3.17.0

Rasa_Sdk 3.17.0 (2026-06-23)

Features

• #1314: Added streaming support to CollectingDispatcher and ActionExecutor.

  Custom actions can now emit responses incrementally using three new async methods on CollectingDispatcher:

  • stream_start() — begins a streaming response.
  • stream_chunk(text, image, buttons, elements, attachment, json_message) — emits a response fragment; accepts the same rich-content fields as utter_message().
  • stream_end() — ends the streaming response.

  Whether actual streaming occurs is determined by the transport, not the action. The supported transports are gRPC and direct custom action execution.
  On streaming transports, chunks are forwarded to the caller immediately as they are produced.
  On non-streaming transports, the same action code works unchanged: chunks are accumulated
  and replayed as individual utter_message() calls when stream_end() is called, so no branching or conditional logic is needed in action code.

  ActionExecutor.run() accepts an optional sink: asyncio.Queue parameter. When provided,
  each streaming event (stream_start, stream_chunk, stream_end, stream_done) is placed in the queue
  as the action runs. Omitting sink keeps the existing unary behaviour. A convenience wrapper run_streaming(action_call, sink) is also available.

  The gRPC transport now implements the WebhookStream server-streaming RPC, yielding WebhookStreamEvent messages (ChunkStart, Chunk, ChunkEnd, final_result, error) as the action produces them. The Chunk proto message has been extended to carry rich content fields (image, custom, attachment, buttons, elements) in addition to text.

  If an action calls stream_start() and stream_chunk() but forgets to call stream_end(), the executor closes the stream automatically and logs a warning.
  The CollectingDispatcher.is_streaming_active property returns True when a stream has been opened but not yet closed.

• #1320: Added graceful barge-in support to streaming custom actions.

  When a user speaks over a streaming assistant response (barge-in), the voice channel can signal the action server via the new AckStreamChunks gRPC RPC. The action server stops producing chunks immediately and waits for the action to finish naturally so that any Rasa events (slot sets, etc.) are still returned.

  New CollectingDispatcher API:

  • cancel_stream() — called by the transport layer on receipt of AckStreamChunks. Sets is_streaming_cancelled to True and causes all subsequent stream_chunk() calls to be silently dropped, so the action stops producing output without needing to be forcefully interrupted.
  • is_streaming_cancelled property — True after cancel_stream() has been called. Action authors can optionally inspect this after stream_end() to suppress side effects that are irrelevant when the response was cut short.

  New gRPC RPC in ActionService:

  • AckStreamChunks(StreamChunkAck) returns (google.protobuf.Empty) — the voice channel sends the response_id of the active streaming response. The action server looks up the dispatcher in its internal registry, cancels the stream, and returns immediately. WebhookStream detects the cancellation on the next iteration, drains the internal queue until the action finishes, and yields a final_result (containing all Rasa events) to the client.

  The response_id field already present on ChunkStart, Chunk, and ChunkEnd events serves as the correlation key between the outbound stream and the inbound AckStreamChunks call.

  After a barge-in, WebhookStream waits up to 30 seconds for the action to finish and deliver final_result. The timeout can be configured via the ACTION_SERVER_STREAM_BARGE_IN_TIMEOUT_SECONDS environment variable. If the action does not complete within that window, an empty final_result is yielded so Rasa Pro can finish the barge-in handler normally, and the task is cancelled so a hung action cannot block the stream indefinitely.

Bugfixes

• #1341, #1342: Widened the websockets dependency constraint from >=10.0,<12.0 to >=10.0,<16 to allow compatibility with newer websockets releases.