fix(client): prevent tool hang by failing pending requests on transpo…

src/mcp/client/session.py

@@ -243,6 +243,15 @@ async def __call__(self, message: IncomingMessage) -> None: ...  # pragma: no br
 
 
 async def _default_message_handler(message: IncomingMessage) -> None:
+    # Transport-level faults (e.g. an SSE read timeout) reach the handler as the
+    # `Exception` arm of `IncomingMessage`. The default used to only checkpoint,
+    # so these vanished with no trace. Log them at ERROR: in-flight requests are
+    # already failed by the dispatcher, but a fault with no request in flight
+    # (an idle SSE stream dropping) would otherwise leave no signal at all.
+    # Server notifications remain a no-op - a handler is optional for those.
+    if isinstance(message, Exception):
+        logger.error("transport error surfaced to message handler: %r", message)
+        return
     await anyio.lowlevel.checkpoint()
 
 

src/mcp/shared/jsonrpc_dispatcher.py

@@ -536,6 +536,11 @@ async def _dispatch(
         are awaited; any other `await` would head-of-line block the read loop.
         """
         if isinstance(item, Exception):
+            # A transport fault fails every in-flight request now, rather than
+            # leaving each waiter to block until its own `opts["timeout"]`
+            # elapses (or forever, if it has none). Done before the observer
+            # runs so a slow observer can't delay freeing the waiters.
+            self._fail_pending(item)
             if self.on_stream_exception is None:
                 logger.debug("transport yielded exception: %r", item)
                 return
@@ -685,19 +690,36 @@ def _spawn(
         else:
             self._tg.start_soon(fn, *args)
 
-    def _fan_out_closed(self) -> None:
-        """Wake every pending `send_raw_request` waiter with `CONNECTION_CLOSED`.
+    def _fan_out_error(self, error: ErrorData) -> None:
+        """Wake every pending `send_raw_request` waiter with `error`.
 
-        Synchronous: callers may be inside a cancelled scope. Idempotent.
+        Synchronous: callers may be inside a cancelled scope. Idempotent - the
+        `_pending` table is cleared, so a second call is a no-op. A waiter that
+        already holds an outcome (buffer-of-1 full) keeps it: the `WouldBlock`
+        is swallowed rather than clobbering the real response with the signal.
         """
-        closed = ErrorData(code=CONNECTION_CLOSED, message="Connection closed")
         for pending in self._pending.values():
             try:
-                pending.send.send_nowait(closed)
+                pending.send.send_nowait(error)
             except (anyio.WouldBlock, anyio.BrokenResourceError, anyio.ClosedResourceError):
                 pass
         self._pending.clear()
 
+    def _fan_out_closed(self) -> None:
+        """Wake every pending `send_raw_request` waiter with `CONNECTION_CLOSED` (EOF/shutdown)."""
+        self._fan_out_error(ErrorData(code=CONNECTION_CLOSED, message="Connection closed"))
+
+    def _fail_pending(self, exc: Exception) -> None:
+        """Wake every pending waiter because the read stream yielded a transport fault.
+
+        Reported as `CONNECTION_CLOSED` for parity with the EOF path (`send_raw_request`
+        raises `MCPError` either way), but the message carries `exc` so callers can see
+        what actually broke - e.g. an `httpx.ReadTimeout` from a low `sse_read_timeout` -
+        instead of the request hanging until its own timeout. The raw `exc` still reaches
+        any `on_stream_exception` observer untouched.
+        """
+        self._fan_out_error(ErrorData(code=CONNECTION_CLOSED, message=f"Transport error: {exc!r}"))
+
     async def _handle_request(
         self,
         req: JSONRPCRequest,

tests/shared/test_