diff-reviewer 0.1.1

AI-based local code diff reviewer that runs on every git commit.

🤖 DiffReviewer — Local AI Code Review on Every Commit

DiffReviewer is a lightweight, local-first AI assistant that reviews your code on every Git commit and provides helpful suggestions about:

• 🔍 Code quality
• ⚙️ Logical issues
• ✨ Readability and improvements

It uses a local GGUF-compatible LLM (like Mistral or StarCoder) through llama.cpp, and runs silently in the background without blocking your workflow.

---

🚀 Features

• ✅ Runs locally (no data is sent to the cloud)
• ✅ Hooks into Git using a global pre-commit hook
• ✅ Queues and reviews changes asynchronously — never blocks your commits
• ✅ Automatically deduplicates model loads (only one review at a time)
• ✅ Diff-aware (only reviews staged changes)
• ✅ Saves feedback in a .diff_review/ folder per repo
• ✅ Supports models like Mistral, StarCoder, DeepSeek etc.
• ✅ Works with or without GPU
• ✅ Highly configurable during setup

---

📦 Installation

🔹 Option 1: From PyPI

pip install diff-reviewer

🔹 Option 2: From GitHub (latest)

pip install git+https://github.com/your-username/DiffReviewer.git

🔹 Option 3: Local development install

git clone https://github.com/ChargedMonk/DiffReviewer.git
cd DiffReviewer
pip install .

💠 One-Time Setup

To set up the global Git hook and configure the reviewer:

diff-reviewer init \
  --model-dir ~/.models \
  --model-file mistral-7b-instruct-v0.2.Q4_K_M.gguf \
  --gdrive-model-url https://drive.google.com/file/d/1IVrCT8mzSNtfUJ5rTyDbLfcxbHzkcX2K/view \
  --text-context 8192 \
  --review-dir .diff_review \
  --prompt-prefix "You are a senior code reviewer. Given the diff and surrounding code context, suggest improvements in code quality, logic, and readability. Be precise and constructive." \
  --max-tokens 1024

This will:

• Register a global Git pre-commit hook at ~/.git-hooks/
• Persist your config to ~/.diff_reviewer_config.json
• Start reviewing staged changes in background after every commit

💾 Model Setup

Download a .gguf model (e.g., from HuggingFace or Google Drive), then place it in your preferred directory:

~/.models/mistral-7b-instruct-v0.2.Q4_K_M.gguf

Then use --model-dir and --model-file during init.

You can also provide a Google Drive download URL via --gdrive-model-url if the model isn't present — it will be downloaded automatically.

---

🧠 How Queued Reviews Work

To avoid system freezes and GPU overload:

✅ DiffReviewer now uses a file-based job queue. Each commit writes a job to ~/.diff_reviewer/queue, and a single background worker handles them one at a time. Why it matters:

⏱ Multiple commits don’t overload your system

🔄 Model is only loaded once per session

💻 Your terminal and Git remain fully responsive

♻️ Crashed or exited workers restart automatically

---

💡 How It Works

1. On commit, the pre-commit hook captures the staged diff.
2. The diff is base64-encoded and enqueued as a job.
3. A background worker processes jobs one by one.
4. Each changed file is reviewed using a local LLM.
5. Suggestions are saved in .diff_review/..review.txt

Example output:

📤 Running code reviewer in background... (log: /tmp/diff_reviewer_1234.log)
✅ Review completed!

Check .diff_review/ in your repo to view suggestions.

---

🧐 Example Suggestion

📝 Suggestions for utils.py:
- The function `process_data` has nested conditionals that could be flattened for readability.
- Consider using `enumerate()` instead of manually managing the index.

---

⚙️ Configuration Options (via diff-reviewer init)

Option	Description
--model-dir	Path to directory where .gguf models are stored
--model-file	Name of the model file to load
--gdrive-model-url	Optional Google Drive URL to download model if missing
--text-context	Max context length to use (e.g., 8192)
--review-dir	Directory to save review suggestions in each repo
--prompt-prefix	Prompt to guide the AI reviewer
--max-tokens	Max tokens in the generated review output

Config is saved in ~/.diff_reviewer_config.json and used automatically on every commit.

---

🐛 Troubleshooting

📌 Review doesn’t run on commit?

• Ensure you ran: diff-reviewer init
• Check that ~/.git-hooks/pre-commit exists and is executable
• Make sure the model .gguf file is correctly named and placed

🤊 Too slow?

• Use smaller quantized models (e.g., Q4_K_M instead of Q8_0)
• Avoid very large models like 13B unless you have a GPU
• Check that your text-context and max-tokens aren't too high

💥 Terminal frozen?

Ensure the reviewer is backgrounded correctly. This is handled automatically via the pre-commit script placed in ~/.git-hooks/.

---

📜 License

MIT License

🙌 Credits

Powered by llama-cpp-python and open-source GGUF models from Hugging Face, Google Drive, and the open source AI community.