Tiered AI code review
Twelve specialists.
One verdict.
review-pro reviews AI-written code in stages. Triage dispatches only the relevant specialists, each runs in its own scoped context, then synthesis returns one verdict. Built for the failure modes AI code actually ships with.
Triage
Classify the diff and dispatch only the specialists that matter for this change. Each one gets scoped context — callers, repo search, schema, consumers — not the whole repository.
Fan-out
The relevant specialists run in parallel. Twelve reviewers, each owns one concern, each returns structured, evidence-backed findings with file, line, impact, and remedy.
Synthesis
Dedupe overlaps, resolve cross-reviewer conflicts by domain ownership, calibrate severity against over-reporting, and emit one verdict — BLOCK, REQUEST CHANGES, or APPROVE.
Capabilities
A review bench, not one big guess.
Every specialist owns a single concern and a strict rubric. A first-class AI-anti-patterns lens catches what generic reviewers miss — hallucinated APIs, over-engineering, ignored helpers, needless dependencies.
12 specialist reviewers
Each owns one concern, with a strict, calibrated rubric and structured output.
16 stack packs
Language & framework signals layered onto the core rubrics at review time. Framework packs compose on a language pack.
4 platforms
One canonical source installs into each tool from a single command. Codex agents auto-transform to TOML; Cursor can /add-plugin the repo directly.
Install
One command. No lock-in.
Install the core into your tool, then add the stack packs you use.
npx review-pro init # interactive: pick platforms + stacks
npx review-pro add python # add a stack to .review-pro/No Node.js? review-pro is plain markdown skills and agents — copy them into your tool’s folder and it works the same. See the manual install guide.
Then restart your tool and run a review — open a branch and ask the session to review it, or invoke the review-pro skill.