`HttpServletStreamableServerTransportProvider` removes session on non-fatal failure

Bug description
We use the MCP Java SDK as via our dependency on Spring AI. Our CI/CD test coverage for our MCP implementation fails regularly with a race condition when dealing with intentional error conditions. We could not reproduce it locally on dev machines, but via logging eventually found a workaround to what appears to be a problem in HttpServletStreamableTransportProvider.

Environment
Versions 1.1.1, 1.1.2, and 2.0.0-M2 of the Java SDK. Prerelease versions of Spring AI 2.0, including 2.0.0-M5 and 2.0.0-M4. Java 25.

Steps to reproduce

See attached for a standalone repro. The error manifested for us as a race condition but always of this form:

  java.lang.RuntimeException: MCP session with server terminated
  at io.modelcontextprotocol.spec.McpClientSession.lambda$dismissPendingResponses$1(McpClientSession.java:128)
  at java.base/java.util.concurrent.ConcurrentHashMap.forEach(ConcurrentHashMap.java:1619)
  at io.modelcontextprotocol.spec.McpClientSession.dismissPendingResponses(McpClientSession.java:126)
  at io.modelcontextprotocol.spec.McpClientSession.close(McpClientSession.java:304)
  at io.modelcontextprotocol.client.LifecycleInitializer$DefaultInitialization.close(LifecycleInitializer.java:225)
  at io.modelcontextprotocol.client.LifecycleInitializer.handleException(LifecycleInitializer.java:257)
  at io.modelcontextprotocol.client.transport.HttpClientStreamableHttpTransport.handleException(HttpClientStreamableHttpTransport.java:231)
  at io.modelcontextprotocol.client.transport.HttpClientStreamableHttpTransport.lambda$sendMessage$33(HttpClientStreamableHttpTransport.java:641)
  at reactor.core.publisher.FluxOnErrorReturn$ReturnSubscriber.onError(FluxOnErrorReturn.java:172)
  at reactor.core.publisher.FluxOnErrorResume$ResumeSubscriber.onError(FluxOnErrorResume.java:107)
  at reactor.core.publisher.Operators.error(Operators.java:198)
  at reactor.core.publisher.FluxError.subscribe(FluxError.java:44)
  at reactor.core.publisher.Flux.subscribe(Flux.java:8888)
  at reactor.core.publisher.FluxOnErrorResume$ResumeSubscriber.onError(FluxOnErrorResume.java:104)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.checkTerminated(FluxFlatMap.java:847)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.drainLoop(FluxFlatMap.java:613)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.drain(FluxFlatMap.java:593)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.onError(FluxFlatMap.java:456)
  at reactor.core.publisher.SerializedSubscriber.onError(SerializedSubscriber.java:125)
  at reactor.core.publisher.FluxRetryWhen$RetryWhenMainSubscriber.whenError(FluxRetryWhen.java:230)
  at reactor.core.publisher.FluxRetryWhen$RetryWhenOtherSubscriber.onError(FluxRetryWhen.java:282)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.checkTerminated(FluxFlatMap.java:847)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.drainLoop(FluxFlatMap.java:613)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.drain(FluxFlatMap.java:593)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.onError(FluxFlatMap.java:456)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.onNext(FluxFlatMap.java:419)
  at reactor.core.publisher.SinkManyEmitterProcessor.drain(SinkManyEmitterProcessor.java:480)
  at reactor.core.publisher.SinkManyEmitterProcessor.tryEmitNext(SinkManyEmitterProcessor.java:278)
  at reactor.core.publisher.SinkManySerialized.tryEmitNext(SinkManySerialized.java:100)
  at reactor.core.publisher.InternalManySink.emitNext(InternalManySink.java:27)
  at reactor.core.publisher.FluxRetryWhen$RetryWhenMainSubscriber.onError(FluxRetryWhen.java:195)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.checkTerminated(FluxFlatMap.java:847)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.drainLoop(FluxFlatMap.java:613)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.drain(FluxFlatMap.java:593)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.onError(FluxFlatMap.java:456)
  at reactor.core.publisher.FluxFlatMap$FlatMapMain.onNext(FluxFlatMap.java:419)
  at reactor.core.publisher.MonoFlatMapMany$FlatMapManyInner.onNext(MonoFlatMapMany.java:251)
  at reactor.core.publisher.FluxCreate$BufferAsyncSink.drain(FluxCreate.java:887)
