Keep AutoDetect channel open after provisional SSE failure

[PranavSenthilnathan]

Motivation and Context

When AutoDetect tries Streamable HTTP and then provisionally falls back to SSE, both transports use the message channel owned by the AutoDetectingClientSessionTransport.

If the provisional SSE connection fails, SseClientSessionTransport currently completes that shared channel. Connect-time protocol fallback may subsequently retry with initialize using the same AutoDetect instance, but its reader can no longer receive the response because the channel has already completed.

Streamable HTTP already distinguishes provisional use from an adopted transport. This change gives SSE equivalent behavior.

This is a prerequisite for handling the AutoDetect scenario in #1765.

Changes

• Track whether SSE has been adopted based on a successful initial POST.
• Leave AutoDetect’s message channel open when a provisional SSE attempt fails.
• Preserve existing channel completion behavior for explicit SSE and successfully adopted SSE transports.
• Add a regression test that fails Streamable HTTP and provisional SSE, then successfully retries using the same AutoDetect transport and message channel.

No public API is added or changed.

How Has This Been Tested?

• dotnet build
• AutoDetect transport tests on net10.0, net9.0, net8.0, and net472
• Full core test suites

The ASP.NET conformance suite could not start locally because Node.js 21.6.2 does not provide fs.globSync, which is required by the pinned conformance package.

Breaking Changes

None.

This pull request description was generated with GitHub Copilot.

[Copilot]

Pull request overview

This PR fixes an AutoDetect transport failure mode where a provisional SSE fallback attempt could complete the shared message channel, preventing subsequent retries (e.g., retrying initialize) from receiving responses via the same AutoDetectingClientSessionTransport. It aligns SSE behavior with Streamable HTTP’s “provisional vs adopted” distinction.

Changes:

• Track when SSE is “adopted” (after a successful POST) and only complete the shared channel on disconnect once adopted.
• Update SSE disconnect handling to avoid completing AutoDetect’s shared channel during provisional SSE probe failures.
• Add a regression test covering Streamable HTTP failure + provisional SSE failure + successful retry using the same AutoDetect instance/channel.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
tests/ModelContextProtocol.Tests/Transport/HttpClientTransportAutoDetectTests.cs	Adds regression coverage ensuring the shared message channel remains usable after provisional SSE failure in AutoDetect mode.
src/ModelContextProtocol.Core/Client/SseClientSessionTransport.cs	Introduces “adoption” tracking and conditional disconnect completion to keep AutoDetect’s shared channel open during provisional SSE failures.

[tarekgh]

nit: this new guard is only covered for the non-adopted path. Consider adding a test for the inverse case as well: when SSE has been adopted (a POST succeeds so _sseAdopted is true), a later disconnect should still complete the shared channel. That protects against a future change that over-suppresses SetDisconnected for an adopted SSE transport. The existing SSE tests assert message exchange rather than completion on disconnect, so this path is currently untested.