https://github.com/matheussricardoo/git-sage
AI-powered git commit generator using local LLMs via Ollama. Runs entirely offline, no API key required.
https://github.com/matheussricardoo/git-sage
ai cli convertional-commits git llm offline ollama rust terminal
Last synced: 3 months ago
JSON representation
AI-powered git commit generator using local LLMs via Ollama. Runs entirely offline, no API key required.
- Host: GitHub
- URL: https://github.com/matheussricardoo/git-sage
- Owner: matheussricardoo
- License: mit
- Created: 2026-02-20T09:53:49.000Z (5 months ago)
- Default Branch: main
- Last Pushed: 2026-02-27T13:44:31.000Z (4 months ago)
- Last Synced: 2026-02-27T18:46:47.706Z (4 months ago)
- Topics: ai, cli, convertional-commits, git, llm, offline, ollama, rust, terminal
- Language: Rust
- Homepage:
- Size: 38.1 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
A CLI tool that uses a local LLM via Ollama to automatically generate Conventional Commit messages from your staged git diffs.
---
### Features
| Feature |
|:---|
| Local LLM — no API key, runs fully offline |
| GPU-accelerated inference via Ollama |
| Follows the Conventional Commits specification |
| Split mode — one commit per file with `-s` |
| Interactive prompt — accept, edit or skip each suggestion |
| Lock file auto-included in the first accepted commit |
| Colored output for quick visual scanning |
| Spinner feedback while the model is generating |
| Diff truncation warning with suggested `--ctx` value |
---
### Technologies
Tokio • Reqwest • Clap • Inquire • Serde • Indicatif • owo-colors • Ollama
---
### Getting Started
#### 1. Install Ollama
```bash
curl -fsSL https://ollama.com/install.sh | sh
```
#### 2. Pull the model
```bash
ollama pull qwen2.5-coder:3b
```
#### 3. Clone & build
```bash
git clone https://github.com/matheussricardoo/git-sage.git
cd git-sage
cargo build --release
```
#### 4. (Optional) Install globally
```bash
cargo install --path .
```
---
### Usage
```bash
# Stage your files first
git add .
# Generate a single commit for all staged files
git-sage
# Generate one commit per file (split mode)
git-sage -s
# Prompt to push after all commits are done
git-sage -s --push
# Use a different model
git-sage -s --model qwen2.5-coder:7b
# Adjust model creativity (0.0 = deterministic, 1.0 = creative)
git-sage -s --temp 0.2
# Increase context window for large diffs
git-sage -s --ctx 4096
# Run fully on CPU (no GPU)
git-sage --gpu 0 --threads 8
```
#### Interactive prompt
For each suggestion you can choose:
```
--- src/main.rs
⠋ Generating commit message...
Suggestion: feat(main): add --push flag and colored output
> Commit this file?
Yes ← commit as-is
No ← skip this file
Edit ← open the message for editing before committing
```
#### Truncation warning
When a diff exceeds the context window, a warning is shown with a suggested fix:
```
Diff truncated (6000/12400 bytes). Use --ctx 4133 for full analysis.
```
---
### Structure
```
src/
├── main.rs # CLI args, commit workflow, split logic, colored output
├── llm.rs # Ollama API client, spinner, truncation logic
├── git.rs # Git diff helpers, generic lock file detection
└── instructions.txt # System prompt for the LLM
```
---
### Model Configuration
All model parameters are controlled via CLI flags — no need to edit source code or recompile.
```bash
git-sage --model --threads --gpu --ctx --temp
```
To see all available models locally:
```bash
ollama list
```
To pull a different model:
```bash
ollama pull
# Examples:
ollama pull qwen2.5-coder:7b
ollama pull deepseek-coder:6.7b
ollama pull llama3.2:3b
```
---
### CLI Flags
All parameters have sensible defaults tuned for a **GTX 1050 Ti (4GB VRAM) + i5-7400 + 8GB RAM** setup.
| Flag | Default | Description |
|:---|:---:|:---|
| `-s, --split` | `false` | Generate one commit per staged file |
| `-t, --temp` | `0.0` | Model temperature (`0.0` = deterministic, `1.0` = creative) |
| `--model` | `qwen2.5-coder:3b` | Ollama model to use |
| `--threads` | `4` | CPU threads (set to your core count) |
| `--gpu` | `99` | GPU layers to offload (`99` = all, `0` = CPU only) |
| `--ctx` | `2048` | Context window in tokens (increase for large diffs) |
| `--push` | `false` | Prompt to push commits to remote after finishing |
---
### Lock File Support
git-sage automatically detects staged lock files and attaches them to the first accepted commit. No manual handling needed. Supported lock files:
| Language | Lock File |
|:---|:---|
| Rust | `Cargo.lock` |
| Node (npm) | `package-lock.json` |
| Node (yarn) | `yarn.lock` |
| Node (pnpm) | `pnpm-lock.yaml` |
| Python (poetry) | `poetry.lock` |
| Python (pipenv) | `Pipfile.lock` |
| Ruby | `Gemfile.lock` |
| Go | `go.sum` |
| PHP | `composer.lock` |
| Nix | `flake.lock` |
---
### Model Selection Guide
Choose based on your available VRAM:
| VRAM | Recommended Model | Notes |
|:---:|:---|:---|
| 2 GB | `qwen2.5-coder:1.5b` | Fast, lower accuracy |
| 4 GB | `qwen2.5-coder:3b` | Best quality/speed ratio for 4 GB cards |
| 6 GB | `qwen2.5-coder:7b` | Noticeably better descriptions |
| 8 GB+ | `deepseek-coder:6.7b` | High quality output |
| CPU only | `qwen2.5-coder:1.5b` | Use `--gpu 0`, expect slower inference |
> The default `qwen2.5-coder:3b` in Q4 quantization uses ~1.9 GB of VRAM and fits entirely on a 4 GB card.
#### Checking your specs
**GPU VRAM (Linux):**
```bash
nvidia-smi --query-gpu=name,memory.total,memory.free --format=csv
```
**CPU core count:**
```bash
nproc
```
**RAM:**
```bash
free -h
```
**Check if Ollama is using your GPU:**
```bash
ollama ps
# The "Processor" column shows GPU % vs CPU %
```
---
### Author
