Choose commit types with a decision model

Use a TypeSafe-compatible decision service to select the Conventional Commit type after LiteLLM drafts the description and optional body. If you leave the decision URL unset, LiteLLM chooses the type as usual.

Before you begin

Install the hook and configure a LiteLLM model with LITELLM_PROXY_MODEL or --model.
Choose either the TypeSafe Jev API or a running Ollaya server.

Use Jev

Get an API key from the TypeSafe dashboard.
Set the decision URL and your key in the environment where the Git hook runs:
```
export TYPESAFE_BASE_URL=https://api.typesafe.ai
export TYPESAFE_API_KEY="your-typesafe-key"
```
Replace the example key with your own. Jev requires the key in an Authorization: Bearer header over HTTPS; the decision URL alone cannot authenticate a request. The default decision model is jev-latest.

Use Ollaya

Install Ollaya and load the default laya model:
```
ollaya run laya "A short text to load the model"
```
This command starts the server and pulls the model if necessary.
Point the hook at the local server:
```
unset TYPESAFE_API_KEY
export TYPESAFE_BASE_URL=http://127.0.0.1:11435
export NO_PROXY=localhost,127.0.0.1
```
Local Ollaya does not require a key unless the server is configured to require one. Unset any Jev key in the hook’s environment before connecting to a local server. The TypeSafe SDK sends a placeholder local key when TYPESAFE_API_KEY is unset. If the server requires a key, set TYPESAFE_API_KEY to that key in the hook’s environment. You can use TYPESAFE_MODEL to choose a different installed model.

Generate a message

Stage your changes and commit normally. The hook asks LiteLLM to generate the description first, then sends it to the configured decision service to choose the type. You still review the completed message before the hook writes it. If the decision service is unavailable or returns an unsupported type, the hook stops without writing the message.

You can also set the URL for one invocation with --decision-url and override the decision model with --decision-model. See the Configuration reference for all options.