96 lines
No EOL
5.4 KiB
Markdown
96 lines
No EOL
5.4 KiB
Markdown
# 2026-06-24 MCP: Transport Working Group (Weekly)
|
|
|
|
## TL;DR
|
|
The working group established an advancement strategy for the streaming
|
|
Specification Enhancement Proposal (SEP) draft and debated introducing
|
|
client-generated conversation correlation IDs to address routing and billing
|
|
concerns. Shaun Smith from Hugging Face presented the streaming pull request,
|
|
leading to discussions about task interaction and JSON RPC over Server-Sent
|
|
Events (SSE) challenges. Ladi Bitto from Microsoft proposed an agnostic
|
|
conversation identifier to resolve costly routing workarounds in Copilot Studio,
|
|
which Kurtis Van Gent agreed to synthesize into a formal problem statement for
|
|
core maintainers. Due to upcoming conferences and holidays, the next weekly
|
|
meeting was canceled.
|
|
|
|
## Detailed Summary
|
|
|
|
### Streaming Brief and Task Interaction
|
|
- Shaun Smith (Hugging Face) presented a Specification Enhancement Proposal
|
|
(SEP) pull request to address tool-calling streaming capabilities,
|
|
particularly for multimodal scenarios, to allow higher-resolution execution
|
|
steering and progress tracking.
|
|
- Mark Roth raised concerns about how streaming would interact with existing
|
|
tasks (e.g., handling cancellation cases) and suggested that streaming
|
|
capabilities might be built directly into tasks rather than introducing a
|
|
distinct mechanism.
|
|
- Shaun Smith noted that the protocol must clearly instruct users on when to use
|
|
streaming versus tasks, or whether tasks themselves should be extended.
|
|
|
|
### JSON RPC and Server-Sent Events (SSE) Challenges
|
|
- Kurtis Van Gent raised a technical concern regarding how auto-streaming would
|
|
work with JSON RPC, as JSON RPC typically requires a complete message before
|
|
processing.
|
|
- Luca Chang noted that while the current events primitive streams standalone
|
|
JSON RPC messages over an SSE stream, it does not manage event association or
|
|
ordering.
|
|
- Mark Roth pointed out that a unidirectional SSE stream presents challenges for
|
|
cancellation and bidirectional steering (e.g., redirecting sub-agents running
|
|
in tool-calling loops).
|
|
- The group agreed that concrete use cases, user journeys, and tool designs must
|
|
be documented early to ensure core maintainer alignment. Shaun Smith plans to
|
|
have a mergeable brief ready in two weeks.
|
|
|
|
### Client-Generated Conversation Correlation IDs
|
|
- Ladi Bitto (Microsoft) proposed introducing a client-generated conversation
|
|
correlation or thread ID to resolve routing issues and support session
|
|
management without re-introducing full protocol-level session handling.
|
|
- Kurtis Van Gent explained that changing the protocol for the current release
|
|
version is impossible because that ship has sailed, but it could target a
|
|
future version. He noted that core maintainers previously voted to remove
|
|
sessions, so a very strong case is required.
|
|
- Ladi Bitto and Shaun Smith discussed why Peter's proposed "state handles" are
|
|
insufficient, noting they are non-deterministic and introduce significant
|
|
financial costs due to extra tool calls, especially impacting platforms like
|
|
Microsoft Copilot Studio.
|
|
- Mike Kistler suggested quantifying the exact frequency and financial costs of
|
|
these state handle workarounds to present a stronger case to core maintainers.
|
|
- Kurtis Van Gent committed to writing a structured problem statement
|
|
summarizing the user journeys and technical limitations of state handles, to
|
|
be reviewed by the working group before presenting to core maintainers.
|
|
- Gabriel Zimmerman offered to brainstorm arguments internally, and agreed to
|
|
share Catie McCaffrey's contact information with Ladi Bitto to ensure
|
|
Microsoft's routing and cost concerns are represented in core maintainer
|
|
discussions.
|
|
|
|
### SDK Release Readiness
|
|
- Mark Roth requested an update on SDK preparation for the upcoming July
|
|
specification release.
|
|
- Kurtis Van Gent shared a support matrix showing that the Go and C# SDKs are
|
|
closest to completion, which Mike Kistler celebrated.
|
|
- Kurtis Van Gent added that Python SDK development is actively progressing
|
|
inside a 2.0 feature branch and is further along than the dashboard shows.
|
|
|
|
## Next Steps & Timeline
|
|
The next weekly meeting has been canceled due to the AI Engineer World's Fair,
|
|
holidays, and external time off. The group will resume in two weeks. In the
|
|
interim, Shaun Smith will incorporate feedback into the streaming brief and
|
|
refine the tool designs. Kurtis Van Gent will develop database and pagination
|
|
examples for the streaming brief and draft a problem statement document
|
|
summarizing the correlation ID use cases. Ladi Bitto will work on summarizing
|
|
Microsoft's correlation ID use cases and cost factors, while Gabriel Zimmerman
|
|
will provide Catie McCaffrey's contact information.
|
|
|
|
## Action Items
|
|
- [ ] **Shaun Smith** Update the streaming brief document incorporating
|
|
comments, pushing back on Peter's resources suggestion, and creating tool
|
|
designs.
|
|
- [ ] **Kurtis Van Gent** Develop database and pagination examples for the
|
|
streaming brief.
|
|
- [ ] **Kurtis Van Gent** Cancel next week's meeting and post a cancellation
|
|
notice in the Discord channel.
|
|
- [ ] **Kurtis Van Gent** Write a problem statement document summarizing the
|
|
correlation ID use cases and the technical limitations of state handles.
|
|
- [ ] **Ladi Bitto** Summarize the correlation ID use cases and provide a
|
|
concise summary of the financial cost factors of the state handle workarounds.
|
|
- [ ] **Gabriel Zimmerman** Share Catie McCaffrey's contact information with
|
|
Ladi Bitto. |