voice agent works

example/demo.ts

@@ -3,56 +3,160 @@ import { VoiceAgent } from "../src";
 import { tool } from "ai";
 import { z } from "zod";
 import { openai } from "@ai-sdk/openai";
+import { writeFile } from "fs/promises";
 
 // 1. Define Tools using standard AI SDK
 const weatherTool = tool({
-    description: 'Get the weather in a location',
+    description: "Get the weather in a location",
     inputSchema: z.object({
-        location: z.string().describe('The location to get the weather for'),
+        location: z.string().describe("The location to get the weather for"),
     }),
     execute: async ({ location }) => ({
         location,
         temperature: 72 + Math.floor(Math.random() * 21) - 10,
+        conditions: ["sunny", "cloudy", "rainy", "partly cloudy"][Math.floor(Math.random() * 4)],
     }),
 });
 
-// 2. Initialize Agent
+const timeTool = tool({
+    description: "Get the current time",
+    inputSchema: z.object({}),
+    execute: async () => ({
+        time: new Date().toLocaleTimeString(),
+        timezone: Intl.DateTimeFormat().resolvedOptions().timeZone,
+    }),
+});
+
+// 2. Initialize Agent with full voice support
 const agent = new VoiceAgent({
-    model: openai('gpt-4o'),
+    // Chat model for text generation
-    instructions: "You are a helpful voice assistant. Use tools when needed.",
+    model: openai("gpt-4o"),
+    // Transcription model for speech-to-text
+    transcriptionModel: openai.transcription("whisper-1"),
+    // Speech model for text-to-speech
+    speechModel: openai.speech("gpt-4o-mini-tts"),
+    // System instructions
+    instructions: `You are a helpful voice assistant. 
+Keep responses concise and conversational since they will be spoken aloud.
+Use tools when needed to provide accurate information.`,
+    // TTS voice configuration
+    voice: "alloy", // Options: alloy, echo, fable, onyx, nova, shimmer
+    speechInstructions: "Speak in a friendly, natural conversational tone.",
+    outputFormat: "mp3",
+    // WebSocket endpoint
     endpoint: process.env.VOICE_WS_ENDPOINT,
+    // Tools
     tools: {
-        getWeather: weatherTool, // Pass the AI SDK tool directly
+        getWeather: weatherTool,
+        getTime: timeTool,
     },
 });
 
 // 3. Handle Events
-agent.on("connected", () => console.log("Connected to WebSocket"));
 
-// Handle incoming audio from AI (play this to user)
+// Connection events
-agent.on("audio", (base64Audio: string) => {
+agent.on("connected", () => console.log("✓ Connected to WebSocket"));
-    // process.stdout.write(Buffer.from(base64Audio, 'base64')); 
+agent.on("disconnected", () => console.log("✗ Disconnected from WebSocket"));
+
+// Transcription events (when audio is converted to text)
+agent.on("transcription", ({ text, language }: { text: string; language?: string }) => {
+    console.log(`[Transcription] (${language || "unknown"}): ${text}`);
 });
 
-// Logs
+// Text events (user input and assistant responses)
-agent.on("text", (msg: { role: string; text: string }) => console.log(`${msg.role}: ${msg.text}`));
+agent.on("text", (msg: { role: string; text: string }) => {
-agent.on("tool_start", ({ name }: { name: string }) => console.log(`[System] Calling ${name}...`));
+    const prefix = msg.role === "user" ? "👤 User" : "🤖 Assistant";
+    console.log(`${prefix}: ${msg.text}`);
+});
 
