Interface ActiveBridgeSession

How this session decides it was addressed (the mode actually in effect after capability resolution).

The voicing agent's MJ: AI Agents id — stamped on this agent's transcript lines for attribution.

The names/aliases this agent answers to (moderator roster + addressed-only fallback).

A short role/description of the voiced agent (moderator relevance context).

The voiced agent's MJ: AI Agent Runs id — tied to each moderator prompt run for observability.

The MJ: AI Agent Sessions row id the bridge is attached to.

The avatar stream's latest init segment, read: its audio track's timescale lets the floor gate count an avatar piece's speaking time. Every stream (a turn, a resumed or replaced model session) opens with its own init.

Whether the model session was already replaced because its avatar could not be shown (once per session).

This session's own bot participant id on the endpoint — the speaker label on its own transcript lines.

The concrete driver instance transporting media.

The server-side channel host wired for this session (Part A), or undefined when the caller supplied none. Held so teardown can close the session's channels.

Epoch-ms when this session connected — drives the max-session-duration cap of the stale sweep.

The context user the session runs as (for participant upserts / teardown writes).

Whether the model's blind auto-response is OFF for this session (multi-agent meeting). When true the bridge is the sole speech trigger — a Speak decision issues one RequestSpokenUpdate. When false the model auto-responds and the engine stays hands-off (forcing would double-fire).

Coalesces the caller's keypad presses into one note for the model (telephony only).

Set once the session is being torn down, so a model-loss signal during teardown is ignored.

Executes ONE server-channel tool call for this session, routed through the channel host. Bound by the runner-constructing layer to RealtimeSessionRunner.ExecuteServerChannelTool. undefined when no channel host was supplied. Never throws — resolves to a structured failure.

Safety timer that force-releases a held floor after FLOOR_MAX_HOLD_MS in case the agent's final transcript never arrives — so one agent can't wedge the room's floor closed. Cleared on a normal release.

Whether this session operates in full-duplex conversational mode.

Whether at least one HUMAN participant has been seen in the room. Auto-leave only fires after a human was present and then all humans left — so a bot that joins ahead of anyone (or a momentarily bot-only roster at startup) is never reaped prematurely. Set by the occupancy check off the driver roster.

Whether this session currently holds the room's speaking floor (multi-agent floor control). Set when the engine claims the floor before triggering speech in meeting mode, cleared when the agent's own final transcript lands (it finished) or on teardown. Single-agent rooms always hold trivially.

Detects human speech in inbound room audio so the room's agents yield to it; present with ActiveBridgeSession.TurnGate.

Whether THIS session is the room's transcript scribe. Exactly one session per room writes the unified meeting transcript (its own speech + everything it hears), so a multi-agent room records one copy, not N. Elected at connect (first session in the room); re-elected if the scribe leaves.

Epoch-ms of the last inbound activity (transcript) — drives the idle cap of the stale sweep.

The participant identity of the most-recent inbound AUDIO frame (when the provider diarizes). Used to attribute a 'user' transcript to the speaker who produced it — since the model transcript itself has no speaker label. Approximate (single-speaker-at-a-time); undefined until the first diarized frame.

Epoch-ms of the last write of the agent session row's LastActiveAt.

High-resolution timestamp (ms) when the most recent user turn completed, for transcript-to-next-audio latency telemetry.

Pending grace timer started when the last human leaves. If a human rejoins within the grace window (e.g. a page refresh) it is cleared; otherwise it fires AIBridgeEngine.StopBridgeSession so the bot leaves the empty room and its co-agent run finalizes. Cleared on any teardown.

The metadata provider bound to this session's entity operations.

Whether a model-session recovery is in flight; inbound audio is dropped meanwhile.

How many times the model session has been re-opened (capped by MAX_MODEL_RECOVERY_ATTEMPTS).

The model-side addressing latch the model's i_am_addressed signal feeds — present only in ModelSide mode.

Barge-in hook (see StartBridgeSessionParams.OnBargeIn).

End-of-session hook (see StartBridgeSessionParams.OnSessionEnded).

This agent's participation style for the room moderator ('proactive' | 'addressed-only').

The provider row backing this session.

The realtime session the driver's media plane is wired to.

Model-session recovery factory (see StartBridgeSessionParams.RecoverRealtimeSession).

The audio-only replacement factory (see StartBridgeSessionParams.RecoverRealtimeSessionWithoutAvatar).

The room key (driver ExternalConnectionId) this session is in — the unified-transcript room grouping.

The camera and screen sources whose frames reached the current model session (SourceID → the label the model was given). A source not in it is new to the model, which is told what it can now see before its first frame; when one in it ends, the model is told, by that label, that it can no longer see it (and only about sources it actually saw). Cleared when a recovered session replaces the model.

The server-executed tool definitions the session's channels contributed at connect (e.g. the Meeting Controls facilitator tools). Empty when no channel host / no contributing channels. The runner-constructing layer registers these as RealtimeSessionRunner.ServerChannelTools.

The AIAgentSessionBridge row id (the durable record).

The TARGET agent this co-agent voices (role/identity source for the moderator), when distinct from ActiveBridgeSession.AgentID.

This session's own transcript sink (see StartBridgeSessionParams.TranscriptSink), when supplied.

The last final transcript turns, kept so a re-opened model session can be told what was said.

The floor gate over this full-duplex model's outbound audio (see FullDuplexTurnGate); absent for turn-based models.

The per-session turn-taking policy gating generation.

Whether this session has been told how to take turns in a shared room (the framing note is sent once, when the room becomes multi-agent).

Executes the turn-taking host tools (i_am_addressed, yield_turn) for this session. Present for full-duplex / model-side sessions in a room; the layer that owns the model session binds it as its local tool handler (the tool DEFINITIONS are TURN_TAKING_TOOL_DEFINITIONS).