| name | reviewer |
| description | Run the adversarial, read-only codex reviewer on the current diff and fix what it finds. Invoke as /reviewer with optional focus text and scope flags — it kicks off codex immediately (spends codex/OpenAI quota) and injects the report for Claude to act on. codex reads the whole repo to dig but CANNOT edit anything. |
| argument-hint | [focus text] [--uncommitted | --commit SHA | --range A..B | --base BRANCH] [--spec PATH_OR_URL] |
| allowed-tools | Bash($HOME/.claude/skills/reviewer/review.sh --quiet*) |
| disable-model-invocation | true |
reviewer
The read-only, adversarial codex reviewer has been run on the current diff — it
reads the whole repo to dig, but cannot edit anything. Its report:
$HOME/.claude/skills/reviewer/review.sh --quiet --focus-stdin <<'__REVIEWER_FOCUS_EOF__'
$ARGUMENTS
__REVIEWER_FOCUS_EOF__
That report is above. Now act on it:
- Triage every finding — real blocker vs nit vs false positive. codex digs
but can be wrong: verify each claim against the actual code before touching
anything.
- Fix the genuine issues yourself (the reviewer never edits). Make the
smallest correct change per issue.
- Account for the rest: for anything you skip, one line saying why (false
positive / out of scope / intended).
Finish with a short summary of what you changed and what you deliberately left.
If codex produced no report (e.g. no changes to review, or it errored), say so
and stop — don't invent findings.
Args after /reviewer are forwarded to the reviewer: free text = focus; scope
flags pick what to review (default: working tree vs merge-base with the base
branch); --spec PATH_OR_URL points codex at the original issue/spec by link.
Full reference in README.md / review.sh --help.