-// 4. Start (wrap in async function since we can't use top-level await)
+// Streaming text delta events (real-time text chunks)
+agent.on("text_delta", ({ text }: { text: string }) => {
+    process.stdout.write(text);
+});
+
+// Tool events
+agent.on("tool_start", ({ name, input }: { name: string; input?: unknown }) => {
+    console.log(`\n[Tool] Calling ${name}...`, input ? JSON.stringify(input) : "");
+});
+
+agent.on("tool_result", ({ name, result }: { name: string; result: unknown }) => {
+    console.log(`[Tool] ${name} result:`, JSON.stringify(result));
+});
+
+// Speech events
+agent.on("speech_start", ({ text }: { text: string }) => {
+    console.log(`[TTS] Generating speech for: "${text.substring(0, 50)}..."`);
+});
+
+agent.on("speech_complete", () => {
+    console.log("[TTS] Speech generation complete");
+});
+
+// Audio events (when TTS audio is generated)
+agent.on("audio", async (audio: { data: string; format: string; uint8Array: Uint8Array }) => {
+    console.log(`[Audio] Received ${audio.format} audio (${audio.uint8Array.length} bytes)`);
+    // Optionally save to file for testing
+    await writeFile(`output.${audio.format}`, Buffer.from(audio.uint8Array));
+});
+
+// Error handling
+agent.on("error", (error: Error) => {
+    console.error("[Error]", error.message);
+});
+
+// 4. Main execution
 (async () => {
-    try {
+    console.log("\n=== Voice Agent Demo ===");
-        // For now: text-only sanity check, no voice pipeline required.
+    console.log("Testing text-only mode (no WebSocket required)\n");
-        await agent.sendText("What is the weather in Berlin?");
 
-        // Optional: connect only when an endpoint is provided.
+    try {
+        // Test 1: Simple text query with streaming
+        console.log("--- Test 1: Weather Query ---");
+        const response1 = await agent.sendText("What is the weather in Berlin?");
+        console.log("\n");
+
+        // Test 2: Multi-turn conversation
+        console.log("--- Test 2: Follow-up Question ---");
+        const response2 = await agent.sendText("What about Tokyo?");
+        console.log("\n");
+
+        // Test 3: Time query
+        console.log("--- Test 3: Time Query ---");
+        const response3 = await agent.sendText("What time is it?");
+        console.log("\n");
+
+        // Show conversation history
+        console.log("--- Conversation History ---");
+        const history = agent.getHistory();
+        console.log(`Total messages: ${history.length}`);
+
+        // Optional: Connect to WebSocket for real-time voice
         if (process.env.VOICE_WS_ENDPOINT) {
+            console.log("\n--- Connecting to WebSocket ---");
             await agent.connect(process.env.VOICE_WS_ENDPOINT);
-            console.log("Agent connected successfully");
+            console.log("Agent connected. Listening for audio input...");
+
+            // Keep the process running to receive WebSocket messages
+            // In a real app, you would stream microphone audio here
         }
     } catch (error) {
         console.error("Agent run failed:", error);
+        process.exit(1);
     }
 })();
 
-// 5. Simulate sending audio (in a real app, stream microphone data here)
+// Example: How to send audio in a real application
-// agent.sendAudio("Base64EncodedPCM16AudioData...");
+// ---------------------------------------------
+// import { readFile } from "fs/promises";
+//
+// // Option 1: Send base64 encoded audio
+// const audioBase64 = (await readFile("recording.mp3")).toString("base64");
+// await agent.sendAudio(audioBase64);
+//
+// // Option 2: Send raw audio buffer
+// const audioBuffer = await readFile("recording.mp3");
+// await agent.sendAudioBuffer(audioBuffer);
+//
+// // Option 3: Transcribe audio directly
+// const transcribedText = await agent.transcribeAudio(audioBuffer);
+// console.log("Transcribed:", transcribedText);

example/ws-server.ts

@@ -1,22 +1,73 @@
 import "dotenv/config";
-import { WebSocketServer } from "ws";
+import { WebSocketServer, WebSocket } from "ws";
+import { readFile } from "fs/promises";
+import { existsSync } from "fs";
 
 const endpoint = process.env.VOICE_WS_ENDPOINT || "ws://localhost:8080";
 const url = new URL(endpoint);
 const port = Number(url.port || 8080);
 const host = url.hostname || "localhost";
 
