This method must only be used for tasks that "could" create a new SpeechHandle. When draining, every task created with this method wi

An ``INFERENCE_DONE`` event carrying the accumulated speech/silence — the shape the silero/inference VAD emits each inference window.

Construct a SpeechStream without triggering the _main_task WebSocket loop. Patches asyncio.create_task during __init__ so the stream doesn't t

Extract and unpack a coordinate pair from Anthropic's [x, y] format.

Send an event to the Ultravox WebSocket.

Create a Pydantic model from a function's signature. (excluding context types)

Patch env vars: keys with None get removed, others get set.

Stub a _RunningTask into the executor so policy methods can be tested without choreographing real execute() lifetimes. Caller must clean up the

A resolved prediction future, as if the transport already answered.

Run _run_worker and return the config that was active when run() was called.

Run emitter, collect events until we've seen expect_finals is_final events.

Low-level helper that serialises event_json and forwards it to the bidirectional stream. Args: event_json (str): The JSON payload

Create an APIStatusError from an HTTP status code. When the message carries extra detail beyond the standard reason phrase, both the message

Build identification headers for inference requests. Always includes User-Agent with SDK version and Python version. Includes X-LiveKit-Room-

Replace the downstream sink, transferring playback listeners and on_attached/on_detached state.

Parse a model string into a model and voice Args: model (str): Model string to parse Returns: tuple[str, str | None]: Model an

Verify the caller once per call. Tools that mutate the booking (modify, cancel) update or clear the cache themselves.

Yield to the event loop until ``predicate()`` is true or we run out of ticks. Replaces fragile ``await asyncio.sleep(0)`` counts now that the

Collect the entire stream into a single response. Example: ```python from livekit.agents import llm resp

Execute a function tool call and return the result.

Best-effort non-async read — safe if no await between read and use.

Developers building audio sinks must call this method when a playback/segment is finished. Segments are segmented by calls to flush()

Truncate the chat context to the last N items in place. Removes leading function calls to avoid partial function outputs. Preserves t

Wait for the past audio segments to finish playing out. Returns: PlaybackFinishedEvent: The event that was emitted when

Drain and close a session, suppressing errors from missing context.

Create a minimal AgentSession without TranscriptSynchronizer to avoid leaked tasks.

Construct a Soniox `SpeechStream` without spawning its real `_main_task`. Patches `asyncio.create_task` (the call site in `RecognizeStream.__init