---
name: voice-assistant
description: Handle voice assistant requests received via the wake-word pipeline. Use when a request arrives tagged as a voice command or comes through the wake-word webhook (/wake endpoint). Respond concisely — responses will be spoken aloud via TTS (Kokoro). Avoid markdown, lists, or formatting that doesn't work in speech. Keep replies to 1–2 sentences unless detail is requested.
---

# Voice Assistant Skill

## Context

This assistant runs on a Mac Mini (LINDBLUM, 10.0.0.200). Requests may arrive:
- Via the `/wake` HTTP webhook (wake word detected by openWakeWord)
- Via Home Assistant Wyoming voice pipeline
- Via direct text input

## Response style for voice

- Speak naturally, as if in conversation
- Keep it short — 1–2 sentences by default
- No bullet points, headers, or markdown
- Say numbers as words when appropriate ("twenty-two degrees" not "22°C")
- Use the character's personality (defined in system prompt)

## TTS pipeline

Responses are rendered by Kokoro ONNX (port 10301, voice: af_heart) and played back through the requesting room's speaker.

## Smart home integration

For device control requests, use the `home-assistant` skill. HA is at 10.0.0.199:8123.

## Wake word webhook

POST to `http://localhost:8080/wake` triggers this context:
```json
{"wake_word": "hey_jarvis", "score": 0.87}
```
After wake, wait for the transcribed utterance from the STT pipeline (Whisper large-v3, port 10300).