+// Message types for type safety
+interface BaseMessage {
+    type: string;
+}
+
+interface TextDeltaMessage extends BaseMessage {
+    type: "text_delta";
+    text: string;
+}
+
+interface ToolCallMessage extends BaseMessage {
+    type: "tool_call";
+    toolName: string;
+    toolCallId: string;
+    input: unknown;
+}
+
+interface ToolResultMessage extends BaseMessage {
+    type: "tool_result";
+    toolName: string;
+    toolCallId: string;
+    result: unknown;
+}
+
+interface AudioMessage extends BaseMessage {
+    type: "audio";
+    data: string; // base64 encoded
+    format: string;
+}
+
+interface ResponseCompleteMessage extends BaseMessage {
+    type: "response_complete";
+    text: string;
+    toolCalls: Array<{ toolName: string; toolCallId: string; input: unknown }>;
+    toolResults: Array<{ toolName: string; toolCallId: string; output: unknown }>;
+}
+
+type AgentMessage =
+    | TextDeltaMessage
+    | ToolCallMessage
+    | ToolResultMessage
+    | AudioMessage
+    | ResponseCompleteMessage;
+
 const wss = new WebSocketServer({ port, host });
 
 wss.on("listening", () => {
-    console.log(`[ws-server] listening on ${endpoint}`);
+    console.log(`[ws-server] 🚀 listening on ${endpoint}`);
+    console.log("[ws-server] Waiting for connections...\n");
 });
 
-wss.on("connection", (socket) => {
+wss.on("connection", (socket: WebSocket) => {
-    console.log("[ws-server] client connected");
+    console.log("[ws-server] ✓ client connected");
+
+    let streamingText = "";
+    let audioChunks: Buffer[] = [];
 
     // Send a sample transcript to test text pipeline end-to-end.
     setTimeout(() => {
+        console.log("[ws-server] -> Sending test transcript...");
         socket.send(
             JSON.stringify({
                 type: "transcript",
@@ -25,19 +76,94 @@ wss.on("connection", (socket) => {
         );
     }, 500);
 
-    socket.on("message", (data) => {
+    socket.on("message", async (data) => {
         try {
-            const msg = JSON.parse(data.toString()) as {
+            const msg = JSON.parse(data.toString()) as AgentMessage;
-                type?: string;
+
-                text?: string;
+            switch (msg.type) {
-            };
+                case "text_delta":
-            console.log("[ws-server] <-", msg);
+                    // Real-time streaming text from the agent
+                    streamingText += msg.text;
+                    process.stdout.write(msg.text);
+                    break;
+
+                case "tool_call":
+                    console.log(`\n[ws-server] 🛠️ Tool call: ${msg.toolName}`);
+                    console.log(`           Input: ${JSON.stringify(msg.input)}`);
+                    break;
+
+                case "tool_result":
+                    console.log(`[ws-server] 🛠️ Tool result: ${msg.toolName}`);
+                    console.log(`           Result: ${JSON.stringify(msg.result)}`);
+                    break;
+
+                case "audio":
+                    // Handle audio response from TTS
+                    const audioBuffer = Buffer.from(msg.data, "base64");
+                    audioChunks.push(audioBuffer);
+                    console.log(
+                        `[ws-server] 🔊 Received audio: ${audioBuffer.length} bytes (${msg.format})`,
+                    );
+
+                    // Optionally save audio to file for testing
+                    // await writeFile(`output_${Date.now()}.${msg.format}`, audioBuffer);
+                    break;
+
+                case "response_complete":
+                    console.log("\n[ws-server] ✅ Response complete");
+                    console.log(`           Text length: ${msg.text.length}`);
+                    console.log(`           Tool calls: ${msg.toolCalls.length}`);
+                    console.log(`           Tool results: ${msg.toolResults.length}`);
+
+                    // Reset for next response
+                    streamingText = "";
+                    audioChunks = [];
+                    break;
+
+                default:
+                    console.log("[ws-server] <- Unknown message:", msg);
+            }
         } catch {
-            console.log("[ws-server] <- raw", data.toString());
+            console.log("[ws-server] <- raw", data.toString().substring(0, 100));
         }
     });
 
