@livekit/agents 1.0.34 → 1.0.36

package/dist/voice/testing/types.d.ts

@@ -0,0 +1,83 @@
+import type { AgentHandoffItem, ChatMessage, ChatRole, FunctionCall, FunctionCallOutput } from '../../llm/chat_context.js';
+import type { Agent } from '../agent.js';
+/**
+ * Event representing an assistant or user message in the conversation.
+ */
+export interface ChatMessageEvent {
+    type: 'message';
+    item: ChatMessage;
+}
+/**
+ * Event representing a function/tool call initiated by the LLM.
+ */
+export interface FunctionCallEvent {
+    type: 'function_call';
+    item: FunctionCall;
+}
+/**
+ * Event representing the output/result of a function call.
+ */
+export interface FunctionCallOutputEvent {
+    type: 'function_call_output';
+    item: FunctionCallOutput;
+}
+/**
+ * Event representing an agent handoff (switching from one agent to another).
+ */
+export interface AgentHandoffEvent {
+    type: 'agent_handoff';
+    item: AgentHandoffItem;
+    oldAgent?: Agent;
+    newAgent: Agent;
+}
+/**
+ * Union type of all possible run events that can occur during a test run.
+ */
+export type RunEvent = ChatMessageEvent | FunctionCallEvent | FunctionCallOutputEvent | AgentHandoffEvent;
+/**
+ * Type guard to check if an event is a ChatMessageEvent.
+ */
+export declare function isChatMessageEvent(event: RunEvent): event is ChatMessageEvent;
+/**
+ * Type guard to check if an event is a FunctionCallEvent.
+ */
+export declare function isFunctionCallEvent(event: RunEvent): event is FunctionCallEvent;
+/**
+ * Type guard to check if an event is a FunctionCallOutputEvent.
+ */
+export declare function isFunctionCallOutputEvent(event: RunEvent): event is FunctionCallOutputEvent;
+/**
+ * Type guard to check if an event is an AgentHandoffEvent.
+ */
+export declare function isAgentHandoffEvent(event: RunEvent): event is AgentHandoffEvent;
+/**
+ * Options for message assertion.
+ */
+export interface MessageAssertOptions {
+    role?: ChatRole;
+}
+/**
+ * Options for function call assertion.
+ */
+export interface FunctionCallAssertOptions {
+    name?: string;
+    args?: Record<string, unknown>;
+}
+/**
+ * Options for function call output assertion.
+ */
+export interface FunctionCallOutputAssertOptions {
+    output?: string;
+    isError?: boolean;
+}
+/**
+ * Options for agent handoff assertion.
+ */
+export interface AgentHandoffAssertOptions {
+    newAgentType?: new (...args: any[]) => Agent;
+}
+/**
+ * Event type literals for type-safe event filtering.
+ */
+export type EventType = 'message' | 'function_call' | 'function_call_output' | 'agent_handoff';
+//# sourceMappingURL=types.d.ts.map

