GitHub - Waraq-Labs/review-for-agent: A local code review tool for giving structured, line-level feedback to AI coding agents — without opening a PR.

review-for-agent

A local code review tool for giving structured, line-level feedback to AI coding agents — without opening a PR.

Review uncommitted changes in a GitHub-style diff view, leave inline comments, and export a Markdown file you can feed directly back to your AI agent.

Why

After an AI coding agent edits files in your repo, there's no lightweight way to give it precise, line-level feedback. Opening a GitHub PR just to leave review comments on local changes is overkill. Copy-pasting file paths and line numbers into chat is error-prone and loses context.

review-for-agent gives you a familiar PR review UI that runs entirely on your machine and outputs a structured Markdown file the agent can consume.

Features

• GitHub-style diff view — unified or split, powered by @pierre/diffs
• Inline comments — click a line number to comment; shift-click to select a range
• File-level comments — feedback on an entire file, not tied to a specific line
• Global review comment — overall feedback at the top of the review
• Structured export — outputs both JSON and agent-friendly Markdown with diff context snippets
• Clipboard integration — copies review my comments on these changes in @rfa/comments_xxxx.md to your clipboard on submit, ready to paste into your agent
• Local — no GitHub, no PRs. Single binary runtime with embedded frontend bundle

Install

Binary releases

Download the latest binary for your platform from the GitHub Releases page.

# macOS (Apple Silicon)
tar -xzf review-for-agent_*_darwin_arm64.tar.gz

# Linux (x86_64)
tar -xzf review-for-agent_*_linux_amd64.tar.gz

# Linux (ARM64)
tar -xzf review-for-agent_*_linux_arm64.tar.gz

# Move to a directory on your PATH
mv review-for-agent /usr/local/bin/

macOS note: The macOS binary is not yet notarized, so Gatekeeper will show a "can't verify" warning. Remove the quarantine attribute to run it:

xattr -d com.apple.quarantine review-for-agent

Notarized binaries are coming soon. If running this makes you uncomfortable, follow the directions to build from source below.

From source

Requires Go 1.23+.

go install github.com/Waraq-Labs/review-for-agent@v0.3.3

Build locally

git clone https://github.com/Waraq-Labs/review-for-agent.git
cd review-for-agent
make setup
make build

Development

Requires Go 1.23+, Node.js, and pnpm.

# terminal 1
make dev-api

# terminal 2
make dev-web

make dev-web runs Vite and proxies /api/* to http://localhost:4000.

Usage

Run from the root of any Git repo with uncommitted changes:

$ review-for-agent
Listening on localhost:4000
Opening http://localhost:4000/review

Your browser opens with a diff of all uncommitted changes. Leave comments, then click Submit Review.

Wrote rfa/comments_5ae2.json
Wrote rfa/comments_5ae2.md

The Markdown file path is copied to your clipboard as:

review my comments on these changes in @rfa/comments_5ae2.md

Paste that into your AI agent's chat to have it address your feedback.

Flags

Reviewing changes

1. Click a line number to add a single-line comment
2. Click a line, then shift-click another to comment on a range
3. Click "+ File comment" in a file header for file-level feedback
4. Type a global comment in the textarea at the bottom
5. Click "Submit Review" to export

Port selection

The server tries port 4000 first, then increments until it finds a free port.

Ignore paths with .rfaignore

If there are paths you never want in local review diffs (for example generated frontend bundles), add a .rfaignore file at repo root.

Example:

review-for-agent reads .rfaignore at startup and excludes matching paths when generating tracked and untracked diffs. Use one pattern per line. Empty lines and lines beginning with # are ignored.

Lockfiles are always excluded by the server, even if they are not listed in .rfaignore:

• package-lock.json
• pnpm-lock.yaml

Build Commands

Output

Comments are written to the rfa/ directory (created if it doesn't exist).

Markdown format

The Markdown file is designed to be consumed by AI agents. It includes a preamble explaining the file structure, then comments grouped by file with diff context snippets:

# Code Review Comments

> **How to read this file:**
> This file contains review comments on uncommitted changes in this repo.
> Comments are grouped by file. Each comment includes a line or line range
> reference and a quoted diff context snippet showing the relevant code.
> File-level comments (not tied to a specific line) appear under a
> "(file-level)" heading. A global comment, if present, appears at the top
> before any file sections.

Overall, nice progress but a few things to address.

## src/api/handler.ts

### Line 42
> +  const result = await fetch(url);
Add error handling here — what happens if the fetch fails?

### Lines 78–85
> +  if (user.role === 'admin') {
> +    grantAll(user);
> +  }
This grants blanket permissions. Should we scope this to the specific resource?

## src/utils/parse.ts (file-level)
This file duplicates logic from src/core/parser.ts — consider consolidating.

JSON format

The JSON file contains the raw structured data for programmatic use:

{
  "globalComment": "Overall, nice progress but a few things to address.",
  "comments": [
    {
      "file": "src/api/handler.ts",
      "startLine": 42,
      "endLine": 42,
      "side": "right",
      "body": "Add error handling here — what happens if the fetch fails?"
    }
  ]
}

Requirements

• Git
• A modern web browser

License

MIT