AI-powered code review and commit message generator using Claude.

revi analyzes your staged git changes and provides specialized code reviews before generating conventional commit messages. It uses the Claude Code SDK to communicate with Claude models through a persistent subprocess connection for intelligent feedback on security, performance, style, error handling, testing, and documentation.

• Intelligent Code Review: Automatically detects relevant review modes based on your changes
• Specialized Review Modes:
  • Security: SQL injection, XSS, authentication issues, secrets exposure
  • Performance: N+1 queries, unnecessary loops, caching opportunities
  • Style: Naming conventions, code patterns, consistency
  • Errors: Missing error checks, swallowed exceptions, edge cases
  • Testing: Untested code paths, missing assertions, coverage gaps
  • Docs: Missing comments, unclear names, API documentation
• Commit Message Generation: Creates conventional commit messages (feat, fix, docs, etc.)
• Interactive TUI: Real-time progress display with review results
• Streaming Responses: See AI output in real-time as reviews progress
• Configurable: Per-project or global configuration via YAML

• Go 1.21 or later
• Git repository with staged changes
• Node.js and npm (for Claude Code CLI installation)
• Claude Code CLI (@anthropic-ai/claude-code)
• Active Claude Max subscription (required for authentication)

First, install the Claude Code CLI globally using npm:

npm install -g @anthropic-ai/claude-code

Verify the installation:

claude --version

Authenticate using your Claude Max subscription credentials:

claude login

This will open a browser window for authentication. Once complete, you can verify authentication:

claude auth check

Install revi using Go:

go install github.com/buker/revi/cmd/revi@latest

Or build from source:

git clone https://github.com/buker/revi.git
cd revi
make build

revi uses the Claude Code CLI for authentication, which delegates to your Claude Max subscription. Authentication is handled automatically once you've completed the claude login step during installation.

No additional environment variables or API tokens are required. The Claude Code CLI manages authentication internally through a secure subprocess connection.

Stage your changes and run revi:

git add .
revi

This will:

1. Analyze staged changes to detect relevant review modes
2. Run specialized reviews in parallel
3. Generate a commit message
4. Prompt for confirmation before committing

Run code review without committing:

revi review

Generate a commit message without review:

revi commit

# Disable code review
revi --no-review

# Don't block on high-severity issues
revi --no-block

# Preview without committing
revi --dry-run

# Run all review modes
revi --all

# Enable/disable specific modes
revi --security --no-style
revi --performance --testing

# Use a different AI model
revi --model claude-sonnet-4-20250514

# Enable debug logging
revi --debug

# Show version
revi version

You can override the default model (Claude Opus 4.5) using:

• Command-line flag: --model claude-sonnet-4-20250514
• Environment variable: export REVI_AI_MODEL=claude-sonnet-4-20250514
• Config file: Set ai.model in .revi.yaml

Available models include:

• claude-opus-4-5-20251101 (default, most capable)
• claude-sonnet-4-20250514 (balanced performance/cost)
• claude-3-5-haiku-20241022 (fastest, lowest cost)

Create .revi.yaml in your project root or ~/.revi.yaml for global settings:

review:
  enabled: true
  block: true  # Block commit on high-severity issues
  modes:
    security: true
    performance: true
    style: true
    errors: true
    testing: true
    docs: true

commit:
  enabled: true

ai:
  model: "claude-opus-4-5-20251101"  # AI model to use

Environment variables are also supported with the REVI_ prefix:

export REVI_AI_MODEL=claude-sonnet-4-20250514
export REVI_REVIEW_BLOCK=false

If you encounter authentication errors:

1. Verify Claude Code CLI is installed: claude --version
2. Check authentication status: claude auth check
3. Re-authenticate if needed: claude login
4. Ensure you have an active Claude Max subscription

revi implements automatic retry with exponential backoff for rate limits. If you see persistent rate limit errors:

1. Wait a few minutes before retrying
2. Consider using a less intensive model (e.g., Haiku)
3. Reduce the number of review modes with --no-style --no-docs

revi automatically retries network errors once. If issues persist:

1. Check your internet connection
2. Verify the Claude Code CLI is functioning: claude --version
3. Try again in a few minutes

1. Mode Detection: revi analyzes your diff using Claude to determine which review modes are relevant. Falls back to heuristic detection if needed.

2. Parallel Reviews: Selected review modes run concurrently, each focused on its specific concerns.

3. Streaming Output: Review progress displays in real-time as Claude processes your code.

4. Issue Reporting: Issues are categorized by severity (high/medium/low) with locations and actionable suggestions.

5. Blocking: By default, high-severity issues block the commit. Use --no-block to override.

6. Commit Generation: Claude generates a conventional commit message based on the actual changes.

cmd/revi/          # Application entry point
internal/
  ai/              # Claude Code SDK client
  cli/             # Command-line interface (cobra)
  commit/          # Commit message generation
  config/          # Configuration management (viper)
  git/             # Git operations (go-git)
  review/          # Review modes, detection, and execution
  tui/             # Terminal UI (bubble tea)

MIT