83 lines
No EOL
4.9 KiB
Markdown
83 lines
No EOL
4.9 KiB
Markdown
# 2026-03-11 Transports WG (Weekly)
|
|
|
|
## TL;DR
|
|
The meeting centered on two major protocol design topics: the schema scope for
|
|
Multi-Round Tool Results (MRTR) and a proposal to eliminate Data Layer Sessions
|
|
entirely. Regarding MRTR, the group debated whether incomplete response features
|
|
should be generic (applicable to all requests) or scoped only to specific
|
|
contexts like tool calls. The consensus favored a generic approach,
|
|
acknowledging the trade-off in SDK complexity, though the question will be
|
|
escalated to David (DSP) for final guidance. The second major discussion
|
|
involved Peter Alexander's "anti-sessions" proposal, which advocates for
|
|
replacing formal protocol sessions with application-specific IDs passed within
|
|
tool calls. This model-maximalist approach gained surprising traction among
|
|
attendees for its simplicity and flexibility. The group agreed to draft a formal
|
|
SEP for this sessionless approach and present both options (sessions vs. no
|
|
sessions) to the core maintainers for early feedback.
|
|
|
|
## Detailed Summary
|
|
|
|
### MRTR Schema Design: Generic vs. Scoped Features
|
|
* **The Dilemma:** Caitie McCaffrey opened the discussion on whether MRTR
|
|
features (like `request state` and `incomplete response data`) should be added
|
|
to the generic result object (applying to all requests) or scoped strictly to
|
|
specific requests (like tool calls).
|
|
* **Arguments for Scoping:** Mark Roth argued that allowing incomplete responses
|
|
on requests where they logically don't make sense (e.g., `ping` or `task/get`)
|
|
forces all clients to handle edge cases unnecessarily. If the protocol allows
|
|
it, clients must support it.
|
|
* **Arguments for Genericity:** Luca Chang and Caitie McCaffrey countered that
|
|
special-casing features introduces bugs and limits future innovation. Mike
|
|
Kistler (speaking for SDK maintainers) noted that a generic implementation at
|
|
the lower layers of the SDK is actually simpler to build.
|
|
* **Consensus & Escalation:** Caitie McCaffrey will draft a summary of the
|
|
options and escalate the question to David (DSP) for a final ruling based on
|
|
protocol design principles.
|
|
|
|
### The "Anti-Sessions" Proposal
|
|
* **The Pitch:** Peter Alexander presented an argument to eliminate Data Layer
|
|
Sessions from the MCP protocol entirely. Instead of `session create`/`destroy`
|
|
mechanics, applications would use specific tool calls (e.g.,
|
|
`create_shopping_cart`) to generate application-specific IDs, which the model
|
|
would then pass into subsequent tool calls.
|
|
* **Benefits:** This approach simplifies the protocol, removes the overhead of
|
|
session lifecycle management for lightweight sub-agents, and provides greater
|
|
flexibility (e.g., sharing a single "basket ID" across multiple sub-agents or
|
|
having multiple active state objects at once).
|
|
* **The "Model Maximalist" View:** Caitie McCaffrey and Kurtis Van Gent raised
|
|
concerns about whether cheaper or open-source LLMs could reliably handle
|
|
passing these IDs. Peter Alexander argued from a "model maximalist"
|
|
perspective, stating that models are increasingly proficient at instruction
|
|
following and variable tracking.
|
|
* **Support:** Mark Roth, Mike Kistler, and Geoff Goodman expressed strong
|
|
support for the proposal, agreeing that sessions are often confusing and that
|
|
leveraging existing tool call mechanisms is more elegant and scalable.
|
|
* **Backward Compatibility:** While acknowledging this is a major breaking
|
|
change, the group agreed that removing unnecessary complexity now is
|
|
preferable to maintaining a flawed feature.
|
|
|
|
### Next Steps & Timeline
|
|
* **Drafting the SEP:** Peter Alexander will write a formal SEP for the
|
|
sessionless proposal by the end of the week.
|
|
* **Core Maintainer Review:** Because the core maintainer meeting precedes the
|
|
next Transports WG meeting, the group will vote asynchronously on Peter's
|
|
draft. Both the original sessions proposal (by Shaun Smith) and the
|
|
anti-sessions proposal will be presented to the core maintainers as an "early
|
|
preview" to build consensus before Dev Days.
|
|
* **Luca's Roadmap:** Luca Chang is out for two weeks. Kurtis Van Gent will
|
|
ensure Luca's priority task-related SEPs ("jump scares" and "task continue")
|
|
are surfaced at the next core maintainer meeting.
|
|
|
|
## Action Items
|
|
- [ ] **Caitie McCaffrey** to draft a summary of the generic vs. scoped MRTR
|
|
options and escalate to David (DSP) via Discord for guidance.
|
|
- [ ] **Peter Alexander** to write a formal SEP detailing the proposal to remove
|
|
sessions from the MCP protocol.
|
|
- [ ] **Kurtis Van Gent** to follow up with Shaun Smith to ensure the original
|
|
sessions SEP is ready to serve as a counter-argument to the anti-sessions
|
|
proposal.
|
|
- [ ] **Caitie McCaffrey** to initiate an asynchronous thread for the WG to
|
|
review and vote on the two session proposals before the core maintainer
|
|
meeting.
|
|
- [ ] **Kurtis Van Gent** to follow up with Luca Chang regarding the priority
|
|
task-related SEPs before Luca's leave. |