GitBuddy

GitBuddy is an AI-driven CLI that generates Conventional Commit messages from your staged Git diff. It can preview the message, commit for you, optionally push, and fall back to auto-staging when you run gitbuddy with no subcommand.

Features

• AI-powered commit subjects: Generate a single Conventional Commit subject from the current staged diff.
• JSON-first default prompt: The default prompt asks models to return {"subject":"..."} for easier parsing.
• Repair fallback: If a model returns analysis, markdown, or malformed output, GitBuddy retries with a repair prompt.
• Multiple vendors: Supports DeepSeek, OpenAI-compatible endpoints, Ollama, and MiniMax.
• Runtime overrides: Temporarily switch vendor, model, or prompt preset without rewriting your saved config.
• Smart diff truncation: Uses the full diff for normal changes, compact summaries for large diffs, and stats for very large diffs.
• Failure diagnostics: Writes detailed invalid-response diagnostics to your GitBuddy config directory while keeping the terminal error short.

Installation

From crates.io

cargo install gitbuddy

To update if you already have installed, run:

cargo install --force gitbuddy

📦 crates.io/crates/gitbuddy

🍺 Homebrew support is coming soon.

Configuration

GitBuddy stores configuration in ~/.config/gitbuddy/config.toml.

Configure a vendor with its default model:

gitbuddy config --api-key <your-api-key> deepseek

Configure a specific model:

gitbuddy config --api-key <your-api-key> --model gpt-4o openai

For local Ollama, the API key can be any non-empty placeholder:

gitbuddy config --api-key local --model llama3.1 ollama

Usage

After staging your changes, run:

gitbuddy ai

Or run gitbuddy with no subcommand to auto-stage all changes and auto-commit without an extra confirmation prompt.

Options

Option	Description
--push	Push the commit to the remote repository after committing. Without this flag, GitBuddy asks before pushing.
--dry-run	Generate the commit message and print the Git commands without creating a commit or pushing.
--prompt <P>, -p <P>	Select a built-in prompt preset: P1, P2, P3, P4, or P5 (default: P1).
--vendor <VENDOR>	Temporarily override the default vendor for this run.
--model <MODEL>, -m <MODEL>	Temporarily override the model for this run.

Examples:

# Auto-stage all changes, generate message, and auto-commit
gitbuddy

# Generate a message from staged changes, confirm before committing, and push
gitbuddy ai --push

# Preview the generated message without committing
gitbuddy ai --dry-run

# Use a different prompt style
gitbuddy ai -p P3

# Temporarily use a specific model
gitbuddy ai --vendor openai --model gpt-4o

Commit Message Format

The default P1 prompt expects the model to return a JSON object with a single subject field:

{"subject":"fix(ai): repair malformed commit subject output"}

GitBuddy extracts and validates the subject before committing. Accepted commit types are feat, fix, docs, style, refactor, perf, test, chore, ci, and build.

If the response cannot be parsed, GitBuddy sends a second repair request. If repair also fails, it writes a diagnostic log such as ~/.config/gitbuddy/llm-diagnostic-<timestamp>.log with the model name, prompt preset, original response, repair response, and prompt previews.

Support models

Vendor	Default Model	Custom Model	Status
DeepSeek	deepseek v4 flash	yes	✓
Ollama	ollama	yes	✓
OpenAI	gpt-3.5-turbo	yes	✓
MiniMax	MiniMax-M2.7	yes	✓

DeepSeek, OpenAI, and Ollama use OpenAI-compatible APIs. MiniMax uses an Anthropic-compatible API. All vendors support custom model names via gitbuddy config --model ... or the runtime --model flag.

Default endpoints are:

Vendor	Default endpoint
DeepSeek	https://api.deepseek.com
Ollama	http://localhost:11434
MiniMax	https://api.minimaxi.com/anthropic
OpenAI	OpenAI SDK default endpoint

Troubleshooting

• Config not found. Run gitbuddy config first.: Configure at least one vendor before running gitbuddy ai.
• No files added to staging!: gitbuddy ai reads staged changes. Run git add ... first, or run gitbuddy to auto-stage all changes.
• LLM did not return a valid conventional commit subject.: Check the diagnostic file path printed in the terminal. The log is written under ~/.config/gitbuddy/ and includes the raw model output plus prompt previews.
• Empty message from a reasoning model: Use a standard chat model, for example deepseek-chat, instead of a model that returns only reasoning content.

Roadmap

• Enhance the User Interface.
• Using configuration file instead of environment variables.
• Support for more AI models.
• Add statistics and analytics for GitBuddy usage of kinds of Models.
• Support HTTP proxy.
• Allow fully custom user-defined prompts.
• Install for using GitBuddy by Git Hooks (without gitbuddy ai).
• Submit a single request to receive multiple options for users to select from.