     socket.on("close", () => {
-        console.log("[ws-server] client disconnected");
+        console.log("[ws-server] ✗ client disconnected\n");
+    });
+
+    socket.on("error", (error) => {
+        console.error("[ws-server] Error:", error.message);
     });
 });
+
+// Graceful shutdown
+process.on("SIGINT", () => {
+    console.log("\n[ws-server] Shutting down...");
+    wss.close(() => {
+        console.log("[ws-server] Server closed");
+        process.exit(0);
+    });
+});
+
+// Helper function to simulate sending audio to the agent
+async function simulateAudioInput(socket: WebSocket, audioPath: string) {
+    if (!existsSync(audioPath)) {
+        console.log(`[ws-server] Audio file not found: ${audioPath}`);
+        return;
+    }
+
+    const audioBuffer = await readFile(audioPath);
+    const base64Audio = audioBuffer.toString("base64");
+
+    console.log(`[ws-server] -> Sending audio: ${audioPath} (${audioBuffer.length} bytes)`);
+    socket.send(
+        JSON.stringify({
+            type: "audio",
+            data: base64Audio,
+        }),
+    );
+}
+
+// Export for use as a module
+export { wss, simulateAudioInput };

src/VoiceAgent.ts

@@ -1,31 +1,58 @@
 import { WebSocket } from "ws";
 import { EventEmitter } from "events";
-import { generateText, LanguageModel, stepCountIs, type Tool } from "ai";
+import {
+  streamText,
+  LanguageModel,
+  stepCountIs,
+  type Tool,
+  type ModelMessage,
+  experimental_transcribe as transcribe,
+  experimental_generateSpeech as generateSpeech,
+  type TranscriptionModel,
+  type SpeechModel,
+} from "ai";
 
 export interface VoiceAgentOptions {
-  model: LanguageModel; /// AI SDK Model (e.g., openai('gpt-4o'))
+  model: LanguageModel; // AI SDK Model for chat (e.g., openai('gpt-4o'))
+  transcriptionModel?: TranscriptionModel; // AI SDK Transcription Model (e.g., openai.transcription('whisper-1'))
+  speechModel?: SpeechModel; // AI SDK Speech Model (e.g., openai.speech('gpt-4o-mini-tts'))
   instructions?: string;
-  stopWhen?: NonNullable<Parameters<typeof generateText>[0]["stopWhen"]>;
+  stopWhen?: NonNullable<Parameters<typeof streamText>[0]["stopWhen"]>;
   tools?: Record<string, Tool>;
   endpoint?: string;
+  voice?: string; // Voice for TTS (e.g., 'alloy', 'echo', 'shimmer')
+  speechInstructions?: string; // Instructions for TTS voice style
+  outputFormat?: string; // Audio output format (e.g., 'mp3', 'opus', 'wav')
 }
 
 export class VoiceAgent extends EventEmitter {
   private socket?: WebSocket;
   private tools: Record<string, Tool> = {};
   private model: LanguageModel;
+  private transcriptionModel?: TranscriptionModel;
+  private speechModel?: SpeechModel;
   private instructions: string;
-  private stopWhen: NonNullable<Parameters<typeof generateText>[0]["stopWhen"]>;
+  private stopWhen: NonNullable<Parameters<typeof streamText>[0]["stopWhen"]>;
   private endpoint?: string;
   private isConnected = false;
+  private conversationHistory: ModelMessage[] = [];
+  private voice: string;
+  private speechInstructions?: string;
+  private outputFormat: string;
+  private isProcessing = false;
 
