diffpector

Diffpector Review Agent

A local code review agent that analyzes Git commits to identify potential problems, code quality issues, and security vulnerabilities. Supports both Ollama and llama.cpp backends.

Features

• Local-Only: Runs entirely on your machine - no cloud dependencies
• Multi-Language Support: Analyzes Go, Java and TypeScript code with symbol-aware context
• Git Integration: Analyzes commits and diffs
• Code Quality Analysis: Identifies potential bugs, security issues, and code smells
• Detailed Reports: Generates comprehensive code review reports
• Flexible Backend: Use Ollama or llama.cpp with OpenAI-compatible API

Installation

Download the latest binary for your platform from Releases:

• Extract the archive
• Move the binary to your PATH (e.g., /usr/local/bin on macOS/Linux)

Prerequisites

LLM Backend Setup

Choose one of the following backends:

Option 1: Ollama (Recommended for ease of use)

1. Install Ollama
2. Download your preferred model: ollama pull qwen2.5-coder:14b
3. Start the server: ollama serve

Option 2: llama.cpp (For more control)

1. Build llama.cpp with the server enabled
2. Download a GGUF model file
3. Start the server with OpenAI-compatible API:

   llama-server -m /path/to/model.gguf --port 8080
   

Git Repository

• Must be run from within a Git repository
• Requires commits to analyze

Important: Run diffpector from your project's root directory (where your .git folder is located). The tool needs to be executed from the repository root to properly analyze symbol context and cross-references.

Configuration

The agent uses default configuration for llama.cpp. Override by creating a diffpectrc.json file in your project root.

llama.cpp Configuration (Default)

{
  "llm": {
    "provider": "openai",
    "base_url": "http://localhost:8080"
  }
}

The model field is optional for llama.cpp since the model is already loaded when you start the server. The api_key field is also optional for local servers.

Ollama Configuration

{
  "llm": {
    "provider": "ollama",
    "model": "qwen2.5-coder:14b",
    "base_url": "http://localhost:11434"
  }
}

For Ollama, you must specify the model field.

Recommended Models

• qwen 3 coder (30b, q4) - best balance between accuracy and performance (if memory constrained use qwen 2.5 coder (14b, q4) instead)