fix(google): broaden Gemini 3 Flash detection

.changeset/gemini-3-flash-thinking.md

@@ -0,0 +1,5 @@
+---
+"@livekit/agents-plugin-google": patch
+---
+
+fix(plugin-google): default Gemini 3 Flash variants to minimal thinking.

.changeset/neat-crabs-watch.md

@@ -0,0 +1,6 @@
+---
+"@livekit/agents": patch
+"@livekit/agents-plugin-google": patch
+---
+
+fix(llm): log silent error from performLLMInference 

agents/src/voice/generation.ts

@@ -476,6 +476,7 @@ export function performLLMInference(
   model?: string,
   provider?: string,
 ): [Task<void>, _LLMGenerationData] {
+  const logger = log();
   const textStream = new IdentityTransform<string | FlushSentinel>();
   const toolCallStream = new IdentityTransform<FunctionCall>();
 
@@ -584,6 +585,8 @@ export function performLLMInference(
         // Abort signal was triggered, handle gracefully
         return;
       }
+      // surface inference silent errors even when this task's rejection is never awaited
+      logger.error({ error }, 'error in llm node');
       throw error;
     } finally {
       llmStreamReader?.releaseLock();

plugins/google/src/llm.ts

@@ -2,13 +2,19 @@
 //
 // SPDX-License-Identifier: Apache-2.0
 import type * as types from '@google/genai';
-import { FunctionCallingConfigMode, type GenerateContentConfig, GoogleGenAI } from '@google/genai';
+import {
+  FunctionCallingConfigMode,
+  type GenerateContentConfig,
+  GoogleGenAI,
+  ThinkingLevel,
+} from '@google/genai';
 import type { APIConnectOptions } from '@livekit/agents';
 import {
   APIConnectionError,
   APIStatusError,
   DEFAULT_API_CONNECT_OPTIONS,
   llm,
+  log,
   shortuuid,
 } from '@livekit/agents';
 import type { ChatModels } from './models.js';
@@ -19,6 +25,16 @@ interface GoogleFormatData {
   systemMessages: string[] | null;
 }
 
+function isGemini3Model(model: string): boolean {
+  const modelLower = model.toLowerCase();
+  return modelLower.includes('gemini-3');
+}
+
+function isGemini3FlashModel(model: string): boolean {
+  const modelLower = model.toLowerCase();
+  return modelLower.includes('gemini-3') && modelLower.includes('flash');
+}
+
 export interface LLMOptions {
   model: string | ChatModels;
   apiKey?: string;
@@ -275,7 +291,38 @@ export class LLM extends llm.LLM {
     }
 
     if (this.#opts.thinkingConfig !== undefined) {
-      extras.thinkingConfig = this.#opts.thinkingConfig;
+      const { includeThoughts, thinkingBudget, thinkingLevel } = this.#opts.thinkingConfig;
+
+      if (isGemini3Model(this.#opts.model)) {
+        // Gemini 3: only supports thinkingLevel
+        if (thinkingBudget !== undefined && thinkingLevel === undefined) {
+          log().warn(
+            `Model ${this.#opts.model} is Gemini 3 which does not support thinkingBudget. ` +
+              `Please use thinkingLevel ('low' or 'high') instead. Ignoring thinkingBudget.`,
+          );
+        }
+        extras.thinkingConfig = {
+          includeThoughts,
+          thinkingLevel:
+            thinkingLevel ??
+            (isGemini3FlashModel(this.#opts.model) ? ThinkingLevel.MINIMAL : ThinkingLevel.LOW),
+        };
+      } else {
+        // Gemini 2.5 and earlier: only supports thinkingBudget
+        if (thinkingLevel !== undefined && thinkingBudget === undefined) {
+          log().warn(
+            `Model ${this.#opts.model} does not support thinkingLevel. ` +
+              `Please use thinkingBudget (number) instead for Gemini 2.5 and earlier models. ` +
+              `Ignoring thinkingLevel.`,
+          );
+          extras.thinkingConfig = { includeThoughts };
+        } else if (thinkingBudget !== undefined) {
+          // Preserve includeThoughts so callers relying on visible reasoning keep receiving it.
+          extras.thinkingConfig = { thinkingBudget, includeThoughts };
+        } else {
+          extras.thinkingConfig = this.#opts.thinkingConfig;
+        }
+      }
     }
 
     if (this.#opts.automaticFunctionCallingConfig !== undefined) {