fix(google): handle content blocking and generation failures

.changeset/google-generation-failures.md

@@ -0,0 +1,5 @@
+---
+'@livekit/agents-plugin-google': patch
+---
+
+fix(google): surface Gemini content blocking and empty generation failures

plugins/google/src/llm.ts

@@ -2,7 +2,12 @@
 //
 // SPDX-License-Identifier: Apache-2.0
 import type * as types from '@google/genai';
-import { FunctionCallingConfigMode, type GenerateContentConfig, GoogleGenAI } from '@google/genai';
+import {
+  FinishReason,
+  FunctionCallingConfigMode,
+  type GenerateContentConfig,
+  GoogleGenAI,
+} from '@google/genai';
 import type { APIConnectOptions } from '@livekit/agents';
 import {
   APIConnectionError,
@@ -296,13 +301,13 @@ export class LLM extends llm.LLM {
   }
 }
 
-const BLOCKED_REASONS = [
-  'SAFETY',
-  'SPII',
-  'PROHIBITED_CONTENT',
-  'BLOCKLIST',
-  'LANGUAGE',
-  'RECITATION',
+const BLOCKED_REASONS: types.FinishReason[] = [
+  FinishReason.SAFETY,
+  FinishReason.SPII,
+  FinishReason.PROHIBITED_CONTENT,
+  FinishReason.BLOCKLIST,
+  FinishReason.LANGUAGE,
+  FinishReason.RECITATION,
 ];
 
 export class LLMStream extends llm.LLMStream {
@@ -380,6 +385,9 @@ export class LLMStream extends llm.LLMStream {
         },
       });
 
+      let responseGenerated = false;
+      let finishReason: types.FinishReason | undefined;
+
       for await (const chunk of response) {
         if (chunk.promptFeedback) {
           throw new APIStatusError({
@@ -391,32 +399,20 @@ export class LLMStream extends llm.LLMStream {
           });
         }
 
-        // Check for blocked reasons first — safety-blocked responses often lack content.parts,
-        // so this must run before the no-content guard to avoid wasting retries.
-        if (
-          chunk.candidates?.[0]?.finishReason &&
-          BLOCKED_REASONS.includes(chunk.candidates[0].finishReason)
-        ) {
-          throw new APIStatusError({
-            message: `Google LLM: generation blocked - ${chunk.candidates[0].finishReason}`,
-            options: {
-              retryable: false,
-              requestId,
+        if (chunk.usageMetadata) {
+          const usage = chunk.usageMetadata;
+          this.queue.put({
+            id: requestId,
+            usage: {
+              completionTokens: usage.candidatesTokenCount || 0,
+              promptTokens: usage.promptTokenCount || 0,
+              promptCachedTokens: usage.cachedContentTokenCount || 0,
+              totalTokens: usage.totalTokenCount || 0,
             },
           });
         }
 
-        if (!chunk.candidates || !chunk.candidates[0]?.content?.parts) {
-          this.logger.warn(`No content in the response: ${JSON.stringify(chunk)}`);
-          if (retryable) {
-            throw new APIStatusError({
-              message: 'Google LLM: no content in the response',
-              options: {
-                retryable: true,
-                requestId,
-              },
-            });
-          }
+        if (!chunk.candidates) {
           continue;
         }
 
@@ -427,40 +423,47 @@ export class LLMStream extends llm.LLMStream {
         }
 
         const candidate = chunk.candidates[0];
-        const finishReason = candidate.finishReason;
+        if (!candidate) {
+          continue;
+        }
 
-        let chunksYielded = false;
-        for (const part of candidate.content!.parts!) {
-          const chatChunk = this.#parsePart(requestId, part);
-          if (chatChunk) {
-            chunksYielded = true;
-            retryable = false;
-            this.queue.put(chatChunk);
-          }
+        if (candidate.finishReason) {
+          finishReason = candidate.finishReason;
         }
 
-        if (finishReason === 'STOP' && !chunksYielded && retryable) {
+        if (candidate.finishReason && BLOCKED_REASONS.includes(candidate.finishReason)) {
           throw new APIStatusError({
-            message: 'Google LLM: no response generated',
+            message: `Google LLM: generation blocked by Gemini: ${candidate.finishReason}`,
             options: {
-              retryable,
+              retryable: false,
               requestId,
             },
           });
         }
 
-        if (chunk.usageMetadata) {
-          const usage = chunk.usageMetadata;
-          this.queue.put({
-            id: requestId,
-            usage: {
-              completionTokens: usage.candidatesTokenCount || 0,
-              promptTokens: usage.promptTokenCount || 0,
-              promptCachedTokens: usage.cachedContentTokenCount || 0,
-              totalTokens: usage.totalTokenCount || 0,
-            },
-          });
+        if (!candidate.content?.parts) {
+          continue;
         }
+
+        for (const part of candidate.content.parts) {
+          const chatChunk = this.#parsePart(requestId, part);
+          responseGenerated = true;
+          if (chatChunk) {
+            retryable = false;
+            this.queue.put(chatChunk);
+          }
+        }
+      }
+
+      if (!responseGenerated) {
+        throw new APIStatusError({
+          message: 'Google LLM: no response generated',
+          options: {
+            body: { finishReason },
+            retryable,
+            requestId,
+          },
+        });
       }
     } catch (error: unknown) {
       if (error instanceof APIStatusError || error instanceof APIConnectionError) {