Describe the bug

In namespace mode, sampling is used to map the user's intent to one of a set of base tools. The MCP protocol allows sampling requests to specify preferences for the language model to use but we currently leave this blank; we'll end up with whatever model is the default determined by the client. It is possible this goes to an expensive model (consuming the user's tokens/time) when a lower cost or free model would work just as well.

Expected behavior

When used for tool selection, sampling requests should specify free ("0x") or low-cost models.

Actual behavior

Sampling requests do not specify any preferred LLM models.

Reproduction Steps

Run Azure MCP in the default (namespace) mode and run a prompt that will invoke sampling.

Environment

No response