feat: add 2026-03-18 weekly meeting notes (#29)
This commit is contained in:
parent
e46ff88e93
commit
09de045cb7
1 changed files with 78 additions and 0 deletions
78
meetings/2026-03-18_weekly.md
Normal file
78
meetings/2026-03-18_weekly.md
Normal file
|
|
@ -0,0 +1,78 @@
|
||||||
|
# 2026-03-18 Transports WG (Weekly)
|
||||||
|
|
||||||
|
## TL;DR
|
||||||
|
The meeting began with a discussion on the acceptable level of breaking changes
|
||||||
|
for the Multi-Round Trip Request (MRTR) approach, focusing heavily on SDK
|
||||||
|
backwards compatibility and the upgrade path for existing servers. The second
|
||||||
|
half of the meeting was dedicated to a fierce debate on the necessity of Data
|
||||||
|
Layer Sessions. Two opposing views emerged: one advocating for formal session
|
||||||
|
creation/deletion methods, and another proposing the complete omission of
|
||||||
|
sessions in favor of MRTR and model-managed identifiers. The lack of consensus
|
||||||
|
on the core purpose and lifetime of a session led to a challenge for attendees
|
||||||
|
to present compelling, non-transport-related use cases for sessions before a
|
||||||
|
final vote next week.
|
||||||
|
|
||||||
|
## Detailed Summary
|
||||||
|
|
||||||
|
### Client Changes & SDK Backwards Compatibility (MRTR)
|
||||||
|
* **The Upgrade Story:** Felix Weinberger (representing the SDK team) raised
|
||||||
|
concerns about how SDKs are supposed to maintain backwards compatibility when
|
||||||
|
servers upgrade to the new MRTR flow. He stressed that SDKs cannot magically
|
||||||
|
solve compatibility issues if the protocol introduces a breaking change.
|
||||||
|
* **The "Bottom-Left Quadrant" Problem:** The group discussed a scenario where a
|
||||||
|
newly built server (supporting the new horizontally scaled MRTR flow)
|
||||||
|
interacts with an older client that only supports the previous SSE-based
|
||||||
|
elicitation flow. Mark Roth argued this scenario doesn't matter because it
|
||||||
|
represents a stateful routing case that currently doesn't work anyway. Felix
|
||||||
|
Weinberger countered that the spec must explicitly state that this scenario is
|
||||||
|
unsupported (i.e., the server simply ignores the old client's elicitation
|
||||||
|
request) to prevent unexpected outcomes.
|
||||||
|
* **Dual APIs vs. Compatibility Shim:** To support clients interacting with both
|
||||||
|
old and new servers, the group debated two paths. Mark Roth strongly preferred
|
||||||
|
building a compatibility shim in the SDK so existing code continues to work.
|
||||||
|
Simon Russell and others argued that a "magic solution" is too complex,
|
||||||
|
suggesting the SDK should instead offer two distinct APIs (one for the old
|
||||||
|
flow, one for the new flow) forcing developers to handle the integration
|
||||||
|
explicitly.
|
||||||
|
* **Next Steps for MRTR:** Caitie McCaffrey will draft a problem statement
|
||||||
|
outlining the options for bridging the API gap. She will meet with the SDK
|
||||||
|
maintainers on Friday to whiteboard the client/server decisions and SDK API
|
||||||
|
changes.
|
||||||
|
|
||||||
|
### The Great Sessions Debate
|
||||||
|
* **The Pro-Sessions Stance:** Shaun Smith presented a proposal to add explicit
|
||||||
|
`session create` and `session delete` methods, along with a `session required`
|
||||||
|
error. In this model, the server generates a Session ID, and the client echoes
|
||||||
|
it back with every request, allowing the server to correlate tasks and
|
||||||
|
maintain opaque application state (e.g., a sandbox environment). Jordan
|
||||||
|
Zimmerman supported this, noting HTTP's 30-year precedent of using cookies for
|
||||||
|
similar handle-based correlation.
|
||||||
|
* **The Anti-Sessions Stance:** Peter Alexander argued that with the upcoming
|
||||||
|
stateless MRTR changes, the historical reasons for sessions (capabilities,
|
||||||
|
subscriptions, multi-turn flows) are gone. For application state (like a
|
||||||
|
shopping cart), he proposed that the model should call a tool (e.g.,
|
||||||
|
`create_basket`), receive an ID, and pass that ID into subsequent tool calls.
|
||||||
|
This removes the overhead of session lifecycle management and allows for
|
||||||
|
greater flexibility, such as sharing a state object across multiple
|
||||||
|
sub-agents.
|
||||||
|
* **The Core Disagreement:** Geoff Goodman highlighted that the primary issue
|
||||||
|
with sessions is the lack of consensus on session lifetime between the client
|
||||||
|
and server. Mark Roth noted that after six months, there is still no agreement
|
||||||
|
on what sessions are actually for, which heavily favors Peter Alexander's
|
||||||
|
proposal to omit them entirely to avoid cementing a flawed feature into the
|
||||||
|
protocol. Stephen Halter pushed back, arguing that removing the feature breaks
|
||||||
|
existing, working scenarios (like VS Code refreshing).
|
||||||
|
* **Next Steps for Sessions:** Kurtis Van Gent challenged the group to bring
|
||||||
|
concrete, non-transport-related use cases for sessions to the next meeting.
|
||||||
|
|
||||||
|
## Action Items
|
||||||
|
- [ ] **Caitie McCaffrey** to write up a problem statement with options for
|
||||||
|
bridging the gap between the current and future SDK APIs.
|
||||||
|
- [ ] **Caitie McCaffrey, Stephen Halter, Max Isbey, Felix Weinberger, and Mark
|
||||||
|
Roth** to meet on Friday to whiteboard and mock up options for SDK API
|
||||||
|
changes.
|
||||||
|
- [ ] **Kurtis Van Gent** to reserve 30 minutes at the start of next week's
|
||||||
|
meeting to definitively tackle the problem of sessions.
|
||||||
|
- [ ] **All** to review their use cases and verify if
|
||||||
|
sessions are still required assuming MRTR and task correlation issues are
|
||||||
|
fixed.
|
||||||
Loading…
Reference in a new issue