The current phase (observability + tests).
A human preempted this agent: mute the rest of the current burst. The floor itself is the coordinator's to release.
The agent's turn ended (its final assistant transcript). Releases the floor — reclassifying a short
utterance as a backchannel so it is not billed against the loop cap — or records an overlay as the
backchannel it was. Always returns the gate to Idle.
What was said, when known (the measured burst duration fills in DurationMs).
Decides what to do with one chunk of the model's outbound audio.
The chunk's size in bytes (PCM16 mono).
Whether to forward, drop, or cut — and whether this chunk took the floor.
Decides what to do with one chunk of the model's outbound speech, given how long it plays: an avatar's media piece counts by its audio track's duration.
A chunk with no speech (0 ms: an avatar piece with video only, or one its driver marks idle, which the engine counts as 0 ms) never starts a burst and never keeps one going, so the face an avatar shows between answers can't take the floor, count as an agent's turn, or hold a muted burst open. Inside a burst it goes with it: forwarded while the agent holds the floor or overlays, dropped once the burst is muted.
How long the chunk's speech plays, in milliseconds.
Whether to forward, drop, or cut — and whether this chunk took the floor.
Returns to Idle with no coordinator side effects (teardown, barge-in).
Per-agent gate over a full-duplex model's outbound audio. One instance per bridged session.