@enfinitos/sdk-audio

EnfinitOS reference SDK for the AUDIO substrate — smart-speaker

• podcast + audio-streaming. Built around the **2026 voice-

assistant LLM-extension pattern**: voice platforms increasingly treat third-party content as plug-in extensions to LLM-driven assistants (Alexa+ uses Claude; Google Assistant with Gemini; Apple Intelligence with the on-device LLM).

Architecture

                  ┌─────────────────────────────────────────┐
                  │   @enfinitos/sdk-renderer-core (TS)     │
                  │   resolve / event-ingest / health       │
                  │   AudioAttentionConstraint              │
                  └─────────────────────────────────────────┘
                                     ▲
                                     │
                  ┌──────────────────┴──────────────────┐
                  │   audio ts-core (AttentionTracker +  │
                  │   AudioClient + listening-session    │
                  │   resolver)                          │
                  └──────────────────┬──────────────────┘
                                     │
   ┌────────────────────┬────────────┴────────────┬─────────────────┐
   │                    │                          │                 │
┌──────────┐      ┌──────────┐               ┌──────────┐     ┌──────────┐
│ Smart    │      │ Podcast  │               │ Streaming│     │ App      │
│ Speaker  │      │ partners │               │ partners │     │ Intents  │
│ (Alexa+/ │      │(Megaph., │               │(Spotify, │     │ (Apple)  │
│ Google/  │      │ Acast,   │               │ Pandora) │     │ Siri+LLM │
│ Apple AI)│      │ Spreaker)│               │          │     │          │
└──────────┘      └──────────┘               └──────────┘     └──────────┘

Getting started

TypeScript core

import { AudioClient } from "@enfinitos/sdk-audio";

const audio = new AudioClient({
  apiBaseUrl: "https://api.enfinitos.com",
  authToken: jwt,
  deviceId: speakerHashedId,
});
await audio.start();

// Begin a listening session — required by the
// AudioAttentionConstraint contract.
const session = audio.beginListeningSession({
  attentionMode: "active-listening",
  channelKind: "podcast-pre-roll",
});

const asset = await audio.resolveNext({ surfaceId: "pre-roll" });
if (asset) {
  await audio.beginPlayback(asset);
  // After playback the wrapper calls reportPlayEnded with dwell.
}

audio.endListeningSession(session.id);

Alexa+ (voice-assistant LLM extension)

import { handler } from "@enfinitos/sdk-audio/smart-speaker/alexa-audio";

export const lambdaHandler = handler({
  apiBaseUrl: "https://api.enfinitos.com",
  authToken: process.env.ENFINITOS_TOKEN!,
  alexaSkillId: process.env.ALEXA_SKILL_ID!,
  llmExtension: {
    mode: "alexa-plus",
    claudeModel: "claude-3-5-sonnet-20240620",
  },
});

Apple App Intents (replacement for SiriKit)

import EnfinitOSAppIntents

struct PlayEnfinitOSPodcast: AppIntent {
    static var title: LocalizedStringResource = "Play EnfinitOS Podcast"
    func perform() async throws -> some IntentResult {
        let client = EnfinitOSAppIntentsClient.shared
        let asset = try await client.resolveNext(surfaceId: "podcast-pre-roll")
        try await client.beginPlayback(asset)
        return .result()
    }
}

2026 platform notes

Alexa+ (LLM-extension pattern)

• Alexa+ is Amazon's LLM-powered Alexa, GA-ed in early 2025 with Claude (Anthropic) as the primary reasoning model. The classic-Alexa template-based skill surface still works; the new path is an LLM extension: the operator registers a set of capabilities + a system-prompt-style description and Alexa+ decides when to invoke them.
• ASK SDK v3 (Node.js, Python) is the supported SDK for the classic flow. The LLM-extension path uses the Alexa Skill API REST endpoints + the Alexa+ "Skill-as-Action" registration.
• AudioPlayer interface is unchanged across the LLM and classic-template paths; the SDK targets AudioPlayer.Play, AudioPlayer.Stop, AudioPlayer.PlaybackStarted/Finished.
• Permissions card flow for user data (location, address) routes through the same OAuth-linked context.

Google Assistant with Gemini (Conversational Actions / Actions Builder)

• Actions on Google v3 SDK is end-of-life as of mid-2024. We do NOT use it.
• New surface: Conversational Actions via Dialogflow CX or Actions Builder API directly (@google/actions-style REST). Google Assistant invokes the action through Gemini's function-calling surface.
• Audio playback uses the Media response object inside the conversational webhook. No WebSocket; it's a turn-based REST pattern.

Apple App Intents (replacement for SiriKit)

• App Intents framework (iOS 16+, hardened in iOS 17) replaces SiriKit Intents Extension which is deprecated for new code. We do NOT use INExtension / INIntentHandler.
• The SDK uses the AppIntent protocol with Swift Concurrency (async throws).
• Apple Intelligence (announced 2024, GA 2026) is the on-device LLM; App Intents are the integration surface for Apple Intelligence's "act on apps" capability.

Podcast partners

Audio-streaming partners

Endpoint surface

Test plan

• TS core + smart-speaker / podcast / streaming bridges: Vitest.
• Apple App Intents: XCTest.