turnStream method

Stream<SyniTurnEvent> turnStream(SyniTurnRequest request)

Streaming counterpart to turn: emits SyniTurnDeltas as the model writes, then exactly one SyniTurnFinal with the same result turn returns. On a slow device the first words arrive as soon as the prompt has been read, instead of after the whole answer. Local-only, like turn; errors (including a runtime failure envelope) surface as stream errors.

Cancelling the subscription stops the generation, so a turn the host gave up on no longer holds the model. The runtime reports tokens only once it is writing: a turn cancelled while still reading its prompt stops at its first token.

Implementation

Stream<SyniTurnEvent> turnStream(SyniTurnRequest request) async* {
  final persona = _requirePersona(mode: SyniExecutionMode.localOnly);
  final capabilities = await _runtime.capabilities();
  if (!capabilities.apiVersions.contains('2.0')) {
    throw UnsupportedError(
      'The installed Syni runtime does not support API 2.0.',
    );
  }
  final installed = currentState as SyniInstalled;
  await for (final chunk in _runtime.runStream(
    request.toRuntimeRequest(persona),
    preset: _presetForSchema(
      request.task.responseSchemaId ?? persona.responseSchemaId,
    ),
    seed: request.generation.seed ?? 0,
  )) {
    if (chunk is rt.SyniRuntimeStreamDelta) {
      yield SyniTurnDelta(chunk.text);
    } else if (chunk is rt.SyniRuntimeStreamFinal) {
      yield SyniTurnFinal(
        _turnResult(
          rt.SyniRuntimeResponse.fromJson(chunk.rawJson),
          request,
          persona,
          installed,
        ),
      );
    }
  }
}