feat(livekit): upgrade Gemini starter

examples/livekit-gemini-hacker-starter/agent/.env.example

@@ -2,3 +2,6 @@ LIVEKIT_API_KEY=<your API Key>
 LIVEKIT_API_SECRET=<your API Secret>
 LIVEKIT_URL=<your LiveKit server URL>
 GOOGLE_API_KEY=<your Google/Gemini API key>
+GEMINI_REALTIME_MODEL=gemini-3.1-flash-live-preview
+GEMINI_IMAGE_MODEL=gemini-3.1-flash-image
+GEMINI_LYRIA_MODEL=models/lyria-realtime-exp

examples/livekit-gemini-hacker-starter/agent/agent.py

@@ -16,9 +16,9 @@ logger = logging.getLogger(__name__)
 # ─────────────────────────────────────────────
 # HACK HERE: swap model IDs to experiment
 # ─────────────────────────────────────────────
-REALTIME_MODEL = "gemini-2.5-flash-native-audio-preview-12-2025"
-IMAGE_MODEL = "gemini-2.5-flash-image"  # Nano Banana
-LYRIA_MODEL = "models/lyria-realtime-exp"
+REALTIME_MODEL = os.getenv("GEMINI_REALTIME_MODEL", "gemini-3.1-flash-live-preview")
+IMAGE_MODEL = os.getenv("GEMINI_IMAGE_MODEL", "gemini-3.1-flash-image")  # Nano Banana 2
+LYRIA_MODEL = os.getenv("GEMINI_LYRIA_MODEL", "models/lyria-realtime-exp")
 
 # ─────────────────────────────────────────────
 # HACK HERE: change the agent's persona
@@ -27,7 +27,7 @@ PERSONA_INSTRUCTIONS = """You are a creative multimodal AI assistant at a Google
 You can see through the user's camera, hear them speak, generate images, and play real-time music.
 
 Your capabilities:
-- generate_image: Create images with Nano Banana (Gemini 2.5 Flash Image). Use this when asked to generate, create, render, or visualize anything.
+- generate_image: Create images with Nano Banana 2 (Gemini 3.1 Flash Image). Use this when asked to generate, create, render, or visualize anything.
 - start_music: Play real-time generative music with Lyria RealTime. Use this for soundtracks, ambience, or any audio atmosphere.
 - stop_music: Stop the current music.
 
@@ -281,9 +281,10 @@ async def entrypoint(ctx: agents.JobContext):
 
     await ctx.connect()
 
+    if REALTIME_MODEL != "gemini-3.1-flash-live-preview":
         try:
             await session.generate_reply(
-            instructions="Greet the user. Let them know you can generate images with Nano Banana and play real-time music with Lyria. Mention they can enable their camera for visual context."
+                instructions="Greet the user. Let them know you can generate images with Nano Banana 2 and play real-time music with Lyria. Mention they can enable their camera for visual context."
             )
         except Exception as exc:
             logger.warning("Initial greeting failed: %s", exc)

examples/livekit-gemini-hacker-starter/agent/pyproject.toml

@@ -4,8 +4,8 @@ version = "0.1.0"
 description = "Gemini hackathon starter — voice, vision, image generation, and real-time music with LiveKit"
 requires-python = ">=3.10,<3.14"
 dependencies = [
-    "livekit-agents[google,images]~=1.4",
-    "google-genai>=1.16.0",
+    "livekit-agents[google,images]>=1.6.4,<1.7",
+    "google-genai>=2.10.0,<3",
     "python-dotenv>=1.0.0",
 ]
 

examples/livekit-gemini-hacker-starter/frontend/app-config.ts

@@ -25,7 +25,7 @@ export const APP_CONFIG_DEFAULTS: AppConfig = {
   companyName: 'Gemini Hackathon',
   pageTitle: 'Gemini Hacker Starter',
   pageDescription:
-    'Voice, vision, image generation, and real-time music with Gemini 2.5 native audio, Nano Banana, and Lyria',
+    'Voice, vision, image generation, and real-time music with Gemini 3.1 Live, Nano Banana 2, and Lyria',
 
   supportsChatInput: true,
   supportsVideoInput: true,

examples/livekit-gemini-hacker-starter/frontend/components/app/welcome-view.tsx

@@ -34,7 +34,7 @@ export const WelcomeView = ({
         <WelcomeImage />
 
         <p className="text-foreground max-w-prose pt-1 leading-6 font-medium">
-          Voice, vision, images, and music — powered by Gemini 3.1, NanoBanana 2, and Lyria
+          Voice, vision, images, and music — powered by Gemini 3.1 Live, NanoBanana 2, and Lyria
         </p>
 
