si-git-commit — Super Intelligent Git Commit

AI-powered git commit messages and workflow automation. The ghsi CLI analyzes your staged changes with a current AI model and writes a Conventional Commits message — then stages, commits, and pushes in one streamlined command.

si-git-commit is the maintained successor to the orphaned ghit/helix tool. It keeps the same high-level workflow, fixes the breakage that killed the original (deprecated model ids, legacy max_tokens / temperature parameters that current models reject), and adds the quality-of-life tooling the original never had — above all, changing your model with ghsi config model instead of hunting down a dotfile to edit by hand.

Features

• One-command workflow: ghsi stages everything, generates a message, asks for confirmation, commits, and pushes.
• Modern model support: OpenAI requests use max_completion_tokens and the model's default temperature, so GPT-5-class models (gpt-5-mini, gpt-5.6-luna, gpt-5) work out of the box — no Unsupported parameter: 'max_tokens' errors, no deprecated defaults.
• Multiple providers: OpenAI, Anthropic, and any OpenAI-compatible endpoint (OpenRouter, local servers, internal gateways).
• Painless configuration: ghsi config model shows your current model and lets you keep or change it interactively; ghsi config path always tells you where the settings TOML lives.
• Streamlined CLI: short commands replace the old shell-alias workarounds, while legacy ghit-style flags keep working.
• Message levels: quiet (one line), normal (subject + bullets), or verbose (detailed body).

Installation

cargo install si-git-commit

This installs the ghsi binary (plus a si-git-commit alias binary). Prebuilt binaries for Linux, macOS, and Windows are attached to each GitHub release.

Or build from source:

git clone https://github.com/encapsa-ai/si-git-commit.git
cd si-git-commit
cargo install --path .

Quick start

# 1. Create your settings file (ghsi prints exactly where it lives)
ghsi config init

# 2. API key: if OPENAI_API_KEY is already exported (e.g. from ~/.zshrc),
#    ghsi finds it automatically — `ghsi config init` will show
#    "✓ OPENAI_API_KEY found" and there is nothing to do.
#    Otherwise export it, or set a ghsi-specific key that overrides
#    the environment with: ghsi config set api_key <your-key>
export OPENAI_API_KEY=<your-key>

# 3. Pick a model — interactively, or directly:
ghsi config model                 # shows current model, offers suggestions
ghsi config model gpt-5.6-luna    # switch in one command

# 4. Make some changes, then:
ghsi        # stage all → generate → confirm → commit → push

Usage

Subcommand shortcuts: ghsi c (commit), ghsi s (stage), ghsi gen / ghsi g (generate).

Useful flags on the commit workflow: -b/--branch <name>, --model <id> and --level quiet|normal|verbose (per-run overrides), --no-push, --dry-run, -y/--yes, --edit.

Coming from ghit? Your aliases still work

The old workflow needed shell aliases because the flags were convoluted:

alias gp="ghit --stage-and-generate --auto"   # old
alias gs="ghit --stage-and-generate"          # old

With ghsi the defaults do what those aliases meant, so the aliases collapse to:

alias gp="ghsi"         # stage all, generate, commit, push
alias gs="ghsi stage"   # stage + generate only

Legacy flags are also accepted directly: ghsi --auto, ghsi --generate, and ghsi --stage-and-generate (-a/-g/-s) behave exactly as they did in ghit.

Configuration

Run ghsi config path to see where your settings file lives. By default it is <config dir>/si-git-commit/config.toml (for example ~/.config/si-git-commit/config.toml on Linux or ~/Library/Application Support/si-git-commit/config.toml on macOS). Set $SI_GIT_COMMIT_CONFIG to use a different path. A hand-created ~/.si-git-commit.toml is also discovered for convenience.

You rarely need to open the file: ghsi config shows everything in it, ghsi config model changes the model, ghsi config set <key> <value> changes anything else, and ghsi config edit opens it in $EDITOR when you do want the file.

Example config.toml:

provider = "openai"            # openai | anthropic | openai-compatible
model = "gpt-5-mini"           # any current model id
api_key = ""                   # prefer OPENAI_API_KEY / ANTHROPIC_API_KEY env vars
message_level = "normal"       # quiet | normal | verbose
max_completion_tokens = 800
# base_url = "https://api.openai.com/v1"   # for openai-compatible endpoints

Environment overrides: OPENAI_API_KEY, ANTHROPIC_API_KEY, SI_GIT_COMMIT_API_KEY, SI_GIT_COMMIT_MODEL, SI_GIT_COMMIT_PROVIDER.

Message levels

• quiet: one-line subject only — fix: handle empty diff
• normal (default): subject plus a short bullet body
• verbose: subject plus a detailed explanatory body

A note on model parameters

Current OpenAI models reject the legacy max_tokens field and custom temperature values. ghsi sends max_completion_tokens and lets each model use its default temperature, which is accepted by both current models and the late GPT-4.1 family. Anthropic requests use Anthropic's current max_tokens parameter. Only current-generation models are suggested by ghsi config model --list, so you will never be pointed at a deprecated id.

Library usage

The generation pipeline is also a library:

use si_git_commit::{Config, Workflow, WorkflowOptions};

#[tokio::main]
async fn main() -> si_git_commit::Result<()> {
    let workflow = Workflow::new(Config::load()?);
    workflow
        .commit_workflow(vec![], WorkflowOptions::default())
        .await
}

See cargo run --example basic_usage for a network-free tour of the prompt building and response parsing helpers.

Troubleshooting

• "no config file found" — run ghsi config init. The error always prints the expected settings file path; ghsi config path shows it too.
• "no API key configured" — export OPENAI_API_KEY (or ANTHROPIC_API_KEY), or set api_key via ghsi config set api_key ….
• Model errors (deprecated / not found) — your model id is gone or mistyped. Run ghsi config model --list and switch with ghsi config model <name>.
• "No changes found to commit" — check git status; ghsi stages everything (including untracked files) by default, so this usually means the tree is genuinely clean.

License

Licensed under either of MIT or Apache-2.0 at your option.