transports-wg/meetings/2026-02-18_weekly.md
2026-02-26 10:59:46 -07:00

5.5 KiB

2026-02-18 Transports WG (Weekly)

TL;DR

The meeting focused on core maintainer feedback for in-flight Specification Enhancement Proposals (SEPs) and refining the Multi-Round Trip Request (MRTR) proposal. PR 2260 (deprecating unsolicited server requests) is expected to pass, while PR 2243 (HTTP Headers) requires revisions to optimize header usage and generalize custom parameters. The bulk of the session centered on MRTR, particularly addressing SDK maintainer concerns regarding backwards compatibility and breaking changes. The group debated using capabilities versus protocol versioning to gracefully migrate clients to the new stateless flow. Technical clarifications were made regarding the use of Tasks for MRTR, handling race conditions idempotently, and generalizing the MRTR pattern beyond just tool calls. The group reached an 80-90% consensus to advance MRTR to a formal SEP, pending a few wording updates and a final review next week. Peter Alexander will also engage the Python and TypeScript SDK teams to ensure smooth architectural alignment.

Detailed Summary

Core Maintainer PR Updates

  • PR 2260 (Out-of-Context Elicitation): Caitie McCaffrey reported that the SEP to eliminate unsolicited server out-of-context elicitation and sampling is expected to pass the official core maintainer vote with only minor comments to address.
  • PR 2243 (HTTP Standardization): Core maintainers requested more time to review this SEP and provided three main points of feedback:
    1. Header Optimization: Can multiple headers be combined, or can headers like tool names be reused for methods/resources?
    2. Custom Parameters: Suggestion to generalize the custom parameter mirroring to cover any field in the MCP body via a key-value map, rather than restricting it to tool parameters.
    3. Encoding Prior Art: Requests for prior art or examples regarding the encoding strategy. Mike Kistler and Kurtis Van Gent will work on these revisions.

MRTR: Backwards Compatibility & SDK Impact

  • Breaking Changes: Peter Alexander relayed significant concerns from SDK maintainers about the MRTR proposal breaking existing client code.
  • Capabilities vs. Versioning: The group debated how to handle the migration. Mark Roth suggested leveraging protocol versioning to eventually deprecate the old flow. Simon Russell and Stephen Halter advocated for capability-based negotiation so clients can opt into the new MRTR flow, allowing servers to choose their implementation while clients support both during the transition.
  • Migration Plan: Caitie McCaffrey emphasized that a clear socialization and migration plan (potentially involving a migration bot) must accompany the SEP to help developers transition before the old flow is officially deprecated.

MRTR: Technical Clarifications

  • Generalizing the Pattern: Mike Kistler pointed out that the document leaned too heavily on tool calls. Mark Roth agreed to update the wording to clarify that the MRTR pattern applies to any client-initiated request (e.g., resources with templates), framed as a split between ephemeral/stateless patterns and persistent/stateful patterns.
  • Progress Notifications: Clarified that in the ephemeral flow, a server can still send progress notifications via the SSE stream before it sends the elicitation request (incomplete result) and closes the stream.
  • Encryption Examples: Kurtis Van Gent requested clarification on encryption requirements. Mark Roth confirmed that AESGCM and assigned JWTs are merely illustrative examples, not mandatory requirements; servers can use other encryption methods.
  • Load Shedding: Simon Russell highlighted a buried feature: a server can send an incomplete response with request state but no input requests, purely for load shedding (asking the client to resume later). Mark Roth will make this feature more prominent in the SEP.

MRTR: Tasks and Race Conditions

  • New Task Method: A new API method was introduced for clients to send input requests back to tasks. Luca Chang and Caitie McCaffrey explained that separating this from the tasks/get (status) call ensures consistent performance and better database observability for the status endpoint.
  • Handling Input Race Conditions: Kurtis Van Gent asked how the protocol handles multiple clients/threads submitting conflicting inputs to the same task. Mark Roth and Caitie McCaffrey argued against adding complex "declined" error states. Instead, the server should process inputs idempotently (e.g., taking the first valid submission). If an input is invalid, the task simply remains in the "input required" state.

Action Items

  • Mark Roth to update the MRTR SEP wording to generalize the pattern beyond tool calls.
  • Mark Roth to clarify in the MRTR SEP that AESGCM and JWT are just examples of encryption, not requirements.
  • Mark Roth to explicitly add the new method for sending input requests to tasks to the list of methods in the SEP.
  • Mark Roth to elevate the visibility of the "load shedding" feature (incomplete response with state but no input request) within the document.
  • Mark Roth to complete all document changes before next week's meeting for a final review.
  • Peter Alexander to meet with the Python and TypeScript SDK maintainers to discuss the MRTR proposal and architectural impacts.
  • Caitie McCaffrey & Group to begin working on the necessary schema changes to formally convert the MRTR document into a proper SEP.