what are the steps to properly close a voice agent conversation with gemini realtime plugin

Thanks for reaching out, adit-kulkarni. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

should i be creations new session/localmedia objects for every vonversation my user has or just have one and keep reusing it?

Thanks — quick answer: create a new session for each conversation. Why: • Each

AgentSession

manages a single, isolated conversation lifecycle (init, media I/O, activity tracking, cleanup). Reusing a session can cause state conflicts. • Local media is managed per session; you don’t need to persist

LocalMedia

yourself. Let the session create and dispose media resources so tracks and state stay clean. If you’d like, tell me which language/SDK you’re using (Python or JS) and I’ll share a short example for creating and tearing down sessions. Sources: Agent session | LiveKit docs | Client SDKs | livekit/livekit_composite | DeepWiki

let me rephrase: We're experiencing crashes when starting a second LiveKit conversation after ending the first one. Looking at the agent-starter-swift example app, I noticed a key architectural difference: Example app approach: // VoiceAgentApp.swift @main struct VoiceAgentApp: App { private let session = Session(...) // Created ONCE at app level var body: some Scene { WindowGroup { AppView() .environmentObject(session) .environmentObject(LocalMedia(session: session)) } } } The Session and LocalMedia are created once and reused - session.end() is called to disconnect, but the same objects are reused for the next session.start(). Our current approach: We create NEW Session and LocalMedia objects for each conversation (in a SwiftUI view or manager). When the conversation ends, these objects are destroyed. When starting a new conversation, we create fresh Session/LocalMedia objects. Question: Is the example app's pattern (creating Session/LocalMedia once at app-level and reusing them) the recommended approach for avoiding native resource conflicts? Should we avoid destroying and recreating Session/LocalMedia objects, and instead reuse a single persistent instance across multiple conversations?

Thanks — I’m responding to your question about whether to reuse Session/LocalMedia or create new ones per conversation. Short answer: Reuse a single app-level

Session

and

LocalMedia

instance across conversations (the example app pattern). This is the recommended approach. Why: • `Session`/`LocalMedia` own native WebRTC and AVFoundation resources (audio units, capture devices, sockets). Recreating them frequently causes race conditions, resource contention, AVAudioSession interruptions, and crashes. • Calling

session.end()

cleanly disconnects and releases session resources while keeping the object usable for a new

session.start()

with a fresh room/token. • Letting the session manage media lifecycle ensures clean teardown and prevents stale native state. Recommended actions: 1. Create

Session

and

LocalMedia

once at app startup (e.g., as EnvironmentObjects). 2. On conversation end, call

session.end()

(drain/cleanup) — do not destroy the objects. 3. For a new conversation, call

session.start()

with the new room URL/token. If you’d like, share crash logs or the sequence you use to end/start sessions (calls and timing). I can point out likely race points and give exact Swift code snippets to switch to the persistent pattern. Sources: client-sdk-swift | livekit/livekit_composite | DeepWiki