GitAICommit

Rust CLI for AI-assisted git commits, pushes, and pull requests.

It analyzes the current repository state, builds a constrained prompt from the relevant diff context, and runs that through a local or compatible LLM backend to generate:

• commit messages
• pull request titles and bodies
• combined commit + push + PR flows

Current Status

The README now reflects the code that is implemented and validated in this repo:

• provider-neutral LLM manager with ollama and openai-compatible
• custom prompt templates with --template
• explicit user context with --context
• real --push, --pr, and --push-pr flows
• default-branch detection
• auto feature-branch creation when pushing from the default branch
• gh pr create integration with existing-PR fallback via gh pr view --json url

Prerequisites

• Rust 1.70+
• Git
• One of:
  • Ollama
  • an OpenAI-compatible HTTP endpoint
• gh if you want real PR creation instead of dry-run PR drafts

Installation

git clone https://github.com/npiesco/gitAICommit.git
cd gitAICommit
cargo install --path .

Quick Start

Basic commit flow:

git-ai-commit

Stage everything first:

git-ai-commit --add-unstaged

Preview only:

git-ai-commit --dry-run --verbose

Generate a PR draft instead of a commit:

git-ai-commit --pr --dry-run

Commit and push:

git-ai-commit --push

Commit, push, and open or reuse a PR:

git-ai-commit --push-pr

Providers

Ollama

Default provider:

git-ai-commit --provider ollama --model qwen3-coder:latest

List local models:

git-ai-commit --list-models

OpenAI-Compatible

Use any OpenAI-compatible endpoint:

git-ai-commit \
  --provider openai-compatible \
  --model your-model-name \
  --base-url http://localhost:11434

With an API key:

git-ai-commit \
  --provider openai-compatible \
  --model gpt-4.1-mini \
  --base-url https://api.openai.com \
  --api-key "$OPENAI_API_KEY"

Notes:

• --model is required for non-Ollama providers.
• If --provider openai-compatible is used without --base-url, the tool defaults to http://localhost:<port>.

CLI Options

git-ai-commit [OPTIONS]

Model options:
    --provider <PROVIDER>          LLM backend: ollama | openai-compatible
    -m, --model <MODEL>            Model name
        --base-url <URL>           Base URL for selected provider
        --api-key <KEY>            API key for selected provider
        --list-models              List models for the selected provider
    -p, --port <PORT>              Ollama port [default: 11434]

Diff options:
    -f, --max-files <COUNT>        Max files included in analysis [default: 10]
    -l, --max-diff-lines <LINES>   Max diff lines / prompt budget unit [default: 50]

Commit options:
    -a, --add-unstaged             Stage unstaged changes first
        --pr                       Generate PR title/body instead of a commit
        --push                     Push after commit
        --push-pr                  Commit, push, then open/reuse PR
        --confirm                  Ask before committing

Customization:
        --template <FILE>          Custom prompt template
        --context <TEXT>           Extra user context for commit/PR generation

Debug:
    -d, --dry-run                  Preview without mutating git state
    -v, --verbose                  Print generated prompt/output details

Advanced:
    -t, --timeout-seconds <SEC>    Generation timeout [default: 60]

Behavior

Commit Generation

• default commit prompts use staged diff summary context
• AI output is sanitized before commit
• the final commit is written through git commit --file
• malformed prompt echoes and shell-wrapper junk are stripped before write

Push Behavior

• if you run --push or --push-pr from a non-default branch, that branch is reused
• if you run from the detected default branch, the tool creates a slugified feature branch first
• branch naming uses user --context first, then the sanitized commit message

PR Behavior

• --pr --dry-run prints a generated TITLE: / BODY: draft
• non-dry-run PR flows call gh pr create --base <default_branch>
• if gh pr create fails because a PR already exists, the tool falls back to gh pr view --json url
• clean worktrees are supported when the current branch is already ahead of the default branch

Custom Templates

Use a custom prompt template file:

git-ai-commit --template ./prompt.txt

The built-in default prompt is intentionally constrained, but custom templates still receive the richer {CONTEXT} payload from the repo analysis path.

Configuration

Config file path:

~/.config/git-ai-commit/config.toml

Example:

provider = "ollama"
model = "qwen3-coder:latest"
base_url = "http://localhost:11434"
api_key = ""
max_files = 10
max_diff_lines = 50
port = 11434
timeout_seconds = 60

Notes:

• CLI flags override config values.
• For openai-compatible, set model explicitly in config or on the command line.

Examples

Commit with explicit context:

git-ai-commit --context "focus on cleanup before release"

Preview a PR from an existing feature branch:

git-ai-commit --pr --dry-run --context "user-facing behavior change"

Commit, push, and open/reuse a PR through local Ollama OpenAI-compatible mode:

git-ai-commit \
  --push-pr \
  --provider openai-compatible \
  --model tinyllama:latest \
  --base-url http://localhost:11434

Testing

Local test loop:

cargo test
cargo fmt --all
cargo clippy --all-targets --all-features -- -D warnings
cargo test
cargo build --release

Live GitHub-backed integration test:

cargo test test_push_pr_with_real_github_repo_creates_then_reuses_pull_request --test interactive_test -- --ignored

That test:

1. creates a disposable private GitHub repo with gh
2. clones it into /tmp
3. runs the real git-ai-commit PR flow
4. verifies PR creation and existing-PR reuse
5. deletes the repo on teardown

License

MIT