Lazy streaming of a single predictor call — the DSPy streamify analog.

run/4 returns an Enumerable that, when enumerated, runs predictor name on program and yields %Dsxir.Stream.Event{} values as tokens arrive, ending with the final %Dsxir.Prediction{} as the last element:

program
|> Dsxir.stream(:answer, %{question: "..."}, listen: [:answer])
|> Enum.each(fn
  %Dsxir.Stream.Event{type: :field_delta, data: token} -> IO.write(token)
  %Dsxir.Prediction{} = prediction -> IO.inspect(prediction)
end)

opts accepts :listen (field-level streaming) and any other predictor opt; the :stream sink is supplied internally. :stream_timeout (default 60s) bounds the wait between events.

A single named predictor is streamed because only Dsxir.Module.Runtime.call/4 propagates per-call opts to the predictor — the multi-node executor does not — and per-predictor listening mirrors how DSPy attaches stream listeners.

Process model

The forward pass runs in a task under Dsxir.TaskSupervisor, replaying the caller's Dsxir.Settings.snapshot/0 (captured eagerly at call time) so the task resolves the same lm/adapter/metadata. The producer-to-consumer hand-off is a mailbox: the sink sends each event to the enumerating process, and the task's return value delivers the prediction last. The consuming process stays responsive; a forward-pass crash surfaces as a raised exception during enumeration rather than killing the consumer, and halting the stream early reaps the task.