you can state it as a single unambiguous git ref or range.
-
Create a temporary artifact directory for the Codex report (e.g. mktemp -d), so the report survives for inspection and never lands in the repo. Done when the directory exists outside the working tree.
-
Run the review read-only via the wrapper, telling Codex to write its full report to the artifact directory:
node C:\Users\shawn\.claude\skills\codex\scripts\codex-run.mjs --read --cd <repo> \
"Review <target>. Focus on <the one or two things that matter for this change>. Write your full report to <artifact-dir>/report.md."
Prompt Codex simply — it is not Claude. State the target, the focus, and where to put the report; skip elaborate role-play and long constraint lists. Done when the run exits and the report exists.
-
Verify the important claims against the code before presenting them. Open the files Codex cites and confirm each finding you intend to relay; drop or explicitly flag anything you can't confirm. Done when every finding in your summary has been checked in-source.
-
Report. If Codex found nothing, say that clearly and name the exact target it inspected — never silently rerun. Done when the user has either verified findings or an explicit clean bill for a named target.
