Generate AI-assisted commit messages from staged Git changes

git

commit-message

llm

cli

Download zip

Author

klaseca

Version

0.1.0

License

Apache-2.0

Repository

github.com/klaseca/genmit

Last updated

2 hours ago

Downloads

Dependencies

#genmit

genmit generates a commit message from the changes already staged in Git. It never stages files and never creates the commit: the CLI prints a message for you to review, while the VS Code extension places it in the Source Control input.

The default provider is Codex CLI, so an existing ChatGPT/Codex login is enough to get started. OpenAI Responses, OpenAI-compatible Chat Completions, Anthropic Messages, and custom command-line harnesses are also supported.

#Quick start

Install the CLI from Mooncakes with an up-to-date MoonBit toolchain:

moon install klaseca/genmit

Install Codex CLI, sign in once, then stage the changes you want described:

codex login
git add <files>
genmit

genmit prints the generated message to standard output. Review it and create the commit normally. The default Codex harness runs codex exec in its read-only sandbox and reuses the existing login. It ignores the Codex user configuration, rules, and project instructions, and disables shell access, hooks, apps, subagents, and web search. Configure its model and reasoning effort through genmit instead.

#CLI

Running genmit without a command is the same as genmit generate.

genmit [options]
genmit generate [options] [--json]
genmit prompt [options]
genmit providers [options] [--json]

Commands:

generate generates and prints a commit message.
prompt prints the direct prompt without contacting the provider.
providers lists the providers available in the active configuration and marks the selected one with *.

providers only reads configuration and does not require a Git repository.

Options:

-C, --cwd <directory> selects the Git repository and the base for a relative config path.
-c, --config <file> uses one TOML file instead of automatic user-config discovery.
-p, --provider <name> selects a configured provider without changing the config file.
--json returns machine-readable output from generate or providers.
-V, --version prints the version.
-h, --help prints help. Use genmit <command> --help for command-specific options.

The --cwd, --config, and --provider options may appear before or after a command. For example, these are equivalent:

genmit --provider openai generate
genmit generate --provider openai

Only staged changes are used. Unstaged and untracked content is not collected by genmit.

genmit generate --json returns a stable result for scripts and editor integrations:

{
  "protocol_version": 1,
  "message": "feat: improve the CLI",
  "provider": "codex",
  "files": ["src/cli.mbt"]
}

#Configuration

Without --config, genmit starts with built-in defaults and then looks for one user config:

$GENMIT_HOME/config.toml, when GENMIT_HOME is set;
otherwise <home>/.config/genmit/config.toml.

On Windows, USERPROFILE is used as the home directory, with HOME as a fallback. If no file exists, the built-in configuration is ready to use with Codex CLI.

Passing --config <file> disables discovery and applies that file over the built-in defaults. Configuration files are not combined with each other. Relative paths are resolved from --cwd, or from the current directory when --cwd is omitted.

A small user config can contain only the settings being changed:

provider = "codex"
language = "en"
style = "conventional"
include_body = true
max_subject_length = 72

[generation]
max_provider_calls = 32
timeout_ms = 600000

[prompt]
extra = ""

The main generation settings are:

language: the language requested for the message;
style: conventional, plain, or gitmoji;
include_body: whether a body is allowed after the subject;
max_subject_length: subject limit from 20 to 200 characters;
prompt.extra: additional instructions;
generation.max_provider_calls and generation.timeout_ms: whole-generation safety limits.

The built-in providers are:

Name	Kind	Connection
codex	codex	Codex CLI with the existing login
openai	openai-responses	OpenAI Responses API
anthropic	anthropic-messages	Anthropic Messages API

The built-in codex and openai providers use gpt-5.6-luna; anthropic uses claude-sonnet-5. Their default reasoning effort is low. A user config may override individual settings without repeating the whole built-in provider:

[providers.codex.config]
model = "gpt-5.6-luna"
reasoning_effort = "low"

To use the built-in OpenAI provider, select it and expose the API key to the process:

provider = "openai"

$env:OPENAI_API_KEY = "..."
genmit

New API providers require an explicit kind, endpoint, and model. Supported API kinds are openai-responses, openai-compatible, and anthropic-messages. Custom command-line integrations are defined as reusable harnesses, then referenced by name from a provider. See genmit.example.toml for both forms.

For a harness, args are followed by the arguments produced by configured parameters and then by tail_args. A parameter is included when its provider value is non-empty, and ${value} is replaced with that value. The prepared prompt is passed through standard input. Harnesses run in the repository root, must remain in the foreground, and should clean up their own child processes when cancelled.

API keys are read from the environment variable named by api_key_env; the key itself should not be written to TOML. reasoning_effort is optional and its valid values depend on the selected model.

#Large staged changes

Small staged diffs use one provider request. Larger diffs are split without dropping content, summarized, and reduced before the final commit message is generated. The default limits are 32 provider calls and 600000 milliseconds for one generation. If the work cannot fit within those limits, genmit fails instead of silently omitting part of the diff.

#VS Code extension

The extension in apps/vscode-extension bundles the JavaScript build of the MoonBit runtime, so the standalone CLI is not required. It adds a Source Control action that generates a message for the selected repository and writes it to the commit input. Provider selection is available from the Source Control overflow menu and the Command Palette.

See the extension README for installation, usage, configuration, cancellation, and troubleshooting.

#Security

Review git diff --cached before generation. A staged diff can contain secrets and is sent to the selected network provider. A custom harness runs in the repository root with the permissions of its executable. The staged-only rule limits what genmit collects; it is not a filesystem sandbox for an external program. The built-in Codex harness requests Codex's read-only sandbox. It also ignores Codex user configuration, rules, and project instructions and disables shell access, hooks, apps, subagents, and web search. This keeps the built-in harness focused on the prompt prepared by genmit. A custom harness can opt into normal Codex behavior when repository-aware generation is desired.

#Development

Build and test the MoonBit module:

moon check
moon test
moon build src --release
moon build src --target wasm

Install the CLI from the current checkout:

moon install ./src

Build and test the extension:

cd apps/vscode-extension
npm install
npm test
npm run package

The extension build compiles src/js_runtime to JavaScript and bundles it with esbuild. For package boundaries and runtime details, see docs/architecture.md.

Action

source

type Action derive(Eq, @moonbitlang/core/debug.Debug
Debug)

CliOptions