  at reactor.core.publisher.FluxCreate$BufferAsyncSink.next(FluxCreate.java:812)
  at reactor.core.publisher.FluxCreate$SerializedFluxSink.next(FluxCreate.java:164)
  at io.modelcontextprotocol.client.transport.ResponseSubscribers$AggregateSubscriber.hookOnComplete(ResponseSubscribers.java:263)
  at reactor.core.publisher.BaseSubscriber.onComplete(BaseSubscriber.java:200)
  at org.reactivestreams.FlowAdapters$FlowToReactiveSubscriber.onComplete(FlowAdapters.java:221)
  at java.net.http/jdk.internal.net.http.LineSubscriberAdapter$LineSubscription.loop(LineSubscriberAdapter.java:430)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$LockingRestartableTask.run(SequentialScheduler.java:182)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$CompleteRestartableTask.run(SequentialScheduler.java:149)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$SchedulableTask.run(SequentialScheduler.java:207)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler.runOrSchedule(SequentialScheduler.java:280)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler.runOrSchedule(SequentialScheduler.java:233)
  at java.net.http/jdk.internal.net.http.LineSubscriberAdapter$LineSubscription.signalComplete(LineSubscriberAdapter.java:193)
  at java.net.http/jdk.internal.net.http.LineSubscriberAdapter.onComplete(LineSubscriberAdapter.java:114)
  at java.net.http/jdk.internal.net.http.common.HttpBodySubscriberWrapper.complete(HttpBodySubscriberWrapper.java:293)
  at java.net.http/jdk.internal.net.http.common.HttpBodySubscriberWrapper.onComplete(HttpBodySubscriberWrapper.java:401)
  at java.net.http/jdk.internal.net.http.ResponseContent$ChunkedBodyParser.accept(ResponseContent.java:220)
  at java.net.http/jdk.internal.net.http.ResponseContent$ChunkedBodyParser.accept(ResponseContent.java:131)
  at java.net.http/jdk.internal.net.http.Http1Response$BodyReader.handle(Http1Response.java:708)
  at java.net.http/jdk.internal.net.http.Http1Response$BodyReader.handle(Http1Response.java:636)
  at java.net.http/jdk.internal.net.http.Http1Response$Receiver.accept(Http1Response.java:528)
  at java.net.http/jdk.internal.net.http.Http1Response$BodyReader.tryAsyncReceive(Http1Response.java:666)
  at java.net.http/jdk.internal.net.http.Http1AsyncReceiver.flush(Http1AsyncReceiver.java:233)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$LockingRestartableTask.run(SequentialScheduler.java:182)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$CompleteRestartableTask.run(SequentialScheduler.java:149)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$TryEndDeferredCompleter.complete(SequentialScheduler.java:324)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$CompleteRestartableTask.run(SequentialScheduler.java:151)
  at java.net.http/jdk.internal.net.http.common.SequentialScheduler$SchedulableTask.run(SequentialScheduler.java:207)
  at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1090)
  at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:614)
  at java.base/java.lang.Thread.run(Thread.java:1474)
  Suppressed: java.lang.Exception: #block terminated with an error
    at reactor.core.publisher.BlockingSingleSubscriber.blockingGet(BlockingSingleSubscriber.java:104)
    at reactor.core.publisher.Mono.block(Mono.java:1773)
    at io.modelcontextprotocol.client.McpSyncClient.callTool(McpSyncClient.java:236)
    at org.labkey.professional.McpServerTest.callTool(McpServerTest.java:358)
    at org.labkey.professional.McpServerTest.callToolExpectingFailure(McpServerTest.java:342)
    at org.labkey.professional.McpServerTest.testBadTableParameters(McpServerTest.java:566)
    at org.labkey.professional.McpServerTest.getSourceForSavedQuery(McpServerTest.java:522)
    at org.labkey.professional.McpServerTest.invokeTools(McpServerTest.java:327)
    at org.labkey.professional.McpServerTest.testAdminAccess(McpServerTest.java:254)
    at java.base/jdk.internal.reflect.DirectMethodHandleAccessor.invoke(DirectMethodHandleAccessor.java:104)
    at java.base/java.lang.reflect.Method.invoke(Method.java:565)
    at org.junit.runners.model.FrameworkMethod$1.runReflectiveCall(FrameworkMethod.java:59)
    at org.junit.internal.runners.model.ReflectiveCallable.run(ReflectiveCallable.java:12)

Expected behavior
The session should continue to process requests and respond to the client without fatal errors.

