Question regarding "step-level forwarding" and the definition of per-agent steps (S)

[arlisgreat]

Hi, I'm trying to better understand the architecture of StreamMA. I have two specific questions regarding the "step-level" design:

1. Interpretation of "step-level forwarding": The project documentation mentions "step-level forwarding." Could you clarify how this is implemented? Does it involve a specific token-level or block-level interception in the inference process?
2. Scaling dimension S: You mention that increasing "per-agent steps" ($S$) is a key scaling dimension, which essentially suggests forcing agents to break down their reasoning into more granular, refined steps.

Could you provide more details on how this is achieved through external prompting or protocol design? Specifically, what mechanisms are used to ensure the model adheres to this finer-grained step-by-step reasoning during the forward pass?Any insights or pointers to the relevant code sections would be greatly appreciated. Thanks!