create-commit

AI-assisted git commit helper for generating well-structured commit messages directly from your working tree.

Features

• Infers commit style from recent history and branch naming conventions (e.g. ticket IDs).
• Summarizes the current diff per file and asks an AI model (OpenAI GPT-5 family or local Ollama) to propose a commit plan.
• Can split changes into multiple commits when appropriate and apply them automatically.
• Supports a dry-run mode for reviewing the generated plan before any commits are created.
• Works with OpenAI or a local Ollama model so you can stay online or offline. Ollama runs an extra diff-based refinement pass to tighten commit summaries and enforce ticket prefixes when required.

Prerequisites

• Rust toolchain (edition 2024 compatible).
• For OpenAI: set OPENAI_API_KEY (and optionally OPENAI_BASE_URL, OPENAI_MODEL).
• For Ollama: run a local Ollama daemon with a chat-capable model pulled (ollama pull llama3.2, etc.). Optional overrides via OLLAMA_BASE_URL and OLLAMA_MODEL.

Usage

cargo run -- --dry-run
# Review the suggested commits, then run without --dry-run:
cargo run --

Key flags:

• --dry-run: only prints the plan produced by the model.
• --allow-dirty-index: proceed even if you already have staged changes.
• --history-limit <n>: adjust how many recent commit messages are sampled.
• --diff-char-limit <n>: cap the characters captured per file diff.
• --provider {openai|ollama}: choose the AI backend (default openai).
• --model <name> / --openai-base-url <url>: override OpenAI configuration without touching env vars.
• --ollama-model <name> / --ollama-base-url <url>: override Ollama configuration when using the local provider.

Using Ollama

# Ensure Ollama is running and your model is available
ollama pull llama3.2

cargo run -- --provider ollama --ollama-model llama3.2 --dry-run

How it works

1. Collects git metadata (branch name, recent commit subjects, working tree status, per-file diffs).
2. Sends that context to the selected AI provider (OpenAI chat completions or Ollama) and asks for a JSON commit plan.
3. Optionally prints the plan (--dry-run) or stages the suggested files and performs commits accordingly.

If the model suggests a commit containing files that have no changes, the tool skips that commit and emits a warning.

Notes

• The tool expects to run inside an existing git repository.
• Ensure you have a clean index unless you pass --allow-dirty-index.
• Review generated commits before pushing. The model responses should be treated as suggestions and verified in your workflow.