Rust: ergonomic SamplingHandler + typed MCP sampling request on sampling.requested

rust/src/generated/session_events.rs と rust/src/handler.rs から始め、SamplingRequestedData と既存の handler traits を比較します。次に、rust/src/session.rs の session-builder の配線と、rust/src/generated/rpc.rs の sampling RPCs を追跡します。型付きの sampling request と登録済みの SamplingHandler が連携して動作し、既存の low-level methods も引き続き利用できれば完了です。

Summary

The Rust binding exposes MCP sampling on the wire — the sampling.requested / sampling.completed events and the handle_pending_sampling (and execute_sampling / cancel_sampling_execution) methods — but there are two gaps for a consumer that wants to service MCP sampling requests itself (i.e. answer an MCP server's sampling/createMessage with its own model):

1. The request content isn't on the typed event. SamplingRequestedData carries only { requestId, serverName, mcpRequestId } — not the MCP CreateMessageRequest params (the messages, modelPreferences, systemPrompt, maxTokens, etc.) that a handler needs to actually produce a completion. So a consumer receives "a sampling request happened" but not what to sample.
2. There's no ergonomic handler. Every other consumer-serviced request has a first-class handler trait (PermissionHandler, ElicitationHandler, McpAuthHandler, UserInputHandler, …) wired on the session builder. Sampling has none — a consumer must subscribe to the raw sampling.requested event and call the low-level, Experimental handle_pending_sampling by hand.

Current state (public Rust binding)

• rust/src/generated/session_events.rs:

  pub struct SamplingRequestedData {
      pub mcp_request_id: serde_json::Value,
      pub request_id: RequestId,
      pub server_name: String,
  }
  pub struct SamplingCompletedData { pub request_id: RequestId }
  

  No sampling request params (messages / model preferences) are present.
• rust/src/generated/rpc.rs: handle_pending_sampling(UIHandlePendingSamplingRequest) -> UIHandlePendingResult, execute_sampling(...), cancel_sampling_execution(...) — all marked Experimental.
• rust/src/handler.rs: ergonomic handler traits exist for permission, elicitation, MCP auth, user input, exit-plan-mode, auto-mode-switch — but not sampling.
• rust/src/session.rs: those handlers are wired on the builder (e.g. mcp_auth_handler); there is no sampling equivalent.

Why this is needed

This is a developer-experience / completeness improvement for consumers that want to fulfill MCP sampling requests with their own inference instead of the default behavior. Today such a consumer:

• can't get the request content from the typed event (has to reach past SamplingRequestedData), and
• has to hand-wire the raw event + Experimental RPC instead of implementing one trait.

It is not a capability blocker (the low-level RPCs exist), so this is a quality-of-life ask, not urgent.

Proposed change

1. Expose the MCP sampling request params on SamplingRequestedData as a typed field (the CreateMessageRequest params: messages, modelPreferences, systemPrompt, includeContext, maxTokens, temperature, stopSequences, metadata), so a handler can read what to sample directly from the event.
2. Add an ergonomic SamplingHandler trait in rust/src/handler.rs mirroring McpAuthHandler:

   #[async_trait]
   pub trait SamplingHandler: Send + Sync + 'static {
       async fn handle(
           &self,
           session_id: SessionId,
           request_id: RequestId,
           request: SamplingRequest, // typed CreateMessageRequest params
       ) -> SamplingResult;          // completion, or None/err to reject
   }
   

   Wire it on the session builder alongside the other handlers, and have the binding translate the trait's result into the existing handle_pending_sampling call (and reject/cancel when the handler declines), so consumers never touch the raw event or the Experimental RPC directly.

Acceptance criteria

• A consumer can register one SamplingHandler on the session builder and receive the typed MCP sampling request (messages + model preferences), returning a completion or a rejection.
• The sampling request params are available as typed data (not only reachable via the untyped/low-level path).
• The existing low-level handle_pending_sampling / execute_sampling methods continue to work for consumers that prefer them.

References

• rust/src/handler.rs — existing handler-trait pattern (McpAuthHandler et al.) to mirror.
• rust/src/generated/session_events.rs — SamplingRequestedData / SamplingCompletedData.
• rust/src/generated/rpc.rs — handle_pending_sampling, execute_sampling, cancel_sampling_execution.
• MCP spec — sampling/createMessage and its CreateMessageRequest params.