Minimal Complete Reproducible example
See attached.

mcp-sdk-bug-repro.zip

cd mcp-sdk-bug-repro
mvn -q test

The test in src/test/java/io/modelcontextprotocol/bug/SessionRemovalBugReproTest.java:

1. Builds a real HttpServletStreamableServerTransportProvider and a minimal McpSyncServer with a single tool.
2. Issues an initialize POST and captures the server-assigned mcp-session-id.
3. Issues a tools/call POST whose response writer is wired to a Writer that accepts the writes (so the SDK successfully serializes the SSE event into the response body) but throws IOException on flush() — exactly what an OutputStream backed by a closed socket does. This sets PrintWriter.trouble = true so writer.checkError() returns true.
4. Asserts that:
   • The full SSE event payload was written to the response (data was delivered).
   • The session was removed from the SDK's sessions map (verified via reflection).
5. Issues a second tools/call POST with the same session id and asserts that the SDK responds 404 Session not found — proving the session is unrecoverable from the client's perspective even though the prior call's response was actually delivered.

The test deterministically reproduces what is otherwise a flaky race between a fast client closing its socket and the server's post-write checkError().

Workaround
Wrap the response in an HttpServletResponseWrapper that returns a PrintWriter whose checkError() always returns false. This suppresses the false positive at the cost of closing the (already broken) post-write disconnect detection. Apply only on the POST path — keep checkError honest for the long-lived GET listening stream.

    private static class _LoggingResponseWrapper extends HttpServletResponseWrapper
    {
        private final String requestUri;
        private _LoggingPrintWriter loggingWriter;

        _LoggingResponseWrapper(HttpServletResponse response, String requestUri)
        {
            super(response);
            this.requestUri = requestUri;
        }

        @Override
        public PrintWriter getWriter() throws IOException
        {
            if (loggingWriter == null)
                loggingWriter = new _LoggingPrintWriter(super.getWriter(), requestUri);
            return loggingWriter;
        }
    }

    private static class _LoggingPrintWriter extends PrintWriter
    {
        private final String requestUri;

        _LoggingPrintWriter(PrintWriter delegate, String requestUri)
        {
            super(delegate);
            this.requestUri = requestUri;
        }

        @Override
        public boolean checkError()
        {
            // Workaround for MCP SDK 1.1.2 bug in HttpServletStreamableServerTransportProvider.sendEvent
            // (https://github.com/modelcontextprotocol/java-sdk). After writing/flushing an SSE event,
            // sendEvent calls writer.checkError() and throws "Client disconnected" if it returns true.
            // The catch block in sendMessage then removes the MCP session from the sessions map - even
            // though the SSE event data was already delivered to the client. For streamable HTTP, each
            // POST is a single-event exchange, so a client closing its socket after receiving the
            // response is normal HTTP/1.1 behavior, not an MCP session termination. The race between
            // the client's socket close and the server's checkError() call is exactly what makes this
            // test flaky on TeamCity. Returning false suppresses the false positive; if the client
            // truly lost data, it will retry on a fresh connection.
            boolean actual = super.checkError();
            if (actual)
            {
                LOG.info("MCP checkError suppressed (returning false): uri={}", requestUri);
            }
            return false;
        }
    }

Bắt đầu với repro đính kèm tại src/test/java/io/modelcontextprotocol/bug/SessionRemovalBugReproTest.java và chạy mvn -q test. Sau đó kiểm tra HttpServletStreamableServerTransportProvider.sendEvent và sendMessage, tập trung vào đường dẫn post-write checkError. Công việc được hoàn tất khi lần flush thất bại được tái hiện không xóa một session sau khi sự kiện SSE của nó đã được gửi, và tools/call thứ hai vẫn có thể sử dụng.