Skip to content

@livekit/agents@1.9.5

Choose a tag to compare

@github-actions github-actions released this 28 Oct 17:46
· 59 commits to main since this release
22da3e8

Patch Changes

  • Type AssemblyAI TTS models and provider options for LiveKit Inference. - #2894 (@rosetta-livekit-bot)

  • Add DecisionModel for provider-neutral batched probability, choice, and score evaluation. - #2735 (@rosetta-livekit-bot)

  • Preserve job context on prewarmed loggers and explicit source locations on child loggers. Report the log caller when OpenTelemetry Pino instrumentation is enabled. - #2882 (@chenghao-mou)

    Reuse each supervisor's logger and bind its job context once at launch.

  • Type language_hint on the inference DeepgramOptions, so a deepgram/nova-3 caller can set modelOptions: { language_hint: ['en', 'es'] } without a type error, and widen DeepgramFluxOptions.language_hint to string | string[] for parity with the other providers. No wire change from the typing: the plugin already spreads modelOptions into the session's extra. - #2826 (@adrian-cowham)

    The plugin now also leaves off the wire a language_hint Deepgram cannot apply, instead of sending one the gateway silently ignores; the key is gated for Deepgram models only, so a hint another provider is given still reaches that provider. Deepgram applies the hint only in Nova-3 multi language mode, so on a non-Flux Deepgram model with any other language (or no language at all) the key is gated on the outgoing frame — both when the session opens and in a mid-session session.update — and a warning names the option, the model and the language, once per session. Only the frame is gated: modelOptions is never modified, so a later language: 'multi', in the same update that supplies the hint or in one of its own, sends the hint again. deepgram/flux-* models keep the hint under any language, since the Flux provider forwards it, and the hint still reaches Deepgram only while the gateway's deepgram_stt_language_hint flag is open.

  • Report running jobs to the server after reconnecting, so it keeps tracking them - #2865 (@davidzhao)

  • Include source code locations and agent names in structured job logs. - #2842 (@andrewnitu)

  • Send a session summary and the session duration in the recording upload header. The summary is the session report JSON without chat history and events. sessionReportToJSON now also emits agent_name, duration_ms, user_turns, errors and close_reason (a livekit.agent.CloseReason name), with null for unknown values, and takes options to omit the chat history and events. SessionReport gains agentName. - #2824 (@jmcclanahan)

  • Claim aligned transcripts for AssemblyAI TTS inference unless word timestamps are disabled. - #2843 (@rosetta-livekit-bot)