Request LLM sampling from the connected client inside a Python MCP tool (with deprecation caveats and fallback)

Documented steps

1. Know the status first: sampling (and roots) are deprecated as of the 2026-07-28 MCP spec revision (SEP-2577) — functional for at least 12 months, but new implementations are advised to call an LLM provider API directly. Treat this as a compat pattern. Docs: https://py.sdk.modelcontextprotocol.io/handlers/sampling-and-roots/
2. Preferred form is the resolver pattern: from mcp.server.mcpserver import Resolve, Sample; from mcp.types import CreateMessageResult, SamplingMessage, TextContent.
3. Write a resolver returning Sample(...): def draft_blurb(title: str) -> Sample: return Sample([SamplingMessage(role="user", content=TextContent(type="text", text=f"Write a one-sentence blurb for {title!r}."))], max_tokens=60) — Sample(messages, max_tokens=...) mirrors sampling/createMessage parameters.
4. Wire it into the tool via Annotated: async def blurb(title: str, draft: Annotated[CreateMessageResult, Resolve(draft_blurb)]) -> str: return draft.content.text if draft.content.type == "text" else "No blurb." Pass tools=/tool_choice= to Sample(...) to receive CreateMessageResultWithTools.
5. Fallback: if the client never declared the sampling capability (sampling.tools when using tools/tool_choice), the call fails with protocol error -32021 before any request is sent — catch it and fall back to a non-LLM default or a plain error message.
6. The legacy direct call ctx.session.create_message(...) still works only on connections with a back-channel and now raises a deprecation warning; the resolver form auto-adapts to the negotiated protocol version.