Stabilize flaky test: McpCompletionTests.testCompletionErrorOnMissingContext

[ashakirin]

McpCompletionTests.testCompletionErrorOnMissingContext is sometimes failing and recovers after simple build restart.
**Error:** Session not found: <sessionId>

Likely causes:
The server removes the session when SSE connection closes, but the client still has pending requests in flight. This is a race condition between:

• SSE connection closing → server removes session
• Client sending next request → "Session not found"

2026-02-09 23:31:09 [HttpClient-1-Worker-0] WARN  i.m.c.t.HttpClientSseClientTransport - SSE stream observed an error
java.io.IOException: chunked transfer encoding, state: READING_LENGTH
	at java.net.http/jdk.internal.net.http.common.Utils.wrapWithExtraDetail(Utils.java:375)
	at java.net.http/jdk.internal.net.http.Http1Response$BodyReader.onReadError(Http1Response.java:660)
	at java.net.http/jdk.internal.net.http.Http1AsyncReceiver.checkForErrors(Http1AsyncReceiver.java:302)
	at java.net.http/jdk.internal.net.http.Http1AsyncReceiver.flush(Http1AsyncReceiver.java:268)
	at java.net.http/jdk.internal.net.http.common.SequentialScheduler$LockingRestartableTask.run(SequentialScheduler.java:205)
	at java.net.http/jdk.internal.net.http.common.SequentialScheduler$CompleteRestartableTask.run(SequentialScheduler.java:149)
	at java.net.http/jdk.internal.net.http.common.SequentialScheduler$SchedulableTask.run(SequentialScheduler.java:230)
	at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1136)
	at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:635)
	at java.base/java.lang.Thread.run(Thread.java:840)
Caused by: java.io.EOFException: EOF reached while reading
	at java.net.http/jdk.internal.net.http.Http1AsyncReceiver$Http1TubeSubscriber.onComplete(Http1AsyncReceiver.java:597)
	at java.net.http/jdk.internal.net.http.SocketTube$InternalReadPublisher$ReadSubscription.signalCompletion(SocketTube.java:648)
	at java.net.http/jdk.internal.net.http.SocketTube$InternalReadPublisher$InternalReadSubscription.read(SocketTube.java:853)
	at java.net.http/jdk.internal.net.http.SocketTube$SocketFlowTask.run(SocketTube.java:181)
	at java.net.http/jdk.internal.net.http.common.SequentialScheduler$SchedulableTask.run(SequentialScheduler.java:230)
	at java.net.http/jdk.internal.net.http.common.SequentialScheduler.runOrSchedule(SequentialScheduler.java:303)
	at java.net.http/jdk.internal.net.http.common.SequentialScheduler.runOrSchedule(SequentialScheduler.java:256)
	at java.net.http/jdk.internal.net.http.SocketTube$InternalReadPublisher$InternalReadSubscription.signalReadable(SocketTube.java:782)
	at java.net.http/jdk.internal.net.http.SocketTube$InternalReadPublisher$ReadEvent.signalEvent(SocketTube.java:965)
	at java.net.http/jdk.internal.net.http.SocketTube$SocketFlowEvent.handle(SocketTube.java:253)
	at java.net.http/jdk.internal.net.http.HttpClientImpl$SelectorManager.handleEvent(HttpClientImpl.java:1402)
	at java.net.http/jdk.internal.net.http.HttpClientImpl$SelectorManager.lambda$run$3(HttpClientImpl.java:1347)
	at java.base/java.util.ArrayList.forEach(ArrayList.java:1511)
	at java.net.http/jdk.internal.net.http.HttpClientImpl$SelectorManager.run(HttpClientImpl.java:1347)
2026-02-09 23:31:09 [ForkJoinPool.commonPool-worker-1] ERROR i.m.c.t.HttpClientSseClientTransport - Error sending message: Sending message failed with a non-OK HTTP code: 404 - {"cause":null,"jsonRpcError":null,"localizedMessage":"Session not found: 7a9438d6-5965-456a-99af-26523f108d68","message":"Session not found: 7a9438d6-5965-456a-99af-26523f108d68","stackTrace":

https://github.com/modelcontextprotocol/java-sdk/actions/runs/21845041347/job/63038831474#step:4:15277

[rsmihir389]

Hey — we have been looking at exactly this kind of failure pattern a lot lately. Race conditions between SSE teardown and in-flight client requests like this usually cost teams more than just the failed test — every rerun burns CI minutes while the underlying race stays unfixed.

A few things that tend to help with this specific pattern:

1. Add a small drain period before the server removes the session — give in-flight requests a window (50–100ms) to complete or time out before the session map is cleaned up
2. Retry with backoff on the client side for 404s on session endpoints specifically, since these are expected during teardown races
3. Use a latch or countdown in the test itself to ensure the client has finished sending before the SSE connection is allowed to close

We built a tool that detects flaky tests like this and shows the dollar cost in CI time: kleore.com. Here is what a report looks like on a real repo: sample report.

60-second install as a GitHub App, free tier shows the cost. Might be useful for catching other intermittent failures in this repo before they pile up.