From 7171f7921970a87bdb5d8704866c57a6acd56418 Mon Sep 17 00:00:00 2001
From: Harshdeep Singh <harshdeepsingh1121@outlook.com>
Date: Fri, 12 Jun 2026 15:50:42 +0530
Subject: [PATCH] test(e2e): add Gemini TTS coverage to provider matrix

Wire Gemini TTS into the existing E2E test infrastructure.

- Add Gemini TTS adapter factory in media-providers
- Add Gemini generateContent audio mock
- Add Gemini to TTS feature support
- Exercise PCM-to-WAV normalization using raw PCM audio responses

This extends existing TTS E2E coverage without changing
runtime behavior or public APIs.
---
 testing/e2e/global-setup.ts            | 67 ++++++++++++++++++++++++++
 testing/e2e/src/lib/feature-support.ts |  2 +-
 testing/e2e/src/lib/media-providers.ts | 10 +++-
 3 files changed, 77 insertions(+), 2 deletions(-)

diff --git a/testing/e2e/global-setup.ts b/testing/e2e/global-setup.ts
index f869df01a..5f1bd7ca9 100644
--- a/testing/e2e/global-setup.ts
+++ b/testing/e2e/global-setup.ts
@@ -43,6 +43,19 @@ export default async function globalSetup() {
   mock.mount('/v1/text-to-speech', elevenLabsTTSMount())
   mock.mount('/v1/speech-to-text', elevenLabsSTTMount())
 
+  // Gemini TTS hits the standard Gemini generateContent endpoint
+  // (POST /v1beta/models/{model}:generateContent) with
+  // responseModalities: ['AUDIO']. aimock's native Gemini audio helper derives
+  // the mime type from the fixture's `format`/`contentType`, so it can't emit
+  // the raw `audio/L16;codec=pcm;rate=24000` PCM that real Gemini TTS returns.
+  // Mount the TTS model's generateContent path directly so we can hand back
+  // PCM and exercise the adapter's PCM→WAV normalization. The path is specific
+  // to the TTS model, so it doesn't intercept Gemini chat/summarize requests.
+  mock.mount(
+    '/v1beta/models/gemini-3.1-flash-tts-preview:generateContent',
+    geminiTTSMount(),
+  )
+
   // Anthropic server_tool_use bug reproduction (issue #604). aimock can't
   // natively synthesize `server_tool_use` / `web_fetch_tool_result` content
   // blocks, so this mount hand-crafts the raw SSE Claude would emit when a
@@ -107,6 +120,14 @@ const FAKE_MP3_BYTES = Buffer.from([
   0xff, 0xfb, 0x90, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 ])
 
+/**
+ * Raw 16-bit little-endian PCM bytes. Gemini TTS returns audio as
+ * `audio/L16;codec=pcm;rate=24000` inlineData, which the adapter wraps in a
+ * RIFF/WAV header before handing it to the browser. The samples are arbitrary
+ * silence — the spec only asserts the `<audio>` element becomes visible.
+ */
+const FAKE_PCM_BYTES = Buffer.alloc(32)
+
 function grokTTSMount(): Mountable {
   return {
     async handleRequest(
@@ -127,6 +148,52 @@ function grokTTSMount(): Mountable {
   }
 }
 
+function geminiTTSMount(): Mountable {
+  return {
+    async handleRequest(
+      req: http.IncomingMessage,
+      res: http.ServerResponse,
+      // aimock strips the mount prefix — pathname will be "/" for an exact match.
+      pathname: string,
+    ): Promise<boolean> {
+      if (pathname !== '/' || req.method !== 'POST') return false
+      await drainBody(req)
+      res.statusCode = 200
+      res.setHeader('Content-Type', 'application/json')
+      // Mirror the Gemini generateContent audio response shape: audio lands as
+      // a single `candidates[0].content.parts[0].inlineData` entry. The PCM
+      // mime type forces the adapter down its PCM→WAV wrapping path.
+      res.end(
+        JSON.stringify({
+          candidates: [
+            {
+              content: {
+                role: 'model',
+                parts: [
+                  {
+                    inlineData: {
+                      mimeType: 'audio/L16;codec=pcm;rate=24000',
+                      data: FAKE_PCM_BYTES.toString('base64'),
+                    },
+                  },
+                ],
+              },
+              finishReason: 'STOP',
+              index: 0,
+            },
+          ],
+          usageMetadata: {
+            promptTokenCount: 5,
+            candidatesTokenCount: 15,
+            totalTokenCount: 20,
+          },
+        }),
+      )
+      return true
+    },
+  }
+}
+
 function grokSTTMount(): Mountable {
   return {
     async handleRequest(
diff --git a/testing/e2e/src/lib/feature-support.ts b/testing/e2e/src/lib/feature-support.ts
index 6d6b950bd..a8ee14fd5 100644
--- a/testing/e2e/src/lib/feature-support.ts
+++ b/testing/e2e/src/lib/feature-support.ts
@@ -180,7 +180,7 @@ export const matrix: Record<Feature, Set<Provider>> = {
   'image-gen': new Set(['openai', 'grok']),
   'audio-gen': new Set(['gemini', 'elevenlabs']),
   'sound-effects': new Set(['elevenlabs']),
-  tts: new Set(['openai', 'grok', 'elevenlabs']),
+  tts: new Set(['openai', 'gemini', 'grok', 'elevenlabs']),
   transcription: new Set(['openai', 'grok', 'elevenlabs']),
   'video-gen': new Set(['openai']),
   // Only Gemini currently surfaces a first-class stateful conversation API via
diff --git a/testing/e2e/src/lib/media-providers.ts b/testing/e2e/src/lib/media-providers.ts
index d399eb33f..0bd5bdd24 100644
--- a/testing/e2e/src/lib/media-providers.ts
+++ b/testing/e2e/src/lib/media-providers.ts
@@ -4,7 +4,11 @@ import {
   createOpenaiTranscription,
   createOpenaiVideo,
 } from '@tanstack/ai-openai'
-import { createGeminiAudio, createGeminiImage } from '@tanstack/ai-gemini'
+import {
+  createGeminiAudio,
+  createGeminiImage,
+  createGeminiSpeech,
+} from '@tanstack/ai-gemini'
 import {
   createGrokImage,
   createGrokSpeech,
@@ -72,6 +76,10 @@ export function createTTSAdapter(
         baseURL: openaiUrl(aimockPort),
         defaultHeaders: headers,
       }),
+    gemini: () =>
+      createGeminiSpeech('gemini-3.1-flash-tts-preview', DUMMY_KEY, {
+        httpOptions: { baseUrl: llmockBase(aimockPort), headers },
+      }),
     grok: () =>
       createGrokSpeech('grok-tts', DUMMY_KEY, {
         baseURL: openaiUrl(aimockPort),