| name | review |
| description | Read-only quality gate for the current change — runs simplify, security-review, a general Claude review, Codex review, and Codex adversarial-review against the base branch and consolidates findings. Use when the user asks to review the change/diff before committing, or invokes /review [--base <branch>] [--auto]. Read-only by default; with --auto (orchestrated pipelines) it applies safe fixes autonomously. |
| model | opus |
| effort | high |
Run the read-only quality gate on the current change, in order. Do NOT fix anything until the
user chooses what to address.
First check the branch: git branch --show-current.
Base branch <base>: use --base <branch> if the user passed one; else derive from the
current branch — if its second-to-last /-segment looks like a version (e.g. 5.1), that's
<base>; otherwise <base> = main. Inspect the change with
git diff --name-only "$(git merge-base origin/<base> HEAD)" — diffing against the merge-base
directly (not origin/<base>...HEAD) includes uncommitted staged/unstaged changes as well as
committed ones, since work-issue checkpoint 4 invokes this before the commit step; a
committed-only diff would silently see nothing to review at exactly that point.
- Fan out three independent checks in parallel — in a single message, launch three
subagents (
Agent tool, subagent_type: general-purpose), each told the current branch,
<base>, and to inspect git diff origin/<base>...HEAD, and to return findings as a plain
list (file:line + one-line description):
- Simplify — invoke the
simplify skill; return its findings verbatim.
- Security — invoke the
security-review skill; return its findings verbatim.
- General — read the diff directly for correctness issues, logic errors, missed edge
cases, and whether it satisfies the issue/spec — no named skill, just a careful read.
Separate subagents means none is biased by having just read another's findings, and all
three run concurrently instead of one after another.
- Codex review — run
/codex:review --base origin/<base>; collect its verbatim findings
(straight, non-adversarial second opinion from a different model).
- Adversarial (Codex) — run
/codex:adversarial-review --base origin/<base>; collect its
verdict + findings (Codex returns findings verbatim and applies no edits).
- Present one consolidated list grouped by source (simplify / security / general / codex
review / adversarial), each finding with
file:line and a severity.
- Use AskUserQuestion to ask which findings to fix. Apply only the chosen fixes — never
auto-apply everything.
--auto mode (for orchestrated pipelines like work-issue): skip AskUserQuestion in step 5.
Autonomously apply findings that are mechanical/safe (spelling, trivially safe idiom fixes) or
clear correctness/security defects with an obvious fix; SKIP judgment/stylistic findings. End
with a ledger of applied vs. skipped findings (the orchestrator records skipped ones in the PR
body). Interactive default behavior is unchanged.
