Code Review

Code Review

Standalone adversarial code review of any diff or PR — and it sharpens developer-workflows' /review when both are installed. The framing is "assume the code has bugs": a reviewer returns a failing test, a DEFECT: file:line, or NO ISSUES FOUND — never prose.

Install

claude plugin install code-review@crickets

On Antigravity, install by path (see Install crickets plugins). The command + both reviewer agents work on both hosts; the cross-model reviewer needs the Gemini CLI, and the evidence-tracker hook is Claude-only (Compatibility).

What it ships

Primitive	Kind	What it does
/code-review	command	adversarially review the current diff or PR — dispatches the reviewer agent(s)
adversarial-reviewer	agent	in-process critic — output is a failing test, a DEFECT: file:line, or NO ISSUES FOUND; prose-only critiques are rejected
adversarial-reviewer-cross	agent	cross-model critic — shells out to the Gemini CLI for a second opinion, and falls back to the in-process reviewer when gemini is absent
evidence-tracker	hook · PreToolUse	default-FAIL evidence gate for /work — blocks flipping a PLAN.md task to [x] until a read matching that task is recorded (Claude-only)

cross-review.sh is the Gemini shell-out behind the cross-model reviewer.

How it composes

• Standalone — review any diff or PR directly with /code-review; requires: [].
• Enhances developer-workflows — soft, two ways: at /review the phase dispatches the adversarial reviewers (the enhances: review declaration); evidence-tracker guards /work's checkbox-flips. Both engage only when developer-workflows is also installed.
• Hosts — the command + both reviewer agents are host-symmetric; the cross-model reviewer degrades gracefully without gemini; the evidence-tracker PreToolUse hook is Claude-only (Antigravity limitations).

Why it works

A reviewer primed to assume bugs exist finds real ones; demanding a failing test or a file:line defect keeps the output actionable instead of a sycophantic "looks good." The cross-model pass escapes the same-model echo chamber — a model reviewing its own code tends to rubber-stamp; a different model has different blind spots. And evidence-tracker stops a task being marked done before the work was actually read. See Why adversarial review.

Related

• First code review — the tutorial.
• Review a change — the how-to.
• Why adversarial review — why the assume-bugs framing works.
• Developer Workflows — the base plugin this enhances at /review.
• Hooks · Plugin anatomy — the evidence-tracker catalog entry + the shared plugin structure.
• Developer Plugin Suite design — the developer-workflows / safety / code-review split.