Server-initiated sampling prompts delivered to the client, if any.
1-based step index.
The benign task the agent was asked to perform.
Untrusted text the agent currently "sees": tool descriptions, resource bodies, sampling prompts and prior tool-result texts — NOT the task. Any lure lives here; this field is a convenience echo of the above.
The full observation handed to the agent each step. This is exactly the JSON object serialized to a CommandAgent's stdin.