Code Review Gate for AI-Written Code: 120+ Scripted Checks

Your agent reviews its own code the way a senior engineer would: with evidence, not vibes.

Every other review skill hands your agent a checklist and hopes. This one ships the tooling: tell your agent "review this PR before I merge" and it runs 120+ deterministic checks on the diff, verifies every claim in the PR description against what the code actually does, scores the change 0-100, gives a MERGE / FIX FIRST / BLOCK verdict, and writes a review you can paste straight into the pull request. Then SKILL.md walks it through the judgement pass no script can do: intent vs implementation, edge cases, concurrency, money, rollout order.

On the bundled sample PR (a Next.js + Supabase + Stripe app, written by a coding agent whose description promises validation, error handling, tests and no breaking changes) the gate returns 0/100 BLOCK with 12 critical findings and 5 of 6 claims unsupported or contradicted. The fixed version of the same PR scores 97/100 MERGE, with 62 findings resolved against the baseline.

What you get

• One command, four outputs: review.md (PR-comment ready), report.html (printable: score ring, verdict, points by category, claims table, findings with the exact line, why it matters and the fix), review.json, and findings.sarif for GitHub code scanning. Exit code 2 on BLOCK so CI can fail.
• Claims vs diff: "added input validation", "handles errors gracefully", "includes tests", "no breaking changes", "fixes #142", "refactor only" are each checked against evidence in the diff and reported VERIFIED / UNSUPPORTED / CONTRADICTED with the lines that prove it. A contradicted description blocks the merge.
• Secrets that actually matter: 40 provider patterns (AWS, Stripe live/test/webhook, OpenAI, Anthropic, GitHub, Google, Slack, SendGrid, Twilio, npm, Supabase...), private-key blocks, database URLs with passwords, committed .env files, Supabase service_role JWTs (the payload is decoded; the public anon key is not flagged), and secret-named variables behind NEXT_PUBLIC_ / VITE_ prefixes that end up in the browser bundle.
• Access control: Next.js route handlers and server actions, Express/Fastify and FastAPI/Flask mutating routes with no auth check, webhooks that never verify a signature, .delete().eq('id', param) without an owner check, Supabase tables created without RLS, USING (true) policies, GRANT ALL TO anon, Firebase rules open to everyone.
• The classics, deterministically: SQL and shell built from strings, eval, dangerouslySetInnerHTML, pickle/yaml.load, open redirects, path traversal, wildcard CORS with credentials, TLS verification off, alg: none, jwt.decode without verify, weak hashes, Math.random() tokens, prices taken from the client, stack traces returned to users, env vars dumped to logs.
• What AI code leaves behind: empty catch blocks, except: pass, promises never awaited, .only/.skip tests, tests without assertions, console.log/print, breakpoints, TODOs, commented-out code, merge-conflict markers, as any, duplicated blocks, 200-line functions, N+1 loops in for + await.
• Change-aware checks: exported symbols removed or re-signed (callers break at runtime), endpoint files deleted, auth / rate-limit / validation / security lines that were removed, destructive migrations (DROP, RENAME, type changes, NOT NULL without default, mass UPDATE), latest/* dependencies, postinstall scripts, unpinned GitHub Actions and script injection.
• Test deltas: 15+ lines of code and no test touched, new functions no test references, test files deleted, snapshot-only tests.
• Baselines: --baseline previous/review.json reports fixed / new / unchanged after a fix loop.
• CI in one command: ci --github prints a workflow that diffs against the base branch, uploads SARIF, comments the review on the PR and fails on BLOCK; ci --pre-push prints a hook. GitLab snippet in the docs.
• Whole-repo audit (--full) for the "what did the agent leave in my codebase" question.
• Noise control: .reviewgate.json (ignored paths, disabled rules, severity overrides, intentionally public routes, allow-listed test values) and // reviewgate-ignore RG-xxx reason on a line. Severity says impact; confidence says how sure the matcher is.

Why it beats prompting

Honest by design

The scripts produce the score and verdict; the agent's judgement findings are listed separately and may override the recommendation, and the review says when they do. Every rule cites its reference (CWE, OWASP, Supabase, Firebase, Stripe, PostgreSQL, GitHub docs) and carries a confidence level. "No findings" means the checks found nothing, not that the code is correct. This is a review aid, not a security assessment.

What's inside

SKILL.md (7-step workflow, judgement pass, honesty rules, quality checklist) · 10 Python scripts (unified-diff parser, 90 line rules, secrets engine, structural checks, test deltas, claim verifier, scoring, Markdown/HTML/JSON/SARIF writers, CLI) · 6 references (every check with severity, confidence and fix; the judgement-pass method and review template; claim types; scoring; config and suppressions; CI recipes) · a sample PR in three states (before, as submitted, fixed) with diffs and descriptions · pre-built sample output.

Requirements

An agent that can run code with Python 3.8+. Standard library only: no installs, no API keys, no network. Works on diffs from git, GitHub (gh pr diff), GitLab or any unified diff.