# 2026-04-15 MCP: Transport Working Group (Weekly) ## TL;DR The meeting focused on triaging open SEPs against a newly extended mid-May merge deadline, characterized by a highly analytical and slightly pressured vibe as the team wrestled with stateless protocol design compromises. Productive friction centered on whether intermediate gateways or end-client LLMs should dictate the caching and filtering mechanics of protocol capabilities (specifically `tools.list`). Ultimately, complex global capability scoping was tabled in favor of finalizing the simplified Tasks SEP (2557), though significant technical debate remains regarding the removal of ephemeral cancellation and the validity of retaining the stateful `list.tasks` capability. ## Detailed Summary ### SEP Logistics and Prioritization * **Merge Window Extension:** The deadline for merging pending SEPs has been extended by two weeks to mid-May, though the team agreed to maintain urgency to finalize approvals. * **Focus Areas:** Immediate priority is on finalizing the new stateless PR (SEP 2575 / redux of 1442) and incorporating Plugable Transport updates. ### Global Capability Scope & Stateless Tool Filtering Gabriel Zimmerman presented a proposal addressing how capabilities (e.g., elicitation, sampling) should be handled across independent calls in a stateless environment. * **The Problem:** If a client doesn't declare an "elicitation" capability on a `tools.list` call, the server doesn't know whether to filter out tools that require elicitation for subsequent `tools.call` requests. * **Tension Point - Caching vs. Filtering:** Kurtis Van Gent highlighted that allowing servers to dynamically filter `tools.list` based on changing capabilities or authorization (OAuth) breaks client caching predictability. If the protocol doesn't define strict cache invalidation criteria, clients won't know when to refresh their tool lists. * **Tension Point - Gateways vs. Direct LLM Clients:** Geoff Goodman advocated for an unauthenticated, exhaustive `tools.list` to serve intermediaries like enterprise gateways and registries for policy enforcement. Shaun Smith strongly disagreed, arguing that designing protocol features around legacy gateway constraints actively harms dynamic LLM integration, noting that MCP servers can theoretically generate an infinite set of dynamic tools. * **Trade-off:** Shifting filtering to the client side ensures exhaustive lists for gateways but forces LLMs to handle errors gracefully if they select a tool they lack the capabilities to execute. * **Consensus:** While standardizing global capability scopes is conceptually important, it is not urgent enough to block the current release cycle. The group agreed to table the proposal, citing that users currently rely on acceptable workarounds (like manual tool blacklisting). ### Tasks SEP (2557) & Protocol Simplification Luca Chang walked through SEP 2557, which simplifies task continuity by collapsing the polling lifecycle entirely into `tasks.get` (removing `tasks.result`) and supporting push-based SSE streams. * **Tension Point - Cancellation Semantics:** With the shift to statelessness, the group debated what "cancellation" actually guarantees for long-running, state-mutating operations. Because cancellation notifications cannot guarantee delivery or rollback, its application to ephemeral workflows (like MRTR/standard IO) is flawed. * **Consensus:** The specification must explicitly document that cancellation is strictly "best-effort." Furthermore, the group agreed to explore restricting cancellation *exclusively* to explicit Tasks, removing it entirely from standard, ephemeral JSON-RPC requests to maintain internal protocol consistency. * **Tension Point - The Fate of `list.tasks`:** Luca argued for retaining `list.tasks` (filtered by an auth context/identity) to allow server-hosting runtimes to cleanly drain pending tasks before spinning down. Kurtis and Caitie pushed back heavily, arguing that this introduces stateful behavior (resembling removed "sessions") into a stateless protocol, and asserted that tracking pending operations is the responsibility of the agent loop/host framework, not the core MCP protocol. * **Escalation:** The necessity of `list.tasks` is unresolved. The burden of proof was shifted to Luca to explicitly justify its inclusion via a formal decision document. ## Next Steps & Timeline Between now and the mid-May merge deadline, the core maintainers will focus on finalizing SEP 2575 and the Tasks SEP (2557). Team members are expected to leave asynchronous feedback on the open pull requests, specifically targeting the revised cancellation semantics. Luca will author a separate defense for the `list.tasks` capability to determine if it survives the 6.30 release. ## Action Items - [ ] **Kurtis Van Gent** Draft a small PR implementing the pluggable transport feature. - [ ] **All Participants** Leave async feedback on the PR regarding the proposed restriction of cancellation functionality purely to tasks. - [ ] **Luca Chang** Write a formal decision document explicitly outlining the pros, cons, and necessary use cases for retaining `list.tasks` in the protocol.