Observer Events

Real-time event types delivered over the observer WebSocket connection during a voice session.

The observer WebSocket delivers real-time events during a voice session - transcripts, state transitions, tool calls, emotion analysis, latency timing, and call control. Each event has a type field that identifies it, and clients can use this field as a discriminator to handle events with type safety. This page is the authoritative catalog of observer events, grouped by category below. For workspace-level events delivered over SSE (calls, text sessions, surfaces, email delivery), see SSE Events.

Event Types

Session Lifecycle

Event Type	Description	Key Fields
session_info	Session metadata sent on connect	call_sid, service_id, workspace_id, current_state, direction, duration_seconds
session_start	Marks the beginning of a session	call_sid, session_id, service_id, workspace_id, initial_state, mode, trace_id
session_end	Marks the end of a session	call_sid, session_id, duration_s, turns, completion_reason, final_state, has_recording

Runtime Lifecycle

Event Type	Description	Key Fields
actor_allocated	The call was assigned an isolated voice runtime	gameserver, node, alloc_latency_ms
media_attached	A caller, agent, or operator media leg attached to the runtime	leg
actor_shutdown	The isolated voice runtime was released	reason, lifetime_s

Transcripts

Event Type	Description	Key Fields
user_transcript	Caller speech, finalized	transcript, state, speaker_id, speaker_role, stt_confidence, eot_confidence, is_barge_in
agent_transcript	Agent speech, finalized	transcript, action, interrupted, filler_type (optional), speaker_id, speaker_role, streaming
agent_transcript_delta	Incremental agent speech token	delta

The filler_type field on agent transcripts is one of: empathy, receipt, working.

State and Navigation

Event Type	Description	Key Fields
state_transition	Agent moved between context graph states	previous_state, next_state, transition_type, annotation

Tools

Event Type	Description	Key Fields
tool_call_started	A tool call began	tool_name, call_id, input, parent_call_id, integration_name, endpoint_name, protocol
tool_call_completed	A tool call finished	tool_name, call_id, output (truncated), succeeded, duration_ms, error_message

Latency and Timing

Event Type	Description	Key Fields
latency	End-to-end latency breakdown	e2e_ttfb_ms, engine_ms, nav_ms, render_ms, audio_ttfb_ms, continuation, is_greeting
nav_timing	Navigation and render timing detail	nav_ms, render_ms, total_ms, input_tokens, output_tokens, model, state, retry_attempts, used_fallback, selected_action

Emotion and Empathy

Event Type	Description	Key Fields
emotion	Caller emotion analysis	dominant, valence, arousal, trend, coherence, scores, speaker (acoustics)
compound_emotion	Blended emotion scores for a turn	compounds (list of name/score pairs), turn_index
empathy_classified	Empathy tier selection	tier, tier_name, should_pause, suppress_filler, filler_type, empathy_baseline

Barge-In and Participants

Event Type	Description	Key Fields
barge_in	Caller interrupted the agent	interrupted_text, discarded_texts, total_barge_ins
participant_joined	A participant entered the conference	participant_id, participant_role, conference_sid
participant_left	A participant left the conference	participant_id, participant_role, conference_sid

Call Control

Event Type	Description	Key Fields
forward_call_resolved	Call forwarding completed	target_number, location_name, result, error
speaker_muted	Speaker mute state changed	muted
injected_event	An external event or guidance was injected into the session	message, sender, event_type

Voice Context

Event Type	Description	Key Fields
voice_context_applied	Voice parameters updated	tts_emotion, tts_speed, tts_volume, filler_enabled, emotion_detection_enabled, reasoning

Keepalive

The stream also emits a periodic ping event with an empty payload to keep the connection alive. Clients can safely ignore it.

Common Fields

Every event frame on the observer stream is a JSON object with this envelope:

Field	Type	Description
type	string	Event type discriminator
data	object	The event-specific fields listed in the tables above
timestamp	number	Unix epoch seconds
seq	integer	Monotonic per-call sequence number, used for ordering and late-join replay

The call itself is identified by the WebSocket URL path, not by an envelope field. call_sid appears inside the data payload of session lifecycle events only, and the call_id field on tool_call_started / tool_call_completed is the tool-call invocation ID, not the call identifier. The connect-time session_info snapshot and the ping keepalive omit timestamp and seq.

Participant Roles

The participant_joined and participant_left events use a participant_role field with these values:

Role	Description
caller	The end user on the call
agent	The AI voice agent
operator	A human operator who joined via escalation

Empathy Tiers

The empathy_classified event includes a tier_name field with these values:

Tier Name	Description
FUNCTIONAL	Standard conversational tone
LIGHT_TOUCH	Brief acknowledgment of caller emotion
FULL_EMPATHY	Extended empathetic response
HOLD_SPACE	Pause and hold space for the caller

Emotion Trends

The emotion event includes a trend field indicating the direction of the caller's emotional state:

Trend	Description
improving	Caller emotion is trending more positive
stable	Caller emotion is not changing significantly
deteriorating	Caller emotion is trending more negative

PHI Annotations

Fields that may contain protected health information are annotated in the schema with x-phi: true. Consumers that persist or log observer payloads can use this annotation to apply redaction or access controls:

Event	Field
user_transcript	transcript
agent_transcript	transcript
agent_transcript_delta	delta
tool_call_started	input
tool_call_completed	output
forward_call_resolved	target_number
session_info	caller_id
barge_in	interrupted_text, discarded_texts
emotion	verified_speaker_id
participant_joined	participant_name, display_name, phone_number
participant_left	display_name, phone_number

Handling Unknown Events

New event types may be added in future releases. Clients should ignore event types they do not recognize rather than treating them as errors. This ensures forward compatibility as the observer stream is extended.