   constructor(options: VoiceAgentOptions) {
     super();
     this.model = options.model;
+    this.transcriptionModel = options.transcriptionModel;
+    this.speechModel = options.speechModel;
     this.instructions =
       options.instructions || "You are a helpful voice assistant.";
     this.stopWhen = options.stopWhen || stepCountIs(5);
     this.endpoint = options.endpoint;
+    this.voice = options.voice || "alloy";
+    this.speechInstructions = options.speechInstructions;
+    this.outputFormat = options.outputFormat || "mp3";
     if (options.tools) {
       this.tools = { ...options.tools };
     }
@@ -38,16 +65,17 @@ export class VoiceAgent extends EventEmitter {
       try {
         const message = JSON.parse(data.toString());
 
-        // Example: Handle transcribed text from the client/STT
+        // Handle transcribed text from the client/STT
         if (message.type === "transcript") {
           await this.processUserInput(message.text);
         }
-        // Handle audio data
+        // Handle raw audio data that needs transcription
         if (message.type === "audio") {
-          this.emit("audio", message.data);
+          await this.processAudioInput(message.data);
         }
       } catch (err) {
         console.error("Failed to process message:", err);
+        this.emit("error", err);
       }
     });
 
@@ -56,12 +84,81 @@ export class VoiceAgent extends EventEmitter {
       this.isConnected = false;
       this.emit("disconnected");
     });
+
+    this.socket.on("error", (error) => {
+      console.error("WebSocket error:", error);
+      this.emit("error", error);
+    });
   }
 
   public registerTools(tools: Record<string, Tool>) {
     this.tools = { ...this.tools, ...tools };
   }
 
+  /**
+   * Transcribe audio data to text using the configured transcription model
+   */
+  public async transcribeAudio(audioData: Buffer | Uint8Array): Promise<string> {
+    if (!this.transcriptionModel) {
+      throw new Error("Transcription model not configured");
+    }
+
+    const result = await transcribe({
+      model: this.transcriptionModel,
+      audio: audioData,
+    });
+
+    this.emit("transcription", {
+      text: result.text,
+      language: result.language,
+    });
+
+    return result.text;
+  }
+
+  /**
+   * Generate speech from text using the configured speech model
+   */
+  public async generateSpeechFromText(text: string): Promise<Uint8Array> {
+    if (!this.speechModel) {
+      throw new Error("Speech model not configured");
+    }
+
+    const result = await generateSpeech({
+      model: this.speechModel,
+      text,
+      voice: this.voice,
+      instructions: this.speechInstructions,
+      outputFormat: this.outputFormat,
+    });
+
+    return result.audio.uint8Array;
+  }
+
+  /**
+   * Process incoming audio data: transcribe and generate response
+   */
+  private async processAudioInput(base64Audio: string): Promise<void> {
+    if (!this.transcriptionModel) {
+      this.emit("error", new Error("Transcription model not configured for audio input"));
+      return;
+    }
+
+    try {
+      const audioBuffer = Buffer.from(base64Audio, "base64");
+      this.emit("audio_received", { size: audioBuffer.length });
+
+      const transcribedText = await this.transcribeAudio(audioBuffer);
+
+      if (transcribedText.trim()) {
+        await this.processUserInput(transcribedText);
+      }
+    } catch (error) {
+      console.error("Failed to process audio input:", error);
+      this.emit("error", error);
+    }
+  }
+
   public async connect(url?: string): Promise<void> {
     return new Promise((resolve, reject) => {
       try {
@@ -85,52 +182,426 @@ export class VoiceAgent extends EventEmitter {
     });
   }
 
-  public async sendText(text: string): Promise<void> {
+  /**
-    await this.processUserInput(text);
+   * Send text input for processing (bypasses transcription)
+   */
+  public async sendText(text: string): Promise<string> {
+    return this.processUserInput(text);
   }
 
