fix: preserve thought_signature across parallel tool calls for Gemini 3+ (#1000)

Author: toubatbrian

.changeset/salty-regions-relate.md

@@ -0,0 +1,5 @@
+---
+'@livekit/agents': patch
+---
+
+preserve thought_signature across parallel tool calls for Gemini 3+ for inference gateway

agents/src/inference/llm.ts

@@ -433,7 +433,10 @@ 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;
+          // Note: We intentionally do NOT reset toolExtra here.
+          // For Gemini 3+, the thought_signature is only provided on the first tool call
+          // in a parallel batch, but must be applied to ALL tool calls in the batch.
+          // We preserve toolExtra so subsequent tool calls inherit the thought_signature.
         }
 
         // Start or continue building the current tool call
@@ -443,9 +446,14 @@ export class LLMStream extends llm.LLMStream {
           this.fncName = tool.function.name;
           this.fncRawArguments = tool.function.arguments || '';
           // Extract extra from tool call (e.g., Google thought signatures)
-          this.toolExtra =
+          // Only update toolExtra if this tool call has extra_content.
+          // Otherwise, inherit from previous tool call (for parallel Gemini tool calls).
+          const newToolExtra =
             // eslint-disable-next-line @typescript-eslint/no-explicit-any
             ((tool as any).extra_content as Record<string, unknown> | undefined) ?? undefined;
+          if (newToolExtra) {
+            this.toolExtra = newToolExtra;
+          }
         } else if (tool.function.arguments) {
           this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;
         }
@@ -464,6 +472,7 @@ export class LLMStream extends llm.LLMStream {
     ) {
       const callChunk = this.createRunningToolCallChunk(id, delta);
       this.toolCallId = this.fncName = this.fncRawArguments = undefined;
+      // Reset toolExtra at the end of the response (not between parallel tool calls)
       this.toolExtra = undefined;
       return callChunk;
     }

agents/src/voice/agent.ts

@@ -325,16 +325,16 @@ export class Agent<UserData = any> {
         );
       }
 
-      // TODO(brian): make parallelToolCalls configurable
       const { toolChoice } = modelSettings;
       const connOptions = activity.agentSession.connOptions.llmConnOptions;
 
+      // parallelToolCalls is not passed here - it will use the value from LLM's modelOptions
+      // This allows users to configure it via: new inference.LLM({ modelOptions: { parallel_tool_calls: false } })
       const stream = activity.llm.chat({
         chatCtx,
         toolCtx,
         toolChoice,
         connOptions,
-        parallelToolCalls: true,
       });
 
       let cleaned = false;