package/dist/voice/testing/types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../../src/voice/testing/types.ts"],"names":[],"mappings":"AAGA,OAAO,KAAK,EACV,gBAAgB,EAChB,WAAW,EACX,QAAQ,EACR,YAAY,EACZ,kBAAkB,EACnB,MAAM,2BAA2B,CAAC;AACnC,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAEzC;;GAEG;AACH,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,SAAS,CAAC;IAChB,IAAI,EAAE,WAAW,CAAC;CACnB;AAED;;GAEG;AACH,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,eAAe,CAAC;IACtB,IAAI,EAAE,YAAY,CAAC;CACpB;AAED;;GAEG;AACH,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,sBAAsB,CAAC;IAC7B,IAAI,EAAE,kBAAkB,CAAC;CAC1B;AAED;;GAEG;AACH,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,eAAe,CAAC;IACtB,IAAI,EAAE,gBAAgB,CAAC;IACvB,QAAQ,CAAC,EAAE,KAAK,CAAC;IACjB,QAAQ,EAAE,KAAK,CAAC;CACjB;AAED;;GAEG;AACH,MAAM,MAAM,QAAQ,GAChB,gBAAgB,GAChB,iBAAiB,GACjB,uBAAuB,GACvB,iBAAiB,CAAC;AAEtB;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,QAAQ,GAAG,KAAK,IAAI,gBAAgB,CAE7E;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,QAAQ,GAAG,KAAK,IAAI,iBAAiB,CAE/E;AAED;;GAEG;AACH,wBAAgB,yBAAyB,CAAC,KAAK,EAAE,QAAQ,GAAG,KAAK,IAAI,uBAAuB,CAE3F;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,QAAQ,GAAG,KAAK,IAAI,iBAAiB,CAE/E;AAED;;GAEG;AACH,MAAM,WAAW,oBAAoB;IACnC,IAAI,CAAC,EAAE,QAAQ,CAAC;CACjB;AAED;;GAEG;AACH,MAAM,WAAW,yBAAyB;IACxC,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAChC;AAED;;GAEG;AACH,MAAM,WAAW,+BAA+B;IAC9C,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,OAAO,CAAC;CACnB;AAED;;GAEG;AACH,MAAM,WAAW,yBAAyB;IAExC,YAAY,CAAC,EAAE,KAAK,GAAG,IAAI,EAAE,GAAG,EAAE,KAAK,KAAK,CAAC;CAC9C;AAED;;GAEG;AACH,MAAM,MAAM,SAAS,GAAG,SAAS,GAAG,eAAe,GAAG,sBAAsB,GAAG,eAAe,CAAC"}

package/dist/voice/testing/types.js

@@ -0,0 +1,19 @@
+function isChatMessageEvent(event) {
+  return event.type === "message";
+}
+function isFunctionCallEvent(event) {
+  return event.type === "function_call";
+}
+function isFunctionCallOutputEvent(event) {
+  return event.type === "function_call_output";
+}
+function isAgentHandoffEvent(event) {
+  return event.type === "agent_handoff";
+}
+export {
+  isAgentHandoffEvent,
+  isChatMessageEvent,
+  isFunctionCallEvent,
+  isFunctionCallOutputEvent
+};
+//# sourceMappingURL=types.js.map

