SEP-2575 HTTP status mapping is timing-sensitive under load

[PranavSenthilnathan]

Description

RawHttpConformanceTests.July2026Post_MissingRequiredCapability_Returns400 intermittently receives HTTP 200 instead of HTTP 400.

The server correctly produces a MissingRequiredClientCapabilityException and sends the corresponding JSON-RPC error. However, the HTTP status can already be committed as 200 before that error reaches the response callback.

Evidence

The failure occurred in:

• https://github.com/modelcontextprotocol/csharp-sdk/actions/runs/30440416457/job/90538136230
• Expected status: BadRequest
• Actual status: OK

The test log shows:

1. requires_sampling threw MissingRequiredClientCapabilityException.
2. The handler took approximately 219 ms.
3. Kestrel completed the request with HTTP 200.
4. The process did not crash; --blame-crash only attached its standard collector.

The same test previously failed with the same 200-versus-400 result on an earlier workflow attempt and then passed when rerun:

• Failed attempt: https://github.com/modelcontextprotocol/csharp-sdk/actions/runs/30432932916/job/90514051740
• Successful retry: https://github.com/modelcontextprotocol/csharp-sdk/actions/runs/30432932916/job/90523495608

The failure has occurred under different configurations and target frameworks, while the other frameworks in the same jobs passed.

Root cause

StreamableHttpPostTransport defers committing response headers for a fixed 250 ms window:

internal static readonly TimeSpan DeferredHeaderFlushGrace =
    TimeSpan.FromMilliseconds(250);

The timer starts after the request is queued for dispatch. If dispatch and handler execution exceed that window, DeferredHeaderFlushAsync invokes the response-starting callback with null and flushes the default HTTP 200 headers.

When the MissingRequiredClientCapability JSON-RPC error arrives later, _httpResponseStarted is already true, so the callback cannot update the HTTP status to 400. The JSON-RPC error is therefore returned over an already-committed HTTP 200 response.

This makes the conformance test dependent on scheduling and machine load rather than solely on server behavior.

Note

This issue was drafted with GitHub Copilot.