AI-powered pull request reviews that don't suck.
Transforms GitHub PR diffs into scroll-based narrative documents with AI-generated explanations, interactive symbol references, and per-section verdicts.

---

⚠️ This project is archived and no longer maintained. It was built as a demo/exploration of AI-powered code review UX. The code is available for reference and learning, but there are no plans for further development or support.

---

What is it?

codereview.dog rethinks the code review experience. Instead of the standard file-by-file diff view, it uses Claude to analyze a pull request and present it as a narrative document — grouping related changes into logical sections with AI-generated explanations woven between the code.

Paste any public GitHub PR URL and get back a structured review with:

• 📚 Logical sections — Changes grouped by concept, not file
• 🧠 AI annotations — Inline explanations between diff lines
• 🔍 Interactive symbols — Hover over types/functions for definitions and docs
• ✅ Per-section verdicts — Approve, revisit, or request changes on each section
• 📊 Progress tracking — Visual progress through the review

Screenshots

Review Overview

AI-generated summary with logical section navigation, file badges, inline annotations, and concern callouts.

Symbol Hover & Annotations

Interactive symbol references with type definitions, package info, and jump-to-definition. AI concern annotations interleaved between diff lines.

Architecture

┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│   React SPA  │────▶│   Go Server  │────▶│  GitHub API  │
│  (Vite/JSX)  │     │   (net/http) │     │              │
└──────────────┘     └──────┬───────┘     └──────────────┘
                            │
                     ┌──────┴───────┐     ┌──────────────┐
                     │   SQLite DB  │     │  Claude API  │
                     │   (cache)    │     │  (Anthropic) │
                     └──────────────┘     └──────────────┘

Layer	Technology	Notes
Frontend	React 19 + Vite 5	Plain JSX, custom CSS, GitHub-dark theme
Backend	Go 1.22+	net/http with standard library router
Database	SQLite	Pure Go via modernc.org/sqlite (no CGO)
AI	Anthropic Claude	Structured JSON output for analysis
Auth	GitHub OAuth2	Web application flow with session cookies

How It Works

1. Paste a PR URL — Any public GitHub pull request
2. Fetch the diff — Server pulls the PR metadata and unified diff from GitHub's API
3. AI analysis — Claude analyzes the diff and produces structured JSON:
   • Groups changes into logical sections
   • Writes narrative introductions for each section
   • Generates inline annotations between diff lines
   • Extracts symbol definitions for hover references
4. Render the review — Frontend presents it as a scrollable narrative document
5. Review — Mark each section as approved, needs-revisit, or request-change

Project Structure

server/             # Go backend
  main.go           # Entry point, routes, middleware
  internal/
    api/            # HTTP handlers (analyze, status, dashboard, submit)
    auth/           # GitHub OAuth + session management
    github/         # GitHub API client
    diff/           # Unified diff parser
    llm/            # Anthropic Claude client
    analysis/       # AI analysis orchestrator + prompts
    db/             # SQLite + migrations
  Makefile

web/                # React frontend
  src/
    pages/          # ReviewPage, LoginPage, DashboardPage
    components/     # DiffBlock, Annotation, SymbolHover, Verdict, etc.
    auth/           # Auth context + route guard
    lib/            # API client, utilities
  vite.config.js

mock/               # Standalone prototype (reference UI)
  src/
    PRReviewApp.jsx # All components + mock data (~2100 lines)

docs/               # Specs & screenshots

Running Locally

Prerequisites

• Go 1.22+
• Node.js 18+
• A GitHub OAuth App
• An Anthropic API key

Quick Start with Docker

git clone https://github.com/jonfriesen/codereview.dog.git
cd codereview.dog

# Create .env with your credentials
cat > .env << EOF
GITHUB_CLIENT_ID=your_github_client_id
GITHUB_CLIENT_SECRET=your_github_client_secret
GITHUB_REDIRECT_URI=http://localhost:8000/auth/github/callback
LLM_API_KEY=your_anthropic_api_key
SESSION_SECRET=$(openssl rand -hex 32)
EOF

docker-compose up --build
# Open http://localhost:8000

Development

# Build everything (frontend + backend)
make

# Or run in dev mode
make dev

# Run tests
make test

Mock Prototype

The mock has a fully working UI with sample data — no backend or API keys needed:

cd mock && npm install && npm run dev
# Open http://localhost:8000

Environment Variables

Variable	Required	Default	Description
PORT	No	8000	Server listen port
APP_ENV	No	development	development or production
GITHUB_CLIENT_ID	Yes (prod)	—	GitHub OAuth app client ID
GITHUB_CLIENT_SECRET	Yes (prod)	—	GitHub OAuth app client secret
GITHUB_REDIRECT_URI	Yes (prod)	—	OAuth callback URL
LLM_API_KEY	Yes (prod)	—	Anthropic API key
LLM_MODEL	No	claude-sonnet-4-20250514	Claude model to use
SESSION_SECRET	Yes (prod)	—	Cookie signing secret (≥32 chars)
DATABASE_PATH	No	./codereview.db	SQLite database path

In development mode, all "required" variables are optional and the app runs with limited functionality.

Key Design Decisions

• Narrative over file-list — Reviews read like a document, not a file browser
• AI annotations inline — Explanations appear between diff lines, not in a sidebar
• Section-based verdicts — Review progress is per-concept, not per-file
• No TypeScript — Plain JSX for simplicity; the codebase isn't large enough to need it
• No CGO — Pure Go SQLite driver for easy cross-compilation and deployment
• GitHub-dark aesthetic — Familiar to developers, easy on the eyes

License

MIT

---

_{Built by @jonfriesen on exe.dev as an exploration of AI-powered code review UX.}