package/dist/voice/testing/types.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/voice/testing/types.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport type {\n  AgentHandoffItem,\n  ChatMessage,\n  ChatRole,\n  FunctionCall,\n  FunctionCallOutput,\n} from '../../llm/chat_context.js';\nimport type { Agent } from '../agent.js';\n\n/**\n * Event representing an assistant or user message in the conversation.\n */\nexport interface ChatMessageEvent {\n  type: 'message';\n  item: ChatMessage;\n}\n\n/**\n * Event representing a function/tool call initiated by the LLM.\n */\nexport interface FunctionCallEvent {\n  type: 'function_call';\n  item: FunctionCall;\n}\n\n/**\n * Event representing the output/result of a function call.\n */\nexport interface FunctionCallOutputEvent {\n  type: 'function_call_output';\n  item: FunctionCallOutput;\n}\n\n/**\n * Event representing an agent handoff (switching from one agent to another).\n */\nexport interface AgentHandoffEvent {\n  type: 'agent_handoff';\n  item: AgentHandoffItem;\n  oldAgent?: Agent;\n  newAgent: Agent;\n}\n\n/**\n * Union type of all possible run events that can occur during a test run.\n */\nexport type RunEvent =\n  | ChatMessageEvent\n  | FunctionCallEvent\n  | FunctionCallOutputEvent\n  | AgentHandoffEvent;\n\n/**\n * Type guard to check if an event is a ChatMessageEvent.\n */\nexport function isChatMessageEvent(event: RunEvent): event is ChatMessageEvent {\n  return event.type === 'message';\n}\n\n/**\n * Type guard to check if an event is a FunctionCallEvent.\n */\nexport function isFunctionCallEvent(event: RunEvent): event is FunctionCallEvent {\n  return event.type === 'function_call';\n}\n\n/**\n * Type guard to check if an event is a FunctionCallOutputEvent.\n */\nexport function isFunctionCallOutputEvent(event: RunEvent): event is FunctionCallOutputEvent {\n  return event.type === 'function_call_output';\n}\n\n/**\n * Type guard to check if an event is an AgentHandoffEvent.\n */\nexport function isAgentHandoffEvent(event: RunEvent): event is AgentHandoffEvent {\n  return event.type === 'agent_handoff';\n}\n\n/**\n * Options for message assertion.\n */\nexport interface MessageAssertOptions {\n  role?: ChatRole;\n}\n\n/**\n * Options for function call assertion.\n */\nexport interface FunctionCallAssertOptions {\n  name?: string;\n  args?: Record<string, unknown>;\n}\n\n/**\n * Options for function call output assertion.\n */\nexport interface FunctionCallOutputAssertOptions {\n  output?: string;\n  isError?: boolean;\n}\n\n/**\n * Options for agent handoff assertion.\n */\nexport interface AgentHandoffAssertOptions {\n  // eslint-disable-next-line @typescript-eslint/no-explicit-any\n  newAgentType?: new (...args: any[]) => Agent;\n}\n\n/**\n * Event type literals for type-safe event filtering.\n */\nexport type EventType = 'message' | 'function_call' | 'function_call_output' | 'agent_handoff';\n"],"mappings":"AA0DO,SAAS,mBAAmB,OAA4C;AAC7E,SAAO,MAAM,SAAS;AACxB;AAKO,SAAS,oBAAoB,OAA6C;AAC/E,SAAO,MAAM,SAAS;AACxB;AAKO,SAAS,0BAA0B,OAAmD;AAC3F,SAAO,MAAM,SAAS;AACxB;AAKO,SAAS,oBAAoB,OAA6C;AAC/E,SAAO,MAAM,SAAS;AACxB;","names":[]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@livekit/agents",
-  "version": "1.0.34",
+  "version": "1.0.36",
   "description": "LiveKit Agents - Node.js",
   "main": "dist/index.js",
   "require": "dist/index.cjs",
@@ -26,7 +26,7 @@
     "README.md"
   ],
   "devDependencies": {
-    "@livekit/rtc-node": "^0.13.22",
+    "@livekit/rtc-node": "^0.13.24",
     "@microsoft/api-extractor": "^7.35.0",
     "@types/fluent-ffmpeg": "^2.1.28",
     "@types/json-schema": "^7.0.15",
@@ -70,7 +70,7 @@
     "zod-to-json-schema": "^3.24.6"
   },
   "peerDependencies": {
-    "@livekit/rtc-node": "^0.13.22",
+    "@livekit/rtc-node": "^0.13.24",
     "zod": "^3.25.76 || ^4.1.8"
   },
   "scripts": {

package/src/inference/llm.ts

@@ -27,7 +27,14 @@ export type OpenAIModels =
   | 'openai/gpt-4o-mini'
   | 'openai/gpt-oss-120b';
 
-export type GoogleModels = 'google/gemini-2.0-flash-lite';
+export type GoogleModels =
+  | 'google/gemini-3-pro-preview'
+  | 'google/gemini-3-flash-preview'
+  | 'google/gemini-2.5-pro'
+  | 'google/gemini-2.5-flash'
+  | 'google/gemini-2.5-flash-lite'
+  | 'google/gemini-2.0-flash'
+  | 'google/gemini-2.0-flash-lite';
 
 export type QwenModels = 'qwen/qwen3-235b-a22b-instruct';
 
@@ -235,6 +242,7 @@ export class LLMStream extends llm.LLMStream {
   private toolIndex?: number;
   private fncName?: string;
   private fncRawArguments?: string;
+  private toolExtra?: Record<string, unknown>;
 
   constructor(
     llm: LLM,
@@ -277,6 +285,7 @@ export class LLMStream extends llm.LLMStream {
     // (defined inside the run method to make sure the state is reset for each run/attempt)
     let retryable = true;
     this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;
+    this.toolExtra = undefined;
 
     try {
       const messages = (await this.chatCtx.toProviderFormat(
@@ -428,6 +437,7 @@ export class LLMStream extends llm.LLMStream {
         if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {
           callChunk = this.createRunningToolCallChunk(id, delta);
           this.toolCallId = this.fncName = this.fncRawArguments = undefined;
+          this.toolExtra = undefined;
         }
 