-  public sendAudio(audioData: string): void {
+  /**
-    if (this.socket && this.isConnected) {
+   * Send audio data to be transcribed and processed
-      this.socket.send(JSON.stringify({
+   * @param audioData Base64 encoded audio data
-        type: "audio",
+   */
-        data: audioData
+  public async sendAudio(audioData: string): Promise<void> {
-      }));
+    await this.processAudioInput(audioData);
-    }
   }
 
-  private async processUserInput(text: string) {
+  /**
+   * Send raw audio buffer to be transcribed and processed
+   */
+  public async sendAudioBuffer(audioBuffer: Buffer | Uint8Array): Promise<void> {
+    const base64Audio = Buffer.from(audioBuffer).toString("base64");
+    await this.processAudioInput(base64Audio);
+  }
+
+  /**
+   * Process user input with streaming text generation
+   * Handles the full pipeline: text -> LLM (streaming) -> TTS -> WebSocket
+   */
+  private async processUserInput(text: string): Promise<string> {
+    if (this.isProcessing) {
+      this.emit("warning", "Already processing a request, queuing...");
+    }
+    this.isProcessing = true;
+
+    try {
       // Emit text event for incoming user input
       this.emit("text", { role: "user", text });
 
-    const result = await generateText({
+      // Add user message to conversation history
+      this.conversationHistory.push({ role: "user", content: text });
+
+      // Use streamText for streaming responses with tool support
+      const result = streamText({
         model: this.model,
         system: this.instructions,
-      prompt: text,
+        messages: this.conversationHistory,
         tools: this.tools,
         stopWhen: this.stopWhen,
+        onChunk: ({ chunk }) => {
+          // Emit streaming chunks for real-time updates
+          // Note: onChunk only receives a subset of stream events
+          switch (chunk.type) {
+            case "text-delta":
+              this.emit("chunk:text_delta", { id: chunk.id, text: chunk.text });
+              break;
+
+            case "reasoning-delta":
+              this.emit("chunk:reasoning_delta", { id: chunk.id, text: chunk.text });
+              break;
+
+            case "tool-call":
+              this.emit("chunk:tool_call", {
+                toolName: chunk.toolName,
+                toolCallId: chunk.toolCallId,
+                input: chunk.input,
+              });
+              break;
+
+            case "tool-result":
+              this.emit("chunk:tool_result", {
+                toolName: chunk.toolName,
+                toolCallId: chunk.toolCallId,
+                result: chunk.output,
+              });
+              break;
+
+            case "tool-input-start":
+              this.emit("chunk:tool_input_start", {
+                id: chunk.id,
+                toolName: chunk.toolName,
+              });
+              break;
+
+            case "tool-input-delta":
+              this.emit("chunk:tool_input_delta", {
+                id: chunk.id,
+                delta: chunk.delta,
+              });
+              break;
+
+            case "source":
+              this.emit("chunk:source", chunk);
+              break;
+          }
+        },
+        onFinish: async (event) => {
+          // Process steps for tool results
+          for (const step of event.steps) {
+            for (const toolResult of step.toolResults) {
+              this.emit("tool_result", {
+                name: toolResult.toolName,
+                toolCallId: toolResult.toolCallId,
+                result: toolResult.output,
+              });
+            }
+          }
+        },
+        onError: ({ error }) => {
+          console.error("Stream error:", error);
+          this.emit("error", error);
+        },
       });
 