         <Button

frontend/src/components/PodView.tsx

@@ -116,6 +116,7 @@ export function PodView({
   const [testingVoice, setTestingVoice] = useState(false);
   const [note, setNote] = useState<string | null>(null);
   const [audioBlocked, setAudioBlocked] = useState(false);
+  const [remoteAudioTracks, setRemoteAudioTracks] = useState(0);
   const [micOn, setMicOn] = useState(false);
   const [leftStreamOpen, setLeftStreamOpen] = useState(() =>
     readStoredBool('podman.myStreamOpen', true),
@@ -146,6 +147,7 @@ export function PodView({
       element.autoplay = true;
       audioElementsRef.current.set(key, element);
       audioRef.current.appendChild(element);
+      setRemoteAudioTracks(audioElementsRef.current.size);
     };
     const removeAudio = (track: RemoteTrack, pub?: RemoteTrackPublication) => {
       const key = pub?.trackSid || track.sid || track.mediaStreamTrack.id;
@@ -153,6 +155,7 @@ export function PodView({
       if (attached) {
         attached.remove();
         audioElementsRef.current.delete(key);
+        setRemoteAudioTracks(audioElementsRef.current.size);
       }
       track.detach().forEach((el) => el.remove());
     };
@@ -165,14 +168,23 @@ export function PodView({
       });
     };
     const onAudio = (track: RemoteTrack, pub: RemoteTrackPublication) => attachAudio(track, pub);
-    const onAudioGone = (track: RemoteTrack, pub: RemoteTrackPublication) => removeAudio(track, pub);
+    const onAudioGone = (track: RemoteTrack, pub: RemoteTrackPublication) =>
+      removeAudio(track, pub);
     const onDisconnected = () => onLeaveRef.current();
     // Browsers block autoplay of incoming audio until a user gesture unlocks it.
     // Surface a button whenever the room can't play sound so PodMan's voice cues
     // are actually heard.
     const onPlaybackChanged = () => setAudioBlocked(!room.canPlaybackAudio);
+    const unlockAudioFromGesture = () => {
+      void room.startAudio().finally(() => setAudioBlocked(!room.canPlaybackAudio));
+    };
     onPlaybackChanged();
     setMicOn(room.localParticipant.isMicrophoneEnabled);
+    window.addEventListener('pointerdown', unlockAudioFromGesture, {
+      once: true,
+      capture: true,
+    });
+    window.addEventListener('keydown', unlockAudioFromGesture, { once: true, capture: true });
 
     room
       .on(RoomEvent.ParticipantConnected, refresh)
@@ -193,8 +205,11 @@ export function PodView({
         .off(RoomEvent.TrackUnsubscribed, onAudioGone)
         .off(RoomEvent.AudioPlaybackStatusChanged, onPlaybackChanged)
         .off(RoomEvent.Disconnected, onDisconnected);
+      window.removeEventListener('pointerdown', unlockAudioFromGesture, { capture: true });
+      window.removeEventListener('keydown', unlockAudioFromGesture, { capture: true });
       audioElementsRef.current.forEach((el) => el.remove());
       audioElementsRef.current.clear();
+      setRemoteAudioTracks(0);
     };
   }, [room, me]);
 
@@ -268,6 +283,8 @@ export function PodView({
     if (!room) return;
     setNote(null);
     try {
+      await room.startAudio().catch(() => {});
+      setAudioBlocked(!room.canPlaybackAudio);
       if (sharing) {
         if (screenTrackRef.current)
           await room.localParticipant.unpublishTrack(screenTrackRef.current);
@@ -585,7 +602,15 @@ export function PodView({
                       <StatusLine
                         label="Audio"
                         value={
-                          beat.on ? (beat.mine ? 'publishing' : `${beat.by} playing`) : 'ready'
+                          audioBlocked
+                            ? 'blocked'
+                            : remoteAudioTracks > 0
+                              ? `${remoteAudioTracks} remote track${remoteAudioTracks === 1 ? '' : 's'}`
+                              : beat.on
+                                ? beat.mine
+                                  ? 'publishing'
+                                  : `${beat.by} playing`
+                                : 'ready'
                         }
                       />
                       <StatusLine label="Agent" value={podmanPresent ? 'watching' : 'waiting'} />