Method generateWithLocalModel

app/src/utils/ModelManager.ts:567–607

* Generate a response using a local model. * Routes to the appropriate underlying manager based on the server sentinel.

(
    server: string,
    messages: LocalLlmMessage[],
    onToken?: (token: string) => void,
    onReasoningToken?: (token: string) => void,
  )

   * Routes to the appropriate underlying manager based on the server sentinel.
   */
  public async generateWithLocalModel(
    server: string,
    messages: LocalLlmMessage[],
    onToken?: (token: string) => void,
    onReasoningToken?: (token: string) => void,
  ): Promise<string> {
    if (server === ModelManager.SKIP_MODEL) {
      return '';
    }

    if (server === ModelManager.BROWSER_LOCAL) {
      const manager = GemmaModelManager.getInstance();
      if (!manager.isReady()) {
        throw new Error('Transformers.js model not loaded');
      }
      const enableThinking = manager.getState().loadSettings?.enableThinking ?? false;
      return manager.generate(messages, onToken, enableThinking, onReasoningToken);
    }

    if (server === ModelManager.LLAMA_CPP_LOCAL && isTauri()) {
      const manager = NativeLlmManager.getInstance();
      if (!manager.isReady()) {
        throw new Error('llama.cpp model not loaded');
      }
      const enableThinking = manager.getState().enableThinking;
      if (!enableThinking || !onReasoningToken) {
        return manager.generate(messages, onToken);
      }
      // Route thinking tokens to onReasoningToken, answer tokens to onToken.
      // Accumulate answer tokens — manager.generate returns the raw backend string which
      // still contains the thinking delimiters, so we can't use it directly.
      let answer = '';
      const collectAndForward = (t: string) => { answer += t; onToken?.(t); };
      const router = makeLlamaCppThinkingRouter(collectAndForward, onReasoningToken);
      await manager.generate(messages, router);
      router.flush();
      return answer;
    }

    throw new Error(`Cannot generate: unsupported server type "${server}"`);
  }

  // ===========================================================================
  // Per-Model Inference Params (remote models only — local models use their own settings)