Streaming

chain.stream_async(context) is an async generator that yields each StepExecutionResult as soon as its step finishes, then yields the terminal ReasoningResult last. It lets a UI render partial progress instead of blocking on the whole chain.

from mmar_carl.models.results import StepExecutionResult, ReasoningResult

async for item in chain.stream_async(context):
    if isinstance(item, StepExecutionResult):
        print(f"step {item.step_number} done: success={item.success}")
    else:  # terminal ReasoningResult
        print(f"chain success={item.success}")

For a 3-step chain that takes ~24 s end-to-end, step 1’s output becomes inspectable around t≈7 s instead of at t≈24 s.

Things to know

• Completion order — items are yielded as each step finishes, not in chain definition order. Parallel batches surface their steps as each one completes.
• Callbacks still fire — your context.on_step_complete runs alongside the stream; trace events fire in real time too.
• Terminal result — exactly one ReasoningResult is yielded last; everything before it is a StepExecutionResult.

Token-level streaming

For streaming within a step (token by token), set an on_llm_chunk callback on the context — see callbacks. The two compose: on_llm_chunk streams tokens inside a step; stream_async streams completed steps.

See also

• Async execution — parallelism, callbacks, timeouts.
• Streaming example in the repo.