fix: cancel in-flight handlers when transport closes in server.run() (#2306)

src/mcp/server/lowlevel/server.py

@@ -387,16 +387,23 @@ class Server(Generic[LifespanResultT]):
                 await stack.enter_async_context(task_support.run())
 
             async with anyio.create_task_group() as tg:
-                async for message in session.incoming_messages:
+                try:
-                    logger.debug("Received message: %s", message)
+                    async for message in session.incoming_messages:
+                        logger.debug("Received message: %s", message)
 
-                    tg.start_soon(
+                        tg.start_soon(
-                        self._handle_message,
+                            self._handle_message,
-                        message,
+                            message,
-                        session,
+                            session,
-                        lifespan_context,
+                            lifespan_context,
-                        raise_exceptions,
+                            raise_exceptions,
-                    )
+                        )
+                finally:
+                    # Transport closed: cancel in-flight handlers. Without this the
+                    # TG join waits for them, and when they eventually try to
+                    # respond they hit a closed write stream (the session's
+                    # _receive_loop closed it when the read stream ended).
+                    tg.cancel_scope.cancel()
 
     async def _handle_message(
         self,
@@ -470,16 +477,32 @@ class Server(Generic[LifespanResultT]):
             except MCPError as err:
                 response = err.error
             except anyio.get_cancelled_exc_class():
-                logger.info("Request %s cancelled - duplicate response suppressed", message.request_id)
+                if message.cancelled:
-                return
+                    # Client sent CancelledNotification; responder.cancel() already
+                    # sent an error response, so skip the duplicate.
+                    logger.info("Request %s cancelled - duplicate response suppressed", message.request_id)
+                    return
+                # Transport-close cancellation from the TG in run(); re-raise so the
+                # TG swallows its own cancellation.
+                raise
             except Exception as err:
                 if raise_exceptions:  # pragma: no cover
                     raise err
                 response = types.ErrorData(code=0, message=str(err))
-
-            await message.respond(response)
         else:  # pragma: no cover
-            await message.respond(types.ErrorData(code=types.METHOD_NOT_FOUND, message="Method not found"))
+            response = types.ErrorData(code=types.METHOD_NOT_FOUND, message="Method not found")
+
+        try:
+            await message.respond(response)
+        except (anyio.BrokenResourceError, anyio.ClosedResourceError):
+            # Transport closed between handler unblocking and respond. Happens
+            # when _receive_loop's finally wakes a handler blocked on
+            # send_request: the handler runs to respond() before run()'s TG
+            # cancel fires, but after the write stream closed. Closed if our
+            # end closed (_receive_loop's async-with exit); Broken if the peer
+            # end closed first (streamable_http terminate()).
+            logger.debug("Response for %s dropped - transport closed", message.request_id)
+            return
 
         logger.debug("Response sent")
 

src/mcp/shared/session.py

@@ -105,7 +105,7 @@ class RequestResponder(Generic[ReceiveRequestT, SendResultT]):
     ) -> None:
         """Exit the context manager, performing cleanup and notifying completion."""
         try:
-            if self._completed:  # pragma: no branch
+            if self._completed:
                 self._on_complete(self)
         finally:
             self._entered = False
@@ -418,7 +418,9 @@ class BaseSession(
             finally:
                 # after the read stream is closed, we need to send errors
                 # to any pending requests
-                for id, stream in self._response_streams.items():
+                # Snapshot: stream.send() wakes the waiter, whose finally pops
+                # from _response_streams before the next __next__() call.
+                for id, stream in list(self._response_streams.items()):
                     error = ErrorData(code=CONNECTION_CLOSED, message="Connection closed")
                     try:
                         await stream.send(JSONRPCError(jsonrpc="2.0", id=id, error=error))