         // Start or continue building the current tool call
@@ -436,6 +446,10 @@ export class LLMStream extends llm.LLMStream {
           this.toolCallId = tool.id;
           this.fncName = tool.function.name;
           this.fncRawArguments = tool.function.arguments || '';
+          // Extract extra from tool call (e.g., Google thought signatures)
+          this.toolExtra =
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            ((tool as any).extra_content as Record<string, unknown> | undefined) ?? undefined;
         } else if (tool.function.arguments) {
           this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;
         }
@@ -454,11 +468,17 @@ export class LLMStream extends llm.LLMStream {
     ) {
       const callChunk = this.createRunningToolCallChunk(id, delta);
       this.toolCallId = this.fncName = this.fncRawArguments = undefined;
+      this.toolExtra = undefined;
       return callChunk;
     }
 
+    // Extract extra from delta (e.g., Google thought signatures on text parts)
+    const deltaExtra =
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      ((delta as any).extra_content as Record<string, unknown> | undefined) ?? undefined;
+
     // Regular content message
-    if (!delta.content) {
+    if (!delta.content && !deltaExtra) {
       return undefined;
     }
 
@@ -466,7 +486,8 @@ export class LLMStream extends llm.LLMStream {
       id,
       delta: {
         role: 'assistant',
-        content: delta.content,
+        content: delta.content || undefined,
+        extra: deltaExtra,
       },
     };
   }
@@ -475,19 +496,37 @@ export class LLMStream extends llm.LLMStream {
     id: string,
     delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,
   ): llm.ChatChunk {
+    const toolExtra = this.toolExtra ? { ...this.toolExtra } : {};
+    const thoughtSignature = this.extractThoughtSignature(toolExtra);
+    const deltaExtra =
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      ((delta as any).extra_content as Record<string, unknown> | undefined) ?? undefined;
+
     return {
       id,
       delta: {
         role: 'assistant',
         content: delta.content || undefined,
+        extra: deltaExtra,
         toolCalls: [
           llm.FunctionCall.create({
             callId: this.toolCallId || '',
             name: this.fncName || '',
             args: this.fncRawArguments || '',
+            extra: toolExtra,
+            thoughtSignature,
           }),
         ],
       },
     };
   }
+
+  private extractThoughtSignature(extra?: Record<string, unknown>): string | undefined {
+    const googleExtra = extra?.google;
+    if (googleExtra && typeof googleExtra === 'object') {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      return (googleExtra as any).thoughtSignature || (googleExtra as any).thought_signature;
+    }
+    return undefined;
+  }
 }

package/src/ipc/job_proc_lazy_main.ts

@@ -136,7 +136,7 @@ const startJob = (
       shutdownTasks.push(callback());
     }
     await Promise.all(shutdownTasks).catch((error) =>
-      logger.error('error while shutting down the job', error),
+      logger.error({ error }, 'error while shutting down the job'),
     );
 
     process.send!({ case: 'done' });

package/src/llm/chat_context.ts

@@ -189,6 +189,12 @@ export class FunctionCall {
 
   createdAt: number;
 
+  extra: Record<string, unknown>;
+  /**
+   * Optional grouping identifier for parallel tool calls.
+   */
+  groupId?: string;
+
   /**
    * Opaque signature for Gemini thinking mode.
    * When using Gemini 3+ models with thinking enabled, this signature must be
@@ -202,6 +208,8 @@ export class FunctionCall {
     args: string;
     id?: string;
     createdAt?: number;
+    extra?: Record<string, unknown>;
+    groupId?: string;
     thoughtSignature?: string;
   }) {
     const {
@@ -210,6 +218,8 @@ export class FunctionCall {
       args,
       id = shortuuid('item_'),
       createdAt = Date.now(),
+      extra = {},
+      groupId,
       thoughtSignature,
     } = params;
     this.id = id;
@@ -217,7 +227,15 @@ export class FunctionCall {
     this.args = args;
     this.name = name;
     this.createdAt = createdAt;
-    this.thoughtSignature = thoughtSignature;
+    this.extra = { ...extra };
+    this.groupId = groupId;
+    this.thoughtSignature =
+      thoughtSignature ??
+      (typeof this.extra.google === 'object' && this.extra.google !== null
+        ? // eslint-disable-next-line @typescript-eslint/no-explicit-any
+          (this.extra.google as any).thoughtSignature ||
+          (this.extra.google as any).thought_signature
+        : undefined);
   }
 
