Member Junction
    Preparing search index...

    Interface RealtimeConfigurationSettings

    The Realtime section — knobs the realtime drivers consume.

    interface RealtimeConfigurationSettings {
        ConnectionMaxSeconds?: number;
        FullDuplex?: boolean;
        IdleSignal?: RealtimeIdleSignal;
        Pricing?: RealtimePricingSettings;
        Reasoning?: RealtimeReasoningSettings;
        ReconnectBufferSeconds?: number;
        RequestedTracks?: readonly RealtimeTrackDescriptor[];
        SessionMaxSeconds?: number;
        Tooling?: RealtimeToolingSettings;
        TrustNativeResumption?: boolean;
        TurnDetection?: RealtimeTurnDetectionSettings;
    }
    Index

    Properties

    ConnectionMaxSeconds?: number

    Maximum duration in seconds of a single underlying provider transport connection (WebSocket or WebRTC call) before the provider closes or resets the socket (for example, 600s on Gemini Live WebSockets, 3600s on OpenAI Realtime, 7200s on xAI).

    Set on MJ: AI Models.ModelConfiguration and overridable on MJ: AI Model Vendors.ModelConfiguration. When elapsed (minus ReconnectBufferSeconds), the realtime framework triggers a proactive in-place reconnect on BaseRealtimeModel / BaseRealtimeClient.

    FullDuplex?: boolean

    Whether this realtime model natively supports full-duplex conversational audio (simultaneous speaking and listening with natural interruptions).

    When true, full-duplex models do not receive turn-taking tools or an energy-VAD floor gate, allowing the provider's native full-duplex model to handle natural turn transitions. Cascades AIModelType -> Vendor ModelDefaults -> AIModel -> AIModelVendor.

    IdleSignal?: RealtimeIdleSignal

    Which server signal means "the session has gone idle and deferred work may be flushed".

    Absent = 'turnComplete', which is every model MJ spoke to before Gemini 3.8 Live Extended Thinking. On a model with asynchronous reasoning, turnComplete arrives while the server is STILL reasoning and still issuing tool calls, so treating it as idle drains queued work at the wrong moment; those models report 'interactionStatus' instead.

    Prices for output the token cost rows do not cover, such as avatar video. Set it on the model-vendor row.

    Reasoning plane settings — dual delegation configuration. Absent defaults to 'local'.

    ReconnectBufferSeconds?: number

    Buffer in seconds before ConnectionMaxSeconds expiry at which the framework initiates a proactive reconnect. Defaults to 10 seconds when omitted.

    RequestedTracks?: readonly RealtimeTrackDescriptor[]

    Media tracks this session asks the model to establish, beyond audio.

    The request side of negotiation, and the reason "video off by default" is structural rather than a default value someone can forget. Absent or empty establishes audio only; a video track exists only because something asked for it. That matters concretely: Gemini 3.8 Live's own turn coverage defaults to including every video frame, billed, so a design where omission means "inherit the provider" would ship a silent cost.

    Requests are intersected with the model's SupportedInboundTracks / SupportedOutboundTracks; anything unsupported resolves to 'unsupported' so the caller falls back deliberately.

    SessionMaxSeconds?: number

    Maximum duration in seconds of a single provider credential or session ceiling when it differs from ConnectionMaxSeconds (for example, 1800s ephemeral token ceiling on Gemini Developer API vs. 600s per WebSocket connection, or 3600s client_secret TTL on xAI vs. 7200s WebSocket session ceiling). Overridable at the MJ: AI Model Vendors level.

    Tool-execution semantics this model permits. Absent = the profile's own defaults.

    TrustNativeResumption?: boolean

    Whether the provider's native session resumption handle/ID reliably restores prior conversation turns on this endpoint.

    • true: pass the native resumption handle/ID and do NOT re-inject conversation history (e.g., Gemini Enterprise / Vertex AI v1, xAI resumption.enabled + conversation_id).
    • false: feed RealtimeConversationMemory history into the reconnected session even if the provider emits a resumption handle (e.g., Gemini Developer API v1alpha ephemeral tokens).
    • null / undefined: defer to the driver's runtime default.

    Catalog-level turn-detection default for this model. Folded into the session Config bag as the turnDetection key BELOW the agent/app config cascade (realtime.session.turnDetection) and the runtime override — the catalog supplies the default, agents/apps/callers refine it.