# 2026-01-21 Transports WG (Weekly) ## TL;DR The group reached a consensus to move forward with deprecating unsolicited server-to-client requests (sampling and elicitations) on the "get stream" to simplify protocol design and address security concerns. This decision was supported by data showing negligible usage in existing repositories. Mark Roth presented a "Multi-Roundtrip Request" proposal designed to replace the current SSE-based elicitation flow with a stateless, request-response model where clients echo back state. The group discussed the technical implications of this change. Finally, the group identified a need to reconcile this new proposal with the existing "Tasks" mechanism and will invite the Tasks author (Luca) to the next meeting. ## Detailed Summary ### Deprecation of Unsolicited Requests * **Proposal to Remove:** Shaun Smith proposed assuming that unsolicited sampling and elicitations (server-to-client requests) are removed from the session track. He argued that eliminating these unsolicited communications resolves complex state transfer issues regarding cookies and session IDs. * **Usage Data:** Peter Alexander shared analysis results from scanning several thousand MCP repositories, finding only one instance of unsolicited sampling usage (which belonged to Shaun Smith). * **Decision:** The group agreed to draft a Specification Enhancement Proposal (SEP) to formally deprecate unsolicited requests on the /GET stream. * **Scope:** The SEP will specifically disallow servers from sending sampling or elicitation requests without an associated client request ID. ### Technical Decisions: Multi-Roundtrip Requests * **Problem Statement:** Mark Roth outlined the inefficiency of the current elicitation flow, where a server opens an SSE stream for elicitation, forcing the client to make a completely independent request that must be manually linked back to the original context by the server. * **Proposed Stateless Flow:** Mark presented a new pattern where the server immediately terminates the initial request with a response indicating "more information needed." The client then sends a completely new request containing the original information plus the elicitation response. * **State Management (Hidden Fields vs. Cookies):** * The group debated how to handle state persistence between the terminated request and the retry. * Gabriel Zimmerman and Simon Russell argued against using HTTP cookies, suggesting a "hidden field" mechanism (opaque blob) is more appropriate for per-tool-call state that should not persist globally. * Mark Roth agreed, noting that cookies fail when multiple tool calls run in parallel within the same session. * **SDK Impact:** Peter Alexander and Simon Russell acknowledged that while switching to this model will be painful for SDK maintainers to implement initially, the resulting architecture is significantly easier to maintain and conceptually simpler. ### Interaction with Tasks * **Overlap Analysis:** The group identified that the "Tasks" mechanism currently duplicates some functionality discussed in the multi-roundtrip proposal, specifically regarding polling and state retention. * **Unified Vision:** Caitie McCaffrey noted that Tasks and Tool Calls are similar but necessary; she described Tasks as a "promise over a network" and suggested that the handling of elicitation responses in both systems needs to align. * **Next Steps:** The group agreed to invite Luca (author of Tasks) to the next meeting to deeply dive into the relationship between the new stateless proposal and the existing Tasks implementation. ## Action Items - [ ] **Shaun Smith** to write a Specification Enhancement Proposal (SEP) to deprecate unsolicited server-to-client requests (sponsored by Peter Alexander). - [ ] **Mark Roth** to post the Multi-Roundtrip Request proposal to GitHub as a Pull Request for community feedback. - [ ] **Caitie McCaffrey** to write a strawman proposal detailing the overlap between "Tasks" and the current multi-roundtrip proposal. - [ ] **Kurtis Van Gent** to invite Luca to the next meeting to discuss the relationship with tasks.