tests/server/test_cancel_handling.py

@@ -6,12 +6,19 @@ import pytest
 from mcp import Client
 from mcp.server import Server, ServerRequestContext
 from mcp.shared.exceptions import MCPError
+from mcp.shared.message import SessionMessage
 from mcp.types import (
+    LATEST_PROTOCOL_VERSION,
     CallToolRequest,
     CallToolRequestParams,
     CallToolResult,
     CancelledNotification,
     CancelledNotificationParams,
+    ClientCapabilities,
+    Implementation,
+    InitializeRequestParams,
+    JSONRPCNotification,
+    JSONRPCRequest,
     ListToolsResult,
     PaginatedRequestParams,
     TextContent,
@@ -90,3 +97,154 @@ async def test_server_remains_functional_after_cancel():
         assert isinstance(content, TextContent)
         assert content.text == "Call number: 2"
         assert call_count == 2
+
+
+@pytest.mark.anyio
+async def test_server_cancels_in_flight_handlers_on_transport_close():
+    """When the transport closes mid-request, server.run() must cancel in-flight
+    handlers rather than join on them.
+
+    Without the cancel, the task group waits for the handler, which then tries
+    to respond through a write stream that _receive_loop already closed,
+    raising ClosedResourceError and crashing server.run() with exit code 1.
+
+    This drives server.run() with raw memory streams because InMemoryTransport
+    wraps it in its own finally-cancel (_memory.py) which masks the bug.
+    """
+    handler_started = anyio.Event()
+    handler_cancelled = anyio.Event()
+    server_run_returned = anyio.Event()
+
+    async def handle_call_tool(ctx: ServerRequestContext, params: CallToolRequestParams) -> CallToolResult:
+        handler_started.set()
+        try:
+            await anyio.sleep_forever()
+        finally:
+            handler_cancelled.set()
+        # unreachable: sleep_forever only exits via cancellation
+        raise AssertionError  # pragma: no cover
+
+    server = Server("test", on_call_tool=handle_call_tool)
+
+    to_server, server_read = anyio.create_memory_object_stream[SessionMessage | Exception](10)
+    server_write, from_server = anyio.create_memory_object_stream[SessionMessage](10)
+
+    async def run_server():
+        await server.run(server_read, server_write, server.create_initialization_options())
+        server_run_returned.set()
+
+    init_req = JSONRPCRequest(
+        jsonrpc="2.0",
+        id=1,
+        method="initialize",
+        params=InitializeRequestParams(
+            protocol_version=LATEST_PROTOCOL_VERSION,
+            capabilities=ClientCapabilities(),
+            client_info=Implementation(name="test", version="1.0"),
+        ).model_dump(by_alias=True, mode="json", exclude_none=True),
+    )
+    initialized = JSONRPCNotification(jsonrpc="2.0", method="notifications/initialized")
+    call_req = JSONRPCRequest(
+        jsonrpc="2.0",
+        id=2,
+        method="tools/call",
+        params=CallToolRequestParams(name="slow", arguments={}).model_dump(by_alias=True, mode="json"),
+    )
+
+    with anyio.fail_after(5):
+        async with anyio.create_task_group() as tg, to_server, server_read, server_write, from_server:
+            tg.start_soon(run_server)
+
+            await to_server.send(SessionMessage(init_req))
+            await from_server.receive()  # init response
+            await to_server.send(SessionMessage(initialized))
+            await to_server.send(SessionMessage(call_req))
+
+            await handler_started.wait()
+
+            # Close the server's input stream — this is what stdin EOF does.
+            # server.run()'s incoming_messages loop ends, finally-cancel fires,
+            # handler gets CancelledError, server.run() returns.
+            await to_server.aclose()
+
+            await server_run_returned.wait()
+
+    assert handler_cancelled.is_set()
+
+
+@pytest.mark.anyio
+async def test_server_handles_transport_close_with_pending_server_to_client_requests():
+    """When the transport closes while handlers are blocked on server→client
+    requests (sampling, roots, elicitation), server.run() must still exit cleanly.
+
+    Two bugs covered:
+      1. _receive_loop's finally iterates _response_streams with await checkpoints
+         inside; the woken handler's send_request finally pops from that dict
+         before the next __next__() — RuntimeError: dictionary changed size.
+      2. The woken handler's MCPError is caught in _handle_request, which falls
+         through to respond() against a write stream _receive_loop already closed.
+    """
+    handlers_started = 0
+    both_started = anyio.Event()
+    server_run_returned = anyio.Event()
+
+    async def handle_call_tool(ctx: ServerRequestContext, params: CallToolRequestParams) -> CallToolResult:
+        nonlocal handlers_started
+        handlers_started += 1
+        if handlers_started == 2:
+            both_started.set()
+        # Blocks on send_request waiting for a client response that never comes.
+        # _receive_loop's finally will wake this with CONNECTION_CLOSED.
+        await ctx.session.list_roots()
+        raise AssertionError  # pragma: no cover
+
+    server = Server("test", on_call_tool=handle_call_tool)
+
+    to_server, server_read = anyio.create_memory_object_stream[SessionMessage | Exception](10)
+    server_write, from_server = anyio.create_memory_object_stream[SessionMessage](10)
+
+    async def run_server():
+        await server.run(server_read, server_write, server.create_initialization_options())
+        server_run_returned.set()
+
+    init_req = JSONRPCRequest(
+        jsonrpc="2.0",
+        id=1,
+        method="initialize",
+        params=InitializeRequestParams(
+            protocol_version=LATEST_PROTOCOL_VERSION,
+            capabilities=ClientCapabilities(),
+            client_info=Implementation(name="test", version="1.0"),
+        ).model_dump(by_alias=True, mode="json", exclude_none=True),
+    )
+    initialized = JSONRPCNotification(jsonrpc="2.0", method="notifications/initialized")
+
+    with anyio.fail_after(5):
+        async with anyio.create_task_group() as tg, to_server, server_read, server_write, from_server:
+            tg.start_soon(run_server)
+
+            await to_server.send(SessionMessage(init_req))
+            await from_server.receive()  # init response
+            await to_server.send(SessionMessage(initialized))
+
+            # Two tool calls → two handlers → two _response_streams entries.
+            for rid in (2, 3):
+                call_req = JSONRPCRequest(
+                    jsonrpc="2.0",
+                    id=rid,
+                    method="tools/call",
+                    params=CallToolRequestParams(name="t", arguments={}).model_dump(by_alias=True, mode="json"),
+                )
+                await to_server.send(SessionMessage(call_req))
+
+            await both_started.wait()
+            # Drain the two roots/list requests so send_request's _write_stream.send()
+            # completes and both handlers are parked at response_stream_reader.receive().
+            await from_server.receive()
+            await from_server.receive()
+
+            await to_server.aclose()
+
+            # Without the fixes: RuntimeError (dict mutation) or ClosedResourceError
+            # (respond after write-stream close) escapes run_server and this hangs.
+            await server_run_returned.wait()