   static create(params: {
@@ -226,6 +244,8 @@ export class FunctionCall {
     args: string;
     id?: string;
     createdAt?: number;
+    extra?: Record<string, unknown>;
+    groupId?: string;
     thoughtSignature?: string;
   }) {
     return new FunctionCall(params);
@@ -241,6 +261,14 @@ export class FunctionCall {
       args: this.args,
     };
 
+    if (Object.keys(this.extra).length > 0) {
+      result.extra = this.extra as JSONValue;
+    }
+
+    if (this.groupId) {
+      result.groupId = this.groupId;
+    }
+
     if (this.thoughtSignature) {
       result.thoughtSignature = this.thoughtSignature;
     }
@@ -627,7 +655,9 @@ export class ChatContext {
           a.name !== b.name ||
           a.callId !== b.callId ||
           a.args !== b.args ||
-          a.thoughtSignature !== b.thoughtSignature
+          a.thoughtSignature !== b.thoughtSignature ||
+          a.groupId !== b.groupId ||
+          JSON.stringify(a.extra) !== JSON.stringify(b.extra)
         ) {
           return false;
         }

package/src/llm/llm.ts

@@ -17,6 +17,7 @@ export interface ChoiceDelta {
   role: ChatRole;
   content?: string;
   toolCalls?: FunctionCall[];
+  extra?: Record<string, unknown>;
 }
 
 export interface CompletionUsage {

package/src/llm/provider_format/openai.test.ts

@@ -258,6 +258,46 @@ describe('toChatCtx', () => {
     ]);
   });
 
+  it('should include provider-specific extra content on tool calls', async () => {
+    const ctx = ChatContext.empty();
+    const msg = ctx.addMessage({ role: 'assistant', content: 'Running tool' });
+
+    const toolCall = FunctionCall.create({
+      id: `${msg.id}/tool_1`,
+      callId: 'call_789',
+      name: 'google_call',
+      args: '{}',
+      extra: { google: { thoughtSignature: 'sig-123' } },
+    });
+    const toolOutput = FunctionCallOutput.create({
+      callId: 'call_789',
+      output: '{"result": "ok"}',
+      isError: false,
+    });
+
+    ctx.insert([toolCall, toolOutput]);
+
+    const result = await toChatCtx(ctx);
+
+    expect(result[0]).toEqual({
+      role: 'assistant',
+      content: 'Running tool',
+      tool_calls: [
+        {
+          type: 'function',
+          id: 'call_789',
+          function: { name: 'google_call', arguments: '{}' },
+          extra_content: { google: { thoughtSignature: 'sig-123' } },
+        },
+      ],
+    });
+    expect(result[1]).toEqual({
+      role: 'tool',
+      tool_call_id: 'call_789',
+      content: '{"result": "ok"}',
+    });
+  });
+
   it('should handle multiple tool calls in one message', async () => {
     const ctx = ChatContext.empty();
 

package/src/llm/provider_format/openai.ts

@@ -17,11 +17,20 @@ export async function toChatCtx(chatCtx: ChatContext, injectDummyUserMessage: bo
       ? await toChatItem(group.message)
       : { role: 'assistant' };
 
