Allow a server to request model generations through its connected client while preserving user control over model choice, context, costs, and tool use. Build a safe sampling loop and inspect the messages exchanged at each step.