-    for (const toolCall of result.toolCalls ?? []) {
+      // Collect the full response text and reasoning
-      this.emit("tool_start", { name: toolCall.toolName });
+      let fullText = "";
+      let fullReasoning = "";
+      const allToolCalls: Array<{
+        toolName: string;
+        toolCallId: string;
+        input: unknown;
+      }> = [];
+      const allToolResults: Array<{
+        toolName: string;
+        toolCallId: string;
+        output: unknown;
+      }> = [];
+      const allSources: Array<unknown> = [];
+      const allFiles: Array<unknown> = [];
+
+      // Process the full stream
+      for await (const part of result.fullStream) {
+        switch (part.type) {
+          // Stream lifecycle
+          case "start":
+            this.sendWebSocketMessage({ type: "stream_start" });
+            break;
+
+          case "finish":
+            this.emit("text", { role: "assistant", text: fullText });
+            this.sendWebSocketMessage({
+              type: "stream_finish",
+              finishReason: part.finishReason,
+              usage: part.totalUsage,
+            });
+            break;
+
+          case "error":
+            this.emit("error", part.error);
+            this.sendWebSocketMessage({
+              type: "stream_error",
+              error: String(part.error),
+            });
+            break;
+
+          case "abort":
+            this.emit("abort", { reason: part.reason });
+            this.sendWebSocketMessage({
+              type: "stream_abort",
+              reason: part.reason,
+            });
+            break;
+
+          // Step lifecycle
+          case "start-step":
+            this.sendWebSocketMessage({
+              type: "step_start",
+              warnings: part.warnings,
+            });
+            break;
+
+          case "finish-step":
+            this.sendWebSocketMessage({
+              type: "step_finish",
+              finishReason: part.finishReason,
+              usage: part.usage,
+            });
+            break;
+
+          // Text streaming
+          case "text-start":
+            this.sendWebSocketMessage({ type: "text_start", id: part.id });
+            break;
+
+          case "text-delta":
+            fullText += part.text;
+            this.sendWebSocketMessage({
+              type: "text_delta",
+              id: part.id,
+              text: part.text,
+            });
+            break;
+
+          case "text-end":
+            this.sendWebSocketMessage({ type: "text_end", id: part.id });
+            break;
+
+          // Reasoning streaming (for models that support it)
+          case "reasoning-start":
+            this.sendWebSocketMessage({ type: "reasoning_start", id: part.id });
+            break;
+
+          case "reasoning-delta":
+            fullReasoning += part.text;
+            this.sendWebSocketMessage({
+              type: "reasoning_delta",
+              id: part.id,
+              text: part.text,
+            });
+            break;
+
+          case "reasoning-end":
+            this.sendWebSocketMessage({ type: "reasoning_end", id: part.id });
+            break;
+
+          // Tool input streaming
+          case "tool-input-start":
+            this.sendWebSocketMessage({
+              type: "tool_input_start",
+              id: part.id,
+              toolName: part.toolName,
+            });
+            break;
+
+          case "tool-input-delta":
+            this.sendWebSocketMessage({
+              type: "tool_input_delta",
+              id: part.id,
+              delta: part.delta,
+            });
+            break;
+
+          case "tool-input-end":
+            this.sendWebSocketMessage({ type: "tool_input_end", id: part.id });
+            break;
+
+          // Tool execution
+          case "tool-call":
+            allToolCalls.push({
+              toolName: part.toolName,
+              toolCallId: part.toolCallId,
+              input: part.input,
+            });
+            this.sendWebSocketMessage({
+              type: "tool_call",
+              toolName: part.toolName,
+              toolCallId: part.toolCallId,
+              input: part.input,
+            });
+            break;
+
+          case "tool-result":
+            allToolResults.push({
+              toolName: part.toolName,
+              toolCallId: part.toolCallId,
+              output: part.output,
+            });
+            this.sendWebSocketMessage({
+              type: "tool_result",
+              toolName: part.toolName,
+              toolCallId: part.toolCallId,
+              result: part.output,
+            });
+            break;
+
+          case "tool-error":
+            this.sendWebSocketMessage({
+              type: "tool_error",
+              toolName: part.toolName,
+              toolCallId: part.toolCallId,
+              error: String(part.error),
+            });
+            break;
+
+          // Sources and files
+          case "source":
+            allSources.push(part);
+            this.sendWebSocketMessage({
+              type: "source",
+              source: part,
+            });
+            break;
+
+          case "file":
+            allFiles.push(part.file);
+            this.sendWebSocketMessage({
+              type: "file",
+              file: part.file,
+            });
+            break;
+        }
       }
 