-    const toolCalls = group.toolCalls.map((toolCall) => ({
-      type: 'function',
-      id: toolCall.callId,
-      function: { name: toolCall.name, arguments: toolCall.args },
-    }));
+    const toolCalls = group.toolCalls.map((toolCall) => {
+      const tc: Record<string, any> = {
+        type: 'function',
+        id: toolCall.callId,
+        function: { name: toolCall.name, arguments: toolCall.args },
+      };
+
+      // Include provider-specific extra content (e.g., Google thought signatures)
+      const googleExtra = getGoogleExtra(toolCall);
+      if (googleExtra) {
+        tc.extra_content = { google: googleExtra };
+      }
+      return tc;
+    });
 
     if (toolCalls.length > 0) {
       message['tool_calls'] = toolCalls;
@@ -53,24 +62,33 @@ async function toChatItem(item: ChatItem) {
       }
     }
 
-    const content =
-      listContent.length == 0
-        ? textContent
-        : textContent.length == 0
-          ? listContent
-          : [...listContent, { type: 'text', text: textContent }];
+    const result: Record<string, any> = { role: item.role };
+    if (listContent.length === 0) {
+      result.content = textContent;
+    } else {
+      if (textContent.length > 0) {
+        listContent.push({ type: 'text', text: textContent });
+      }
+      result.content = listContent;
+    }
 
-    return { role: item.role, content };
+    return result;
   } else if (item.type === 'function_call') {
+    const tc: Record<string, any> = {
+      id: item.callId,
+      type: 'function',
+      function: { name: item.name, arguments: item.args },
+    };
+
+    // Include provider-specific extra content (e.g., Google thought signatures)
+    const googleExtra = getGoogleExtra(item);
+    if (googleExtra) {
+      tc.extra_content = { google: googleExtra };
+    }
+
     return {
       role: 'assistant',
-      tool_calls: [
-        {
-          id: item.callId,
-          type: 'function',
-          function: { name: item.name, arguments: item.args },
-        },
-      ],
+      tool_calls: [tc],
     };
   } else if (item.type === 'function_call_output') {
     return {
@@ -84,6 +102,15 @@ async function toChatItem(item: ChatItem) {
   throw new Error(`Unsupported item type: ${item['type']}`);
 }
 
+function getGoogleExtra(
+  item: Partial<{ extra?: Record<string, unknown>; thoughtSignature?: string }>,
+): Record<string, unknown> | undefined {
+  const googleExtra =
+    (item.extra?.google as Record<string, unknown> | undefined) ||
+    (item.thoughtSignature ? { thoughtSignature: item.thoughtSignature } : undefined);
+  return googleExtra;
+}
+
 async function toImageContent(content: ImageContent) {
   const cacheKey = 'serialized_image'; // TODO: use hash of encoding options if available
   let serialized: SerializedImage;

package/src/llm/provider_format/utils.ts

@@ -133,7 +133,11 @@ export function groupToolCalls(chatCtx: ChatContext) {
 
     if (isAssistantMessage || isFunctionCall) {
       // only assistant messages and function calls can be grouped
-      const groupId = item.id.split('/')[0]!;
+      // For function calls, use group_id if available (for parallel function calls),
+      // otherwise fall back to id-based grouping for backwards compatibility
+      const groupId =
+        item.type === 'function_call' && item.groupId ? item.groupId : item.id.split('/')[0]!;
+
       if (itemGroups[groupId] === undefined) {
         itemGroups[groupId] = ChatItemGroup.create();
 

package/src/voice/agent_activity.ts

@@ -1350,11 +1350,14 @@ export class AgentActivity implements RecognitionHooks {
     );
     tasks.push(llmTask);
 
-    const [ttsTextInput, llmOutput] = llmGenData.textStream.tee();
-
     let ttsTask: Task<void> | null = null;
     let ttsStream: ReadableStream<AudioFrame> | null = null;
+    let llmOutput: ReadableStream<string>;
+
     if (audioOutput) {
+      // Only tee the stream when we need TTS
+      const [ttsTextInput, textOutput] = llmGenData.textStream.tee();
+      llmOutput = textOutput;
       [ttsTask, ttsStream] = performTTSInference(
         (...args) => this.agent.ttsNode(...args),
         ttsTextInput,
@@ -1362,6 +1365,9 @@ export class AgentActivity implements RecognitionHooks {
         replyAbortController,
       );
       tasks.push(ttsTask);
+    } else {
+      // No TTS needed, use the stream directly
+      llmOutput = llmGenData.textStream;
     }
 
     await speechHandle.waitIfNotInterrupted([speechHandle._waitForScheduled()]);
@@ -1421,12 +1427,16 @@ export class AgentActivity implements RecognitionHooks {
     //TODO(AJS-272): before executing tools, make sure we generated all the text
     // (this ensure everything is kept ordered)
 
