Overview

review-pro is a tiered AI code-review system: triage → relevant specialist reviewers → synthesis. It reviews code written by AI agents, using AI agents. It is built to catch the failure modes AI-generated code actually ships with: code that ignores what the repository already knows, needless dependencies, over-engineering, hallucinated APIs.

This CLI installs the review-pro core (13 reviewer skills + subagents + the review-pro orchestrator) into your agent tool, and the stack packs into your repo’s .review-pro/. Supported targets: opencode · Claude Code · Cursor · Codex.

Install

Claude Code: install as a plugin

Adds this repository as a plugin marketplace, then installs the core from it. No Node.js required. Stack packs still come from the CLI.

/plugin marketplace add tufantunc/review-pro
/plugin install review-pro@review-pro

Any platform: installer CLI

One-time. Interactive — it asks which platforms and which stack packs to install.

npx review-pro init                           # interactive: pick platforms + stacks

Non-interactive (CI)? Use --target <platform|all|auto> and --no-stacks.

Per-platform notes

• opencode — skills to ~/.config/opencode/skills/, agents to ~/.config/opencode/agents/ ($OPENCODE_HOME overrides).
• Claude Code — skills to ~/.claude/skills/, agents to ~/.claude/agents/ (the agent skills: field preloads its rubric).
• Cursor — installed via Cursor /add-plugin; run /add-plugin https://github.com/tufantunc/review-pro in Cursor (the repo ships a .cursor-plugin/plugin.json).
• Codex — skills to ~/.agents/skills/, agents auto-transformed to ~/.codex/agents/*.toml. Codex auto-discovers skills; agents inherit the session skill set.

Manual install (no Node.js)

review-pro is plain markdown skills and agents — no runtime dependency. If you don’t have Node.js (or prefer not to use the CLI), copy the files by hand:

1. Core — copy core/skills/<name>/ into your tool’s skills folder, and core/agents/*.md into its agents folder (paths per platform above).
2. Stack packs — copy stacks/<pack>/ into the reviewed repo’s .review-pro/<pack>/.

# example: opencode core + python stack, by hand
cp -R core/skills/*           ~/.config/opencode/skills/
cp    core/agents/*.md        ~/.config/opencode/agents/
mkdir -p .review-pro && cp -R stacks/python .review-pro/python

The CLI just automates these copies (plus the Codex .md→.toml transform and version tracking). Manual and CLI installs are equivalent.

Stack packs

Packs layer language/framework-specific signals onto the reviewers at review time. Install the ones a repo uses into its .review-pro/. Framework/domain packs compose on top of a language pack. Beside the packs, a repository can keep its own .review-pro/rules.md: rules such as “when X changes, Y must change too”, which every review reads from the merge base and reports rule by rule.

npx review-pro                # interactive multi-select
npx review-pro add python     # one pack, CI-safe
npx review-pro list           # catalog + installed + drift

Catalog (16): python · go · rust · typescript-react · node · php · dotnet · kotlin · swift · flutter · capacitor · nextjs · tanstack-start · react-native · wordpress · ai-ml. Examples: a Next.js repo activates typescript-react + nextjs; an LLM app activates python + ai-ml.

Commands

Run a review

After install + restart, in the repo you want to review (on a feature branch):

• ask the session to review this branch with review-pro, or
• invoke the review-pro skill directly.

The agent runs the whole pipeline natively: git diff, reads changed files, reads active stacks from .review-pro/ at the merge base, dispatches the relevant reviewer subagents with their stack signals, has an independent verifier try to refute up to 8 Medium or higher code findings, reports which changed files the reviewers say they examined, and synthesizes one verdict: BLOCK / REQUEST CHANGES / APPROVE. No env vars, no scripts to run at review time.

The 13 reviewers

security · correctness · craft · ai-antipatterns · dry · performance · backend · frontend · a11y · db · api-contract · tests · spec. The ai-antipatterns reviewer is first-class: ignored existing helpers, needless dependencies, invented config/env keys, and hallucinated APIs/symbols/imports. The spec reviewer measures the diff against the issue it came from: requirements missing or half done, requirements implemented wrongly, and work nobody asked for. Reported as its own axis, never merged with the code findings.

FAQ

Does it need Node.js?

Only the installer CLI does (Node ≥ 18). The plugin itself is plain markdown skills and agents, so manual install works without Node on any platform.

How is this different from a single “review this” prompt?

A single prompt asks one agent to cover every concern at once. review-pro gives each concern its own reviewer and rubric, requires every finding to cite where its evidence was found, and has a separate agent try to refute the serious ones. Synthesis then calibrates severity against over-reporting.

Where do stack packs come from?

The catalog ships inside this package, snapshotted at publish time. New packs appear after a new version is published. See stacks/CONTRIBUTING.md in the repo to add your own.

Is it really cross-platform?

Yes — one canonical source installs into opencode, Claude Code, Cursor, and Codex. Codex agents are auto-transformed to TOML; the others consume the same markdown agents.