AdaptiveInterruptionDetector ignores LIVEKIT_INFERENCE_API_KEY when LIVEKIT_API_KEY is set

[ralonsom]

Summary

@livekit/agents exposes LIVEKIT_INFERENCE_API_KEY / LIVEKIT_INFERENCE_API_SECRET so users with a self-hosted SFU and a separate LK Cloud project for inference can authenticate against the inference gateway with different credentials. This works correctly for inference/llm.ts, inference/tts.ts and inference/stt.ts, but is broken in inference/interruption/interruption_detector.ts: when LIVEKIT_API_KEY is set (e.g. SFU creds), LIVEKIT_INFERENCE_API_KEY is silently ignored and the SDK signs the inference JWT with the SFU credentials, causing WebSocket connection rejected with status 401 and a fallback to VAD-based interruption.

Root cause

In agents/src/inference/interruption/defaults.ts the interruptionOptionDefaults object is initialised at module load with:

```ts
apiKey: process.env.LIVEKIT_API_KEY || "",
apiSecret: process.env.LIVEKIT_API_SECRET || "",
```

Then in agents/src/inference/interruption/interruption_detector.ts:

```ts
const { apiKey, apiSecret, ... } = { ...interruptionOptionDefaults, ...options };
// ...
lkApiKey = apiKey ?? process.env.LIVEKIT_INFERENCE_API_KEY ?? process.env.LIVEKIT_API_KEY ?? '';
```

Because apiKey already received process.env.LIVEKIT_API_KEY from the defaults, the ?? chain never falls back to LIVEKIT_INFERENCE_API_KEY.

The other three inference plugins use the correct precedence with ||:

```ts
// llm.ts, tts.ts, stt.ts
const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;
```

Reproduction

1. Self-host the LiveKit SFU (creds A).
2. Create an LK Cloud project and grab a separate set of creds (creds B).
3. Set both pairs in the agent environment:
   ```
   LIVEKIT_URL=ws://my-self-hosted:7880
   LIVEKIT_API_KEY=A_key
   LIVEKIT_API_SECRET=A_secret
   LIVEKIT_INFERENCE_API_KEY=B_key
   LIVEKIT_INFERENCE_API_SECRET=B_secret
   ```
4. Configure `interruption: { mode: 'adaptive' }` and place a call.
5. Worker registers fine against the SFU, but on the first overlap the adaptive detector fails with `WebSocket connection rejected with status 401` and falls back to VAD.

Suggested fix

Align defaults.ts with the precedence used by the other inference plugins:

```ts
apiKey: process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY || "",
apiSecret: process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET || "",
```

Or alternatively keep defaults.ts neutral (empty strings) and let the ?? chain in interruption_detector.ts actually fall through to the env vars.

Environment

• `@livekit/agents` 1.3.0
• Node 24
• LK Cloud project for inference; self-hosted SFU (`livekit/livekit-server:v1.11.0`)