# Local ASR Sidecar Protocol

Mastermind local-first mode expects speech-to-text to run as an external local
streaming service. The app connects to the service over WebSocket and sends
16 kHz mono PCM audio.

## Default Endpoint

```text
ws://127.0.0.1:8765/v1/asr/stream
```

The endpoint is configurable in the Local AI settings.

## Client Start Message

After the WebSocket opens, the app sends a JSON start frame:

```json
{
  "type": "start",
  "sampleRate": 16000,
  "channels": 1,
  "encoding": "pcm_s16le",
  "language": "en-US"
}
```

After that, the app sends binary frames containing raw little-endian signed
16-bit PCM audio at 16 kHz.

## Sidecar Events

The sidecar should send JSON text frames:

```json
{ "type": "ready" }
```

```json
{ "type": "partial", "text": "intermediate transcript" }
```

```json
{ "type": "final", "text": "final transcript" }
```

```json
{ "type": "error", "error": "human-readable error" }
```

Only `final` transcript events enter local semantic reduction. `partial` events
are transient status and must not be persisted.

## v1 Scope

- Required language modes: English (`en-US`), Russian (`ru-RU`), and automatic
  or mixed Russian-English recognition (`auto` or an implementation-equivalent
  mode).
- Sidecar implementation is external to this repository.
- The app does not require a specific Nemotron, NeMo, Riva, or ONNX runtime as
  long as the WebSocket protocol above is implemented.
- The endpoint must be loopback-only. LAN and internet ASR endpoints are not
  Local Providers.

## Native macOS Client Compatibility

The Swift-native app uses the same WebSocket protocol as the Electron client.
Audio frames are raw 16 kHz mono signed 16-bit little-endian PCM.

Microphone and system audio use independent client connections so the sidecar
does not mix channel roles. The sidecar does not need to know whether a client
is Electron or Swift; Mastermind associates each connection with its Source.

Full transcripts are ephemeral sensitive input. The Swift client consumes a
`final` event, derives structured local context, and then releases transcript
content instead of storing it in Assistant History or the Context Graph.