-    const onToolExecutionStarted = (_: FunctionCall) => {
-      // TODO(brian): handle speech_handle item_added
+    const onToolExecutionStarted = (f: FunctionCall) => {
+      speechHandle._itemAdded([f]);
+      this.agent._chatCtx.items.push(f);
+      this.agentSession._toolItemsAdded([f]);
     };
 
-    const onToolExecutionCompleted = (_: ToolExecutionOutput) => {
-      // TODO(brian): handle speech_handle item_added
+    const onToolExecutionCompleted = (out: ToolExecutionOutput) => {
+      if (out.toolCallOutput) {
+        speechHandle._itemAdded([out.toolCallOutput]);
+      }
     };
 
     const [executeToolsTask, toolOutput] = performToolExecutions({
@@ -1501,6 +1511,7 @@ export class AgentActivity implements RecognitionHooks {
         });
         chatCtx.insert(message);
         this.agent._chatCtx.insert(message);
+        speechHandle._itemAdded([message]);
         this.agentSession._conversationItemAdded(message);
       }
 
@@ -1528,6 +1539,7 @@ export class AgentActivity implements RecognitionHooks {
       });
       chatCtx.insert(message);
       this.agent._chatCtx.insert(message);
+      speechHandle._itemAdded([message]);
       this.agentSession._conversationItemAdded(message);
       this.logger.info(
         { speech_id: speechHandle.id, message: textOut.text },
@@ -1612,28 +1624,18 @@ export class AgentActivity implements RecognitionHooks {
     if (shouldGenerateToolReply) {
       chatCtx.insert(toolMessages);
 
-      const handle = SpeechHandle.create({
-        allowInterruptions: speechHandle.allowInterruptions,
-        stepIndex: speechHandle._stepIndex + 1,
-        parent: speechHandle,
-      });
-      this.agentSession.emit(
-        AgentSessionEventTypes.SpeechCreated,
-        createSpeechCreatedEvent({
-          userInitiated: false,
-          source: 'tool_response',
-          speechHandle: handle,
-        }),
-      );
+      // Increment step count on SAME handle (parity with Python agent_activity.py L2081)
+      speechHandle._numSteps += 1;
 
       // Avoid setting tool_choice to "required" or a specific function when
       // passing tool response back to the LLM
       const respondToolChoice = draining || modelSettings.toolChoice === 'none' ? 'none' : 'auto';
 
+      // Reuse same speechHandle for tool response (parity with Python agent_activity.py L2122-2140)
       const toolResponseTask = this.createSpeechTask({
         task: Task.from(() =>
           this.pipelineReplyTask(
-            handle,
+            speechHandle,
             chatCtx,
             toolCtx,
             { toolChoice: respondToolChoice },
@@ -1643,13 +1645,13 @@ export class AgentActivity implements RecognitionHooks {
             toolMessages,
           ),
         ),
-        ownedSpeechHandle: handle,
+        ownedSpeechHandle: speechHandle,
         name: 'AgentActivity.pipelineReply',
       });
 
       toolResponseTask.finally(() => this.onPipelineReplyDone());
 
-      this.scheduleSpeech(handle, SpeechHandle.SPEECH_PRIORITY_NORMAL, true);
+      this.scheduleSpeech(speechHandle, SpeechHandle.SPEECH_PRIORITY_NORMAL, true);
     } else if (functionToolsExecutedEvent.functionCallOutputs.length > 0) {
       for (const msg of toolMessages) {
         msg.createdAt = replyStartedAt;