# Llama.cpp Provider Extension for Pi

This extension dynamically registers models from a llama.cpp `llama-server` instance with their context window sizes.

## Installation

### Option 1: Quick Test
```bash
pi -e /path/to/pi_agent_llama_provider/llama-cpp-provider.ts
```

### Option 2: Auto-Discovery
Copy the extension to your global extensions directory:
```bash
cp llama-cpp-provider.ts ~/.pi/agent/extensions/
```

Then run `/reload` in pi to load it automatically.

### Option 3: Custom Server URL
Set the `LLAMA_SERVER_URL` environment variable before starting pi:
```bash
export LLAMA_SERVER_URL="http://your-server:port/v1"
pi
```

Or modify the `LLAMA_SERVER_URL` constant in the extension file.

## Usage

Once loaded, the extension will:

1. Fetch all models from the llama.cpp server
2. Parse the context window size from each model's preset
3. Register them as a `llama-cpp` provider
4. Make them available via `/model` or `Ctrl+L`

### Commands

- `/llama-reload` - Reload and re-load models from the server
- `/llama-list` - List all available models with their context sizes

### Known Issues

If you run this extension as your only provider for pi agent, you'll get a
warning about no loaded models. My fix was to use an empty model list in the
default models.ini file:
https://github.com/badlogic/pi-mono/blob/main/packages/coding-agent/docs/models.md