-    // Emit text event for assistant response
+      // Add assistant response to conversation history
-    this.emit("text", { role: "assistant", text: result.text });
+      if (fullText) {
+        this.conversationHistory.push({ role: "assistant", content: fullText });
+      }
 
-    // Send the response back (either text to be TTSed or tool results)
+      // Generate speech from the response if speech model is configured
+      if (this.speechModel && fullText) {
+        await this.generateAndSendSpeech(fullText);
+      }
+
+      // Send the complete response
+      this.sendWebSocketMessage({
+        type: "response_complete",
+        text: fullText,
+        reasoning: fullReasoning || undefined,
+        toolCalls: allToolCalls,
+        toolResults: allToolResults,
+        sources: allSources.length > 0 ? allSources : undefined,
+        files: allFiles.length > 0 ? allFiles : undefined,
+      });
+
+      return fullText;
+    } finally {
+      this.isProcessing = false;
+    }
+  }
+
+  /**
+   * Generate speech and send audio via WebSocket
+   */
+  private async generateAndSendSpeech(text: string): Promise<void> {
+    if (!this.speechModel) return;
+
+    try {
+      this.emit("speech_start", { text });
+
+      const audioData = await this.generateSpeechFromText(text);
+      const base64Audio = Buffer.from(audioData).toString("base64");
+
+      // Send audio via WebSocket
+      this.sendWebSocketMessage({
+        type: "audio",
+        data: base64Audio,
+        format: this.outputFormat,
+      });
+
+      // Also emit for local handling
+      this.emit("audio", {
+        data: base64Audio,
+        format: this.outputFormat,
+        uint8Array: audioData,
+      });
+
+      this.emit("speech_complete", { text });
+    } catch (error) {
+      console.error("Failed to generate speech:", error);
+      this.emit("error", error);
+    }
+  }
+
+  /**
+   * Send a message via WebSocket if connected
+   */
+  private sendWebSocketMessage(message: Record<string, unknown>): void {
     if (this.socket && this.isConnected) {
-      this.socket.send(
+      this.socket.send(JSON.stringify(message));
-        JSON.stringify({
-          type: "response",
-          text: result.text,
-          toolCalls: result.toolCalls,
-          toolResults: result.toolResults,
-        }),
-      );
     }
   }
 
+  /**
+   * Start listening for voice input
+   */
   startListening() {
     console.log("Starting voice agent...");
+    this.emit("listening");
+  }
+
+  /**
+   * Stop listening for voice input
+   */
+  stopListening() {
+    console.log("Stopping voice agent...");
+    this.emit("stopped");
+  }
+
+  /**
+   * Clear conversation history
+   */
+  clearHistory() {
+    this.conversationHistory = [];
+    this.emit("history_cleared");
+  }
+
+  /**
+   * Get current conversation history
+   */
+  getHistory(): ModelMessage[] {
+    return [...this.conversationHistory];
+  }
+
+  /**
+   * Set conversation history (useful for restoring sessions)
+   */
+  setHistory(history: ModelMessage[]) {
+    this.conversationHistory = [...history];
+  }
+
+  /**
+   * Disconnect from WebSocket
+   */
+  disconnect() {
+    if (this.socket) {
+      this.socket.close();
+      this.socket = undefined;
+      this.isConnected = false;
+    }
+  }
+
+  /**
+   * Check if agent is connected to WebSocket
+   */
+  get connected(): boolean {
+    return this.isConnected;
+  }
+
+  /**
+   * Check if agent is currently processing a request
+   */
+  get processing(): boolean {
+    return this.isProcessing;
   }
 }

src/index.ts

@@ -1 +1 @@
-export { VoiceAgent } from "./VoiceAgent";
+export { VoiceAgent, type VoiceAgentOptions } from "./VoiceAgent";