Use nested event type when routing transcript chunks

handleWebsocketMessage checks eventData.type === CC_EVENTS.REAL_TIME_TRANSCRIPTION, but this handler otherwise treats notifications as wrapped payloads where the actionable event is in eventData.data.type (it re-emits eventData.data.type just above). If transcript notifications arrive in that common wrapped format, this condition never matches, so task.emit(CC_EVENTS.REAL_TIME_TRANSCRIPTION, ...) is skipped and task-level listeners never receive live transcript updates.

Useful? React with 👍 / 👎.

Make AI feature lookup non-blocking during config load

getAgentConfig() now waits on aiFeatureResourcesPromise inside Promise.all, and getAIFeatureResources() throws on any non-200. That means a 404/403/temporary failure on /v2/ai-feature will reject the entire config load and fail agent registration, even though this is an optional transcript feature and the parser already has a disabled fallback when feature data is missing. This introduces a full login outage path tied to an ancillary API.

Useful? React with 👍 / 👎.

Emit transcript alias instead of firing same event twice

In TaskManager.setupTaskListeners, the REAL_TIME_TRANSCRIPTION branch emits CC_EVENTS.REAL_TIME_TRANSCRIPTION twice, so every transcript chunk is delivered twice to normal listeners (for example the sample app appends each chunk twice), while the intended backward-compatible alias is never emitted. The inline comment says the second emit is an alias, and the new unit test also expects 'realtimeTranscription', so this currently causes duplicated processing and misses alias subscribers.

Useful? React with 👍 / 👎.

Resolve transcript notifications against stored task IDs

This branch looks up the task by message.data.data.conversationId, but every write to taskCollection in this class uses payload.interactionId. The new HistoricTranscriptsResponse type in src/types.ts:866-870 also models conversationId separately from interactionId, so any deployment where those IDs differ will drop every realtime transcript chunk because eventContext.task is undefined and nothing gets emitted to the task.

Useful? React with 👍 / 👎.

Don't stop transcription on every participant exit

PARTICIPANT_LEFT_CONFERENCE is not equivalent to the conversation ending: src/services/task/state-machine/guards.ts:145-169 uses that event to downgrade from conference back to CONNECTED when the current agent and customer are still on the main leg. Mapping it to STOP means a consulted agent leaving a three-way call will shut transcription off for the rest of the live customer conversation.

Useful? React with 👍 / 👎.

Gate transcript control requests to voice interactions

requestRealTimeTranscripts() only checks the backend event type before calling the AI assistant. The task manager already processes AGENT_CONTACT_ASSIGNED for chat tasks too (see test/unit/spec/services/task/TaskManager.ts:1809-1835), so normal chat/email assignment and wrapup events will now POST GET_TRANSCRIPTS even though there is no audio stream to transcribe. That turns every digital interaction into avoidable AI-assistant errors and log noise.

Useful? React with 👍 / 👎.

Start transcript streaming on AGENT_CONTACT hydration

TRANSCRIPT_EVENT_MAP only sends START for AGENT_CONTACT_ASSIGNED and consult events, but TaskManager.handleAgentContact() explicitly recreates already-active tasks from AGENT_CONTACT during multi-session/reconnect recovery. In that path the call is already live and may never emit another AGENT_CONTACT_ASSIGNED, so after a refresh/relogin the SDK rebuilds the task but never posts GET_TRANSCRIPTS START, and the agent stops receiving live transcript chunks for the rest of the interaction.

Useful? React with 👍 / 👎.

What's the reason for keeping the values same?

Stop transcript streaming on terminal transfer/end paths

STOP is only mapped for AGENT_WRAPUP, AGENT_CONSULT_ENDED, and PARTICIPANT_LEFT_CONFERENCE, but several voice flows finish without any of those events. contact.ts binds blind/vteam/consult transfer success directly to AGENT_*TRANSFERRED, and the task state machine can move TRANSFER_SUCCESS, TRANSFER_CONFERENCE_SUCCESS, or some CONTACT_ENDED cases straight to WRAPPING_UP/TERMINATED. In those scenarios the agent has left the conversation, but no stop request is ever posted, so the transcript stream can remain active after the task is effectively over.

Useful? React with 👍 / 👎.

Skip transcript control when realtimeTranscripts is disabled

requestRealTimeTranscripts() only checks for an ApiAIAssistant instance before posting GET_TRANSCRIPTS. The SDK already loads agentConfig.aiFeature, and ApiAIAssistant.fetchHistoricTranscripts() refuses to run when realtimeTranscripts.enable is false, so this looks like an optional feature. As written, tenants where /v2/ai-feature is empty or has realtimeTranscripts.enable === false will still send transcript control requests on every mapped voice event, creating avoidable AI-assistant errors/noise for every call.

Useful? React with 👍 / 👎.

Maybe we can keep the Flag suffix

It will be multiple flags not just real time transcript

Right, I was talking about the variable name - aiFeature to aiFeatureFlags

Can wccApiGatewayUrl be hostname as well as a full url ?

It's hostname

Then we wouldn't require the try-catch. The code in the catch should be enough and the name wccApiGatewayUrl should also be something like wccApiGatewayHostname

What's the reason for keeping the values same?

Should we keep the event name consistent with other event names ?

This event is received differently in terms the payload structure is different, we are not keeping it as part of task object. It is independently present inside cc object so kept the name different in that manner

Should we add a comment in the code describing why we have kept it different ? will be easy for other devs to understand

Remove undeclared error enum from transcript fetch catch

fetchHistoricTranscripts() references AI_ASSISTANT_ERRORS.BASE_URL_NOT_AVAILABLE, but AI_ASSISTANT_ERRORS is not defined or imported in this module. Any failure that enters this catch block (for example a network error or base-URL resolution failure) will throw a new ReferenceError at this line, masking the real cause and bypassing the existing getErrorDetails(...) error normalization path.

Useful? React with 👍 / 👎.

Guard transcript requests until agent ID is initialized

requestRealTimeTranscripts() sends transcript control events with this.agentId but never verifies that it has been set. TaskManager starts consuming websocket events immediately, while cc.connectWebsocket() sets taskManager.setAgentId(...) only after awaiting getAgentConfig(...); during that window (for example reconnect/hydration with active interactions), sendEvent(...) can be called with an undefined agentId, causing failed START/STOP requests and missed transcript streaming.

Useful? React with 👍 / 👎.

Guard interaction lookup when data is missing

prepareEventContext now accepts top-level event types via message.data?.type || message.type, but it immediately dereferences message.data to compute interactionId. If a websocket notification arrives as a top-level event without a data object (for example a bare {"type":"Welcome"} payload), this will throw at runtime and abort message handling for that event path. Add a null check before reading message.data.interactionId / message.data.data.conversationId so non-task top-level events are safely ignored.

Useful? React with 👍 / 👎.