4.1 KiB
2026-02-04 Transports WG (Weekly)
TL;DR
The group focused on finalizing the Multi-Round Trip Request (MRTR) proposal to
prepare it for submission as a public Specification Enhancement Proposal (SEP).
Mark Roth led the review of the drafted SEP, clarifying the distinction between
ephemeral (stateless) and persistent (stateful/tasks) workflows. A key technical
debate centered on data structures for dependent requests (Map vs. Array) and
the role of server-side state. The group achieved consensus to keep the "State"
mechanism (related to proposal 1685) as a separate but parallel SEP to keep the
MRTR proposal focused. Additionally, the group identified a missing method name
in the persistent task workflow and agreed to define a specific method (e.g.,
tasks/provide_input) to handle client responses. The meeting concluded with an
agreement to push the MRTR proposal to a public SEP after incorporating the
discussed tweaks.
Detailed Summary
MRTR Proposal: Ephemeral vs. Persistent Workflows
- Clarification of Scope: Mark Roth presented the updated SEP, distinguishing between "ephemeral tool workflows" (independent requests, no server state) and "persistent tool workflows" (requiring server-side state/Tasks).
- State Accumulation: Caitie McCaffrey and Mark Roth discussed refining the document to focus on message patterns rather than over-defining server-side state storage, ensuring the spec describes how messages flow rather than why a server might store state.
- Schema Updates: The group proposed adding a new
incomplete responsetype to the JSON-RPC response schema. This allows any node to return a response indicating that more information (dependent requests) is needed to complete the operation.
Technical Decisions: Data Structures and Keys
- Map vs. Array: A significant debate occurred regarding the structure of dependent requests. Mark Roth advocated for a Map (using keys) to allow servers to safely associate responses with requests during rolling upgrades or restarts.
- Key Generation: Gabriel Zimmerman and Mike Kistler debated the complexity of generating these keys. The group discussed whether the server needs to see the full original request to validate the response or if a key/hash is sufficient.
- Decision: The group leaned towards allowing the server to use keys for association, noting that if full request context is needed, it can be handled via the separate State mechanism.
State Mechanism (Proposal 1685)
- Separation of Concerns: The group agreed to separate the "State" mechanism (passing opaque blobs/tokens back and forth, similar to HTTP cookies or hidden fields) into its own SEP (referencing proposal 1685).
- Parallel Tracking: While separate, both SEPs (MRTR and State) will be tracked and moved through the process together to ensure they work seamlessly, as State is often necessary for complex multi-round trip scenarios.
- Security: Simon Russell suggested encouraging the use of strings for state so they can be signed by the server to prevent client tampering.
Persistent Workflow (Tasks)
- Missing Method Name: The group identified a gap in the persistent workflow where the client sends task results back to the server (Step 7). Currently, this step lacks a defined JSON-RPC method.
- Naming Proposals: Mark Roth and Caitie McCaffrey discussed potential names
like
tasks/provide_inputortasks/provide_dependent_results. They agreed to select a name to ensure proper routing and logging, treating the "Input Required" state as a form of recoverable error. - Next Steps: The group agreed to move forward with the MRTR proposal as a public SEP, with the persistent workflow requiring a specific method name for the client response.
Action Items
- Mark Roth and Caitie McCaffrey to incorporate the agreed-upon tweaks (including the method name for tasks) and push the proposal to a public SEP.
- @gjz22 is going to present discussion around SEP-1685
- Caitie McCaffrey had requested that we try to check in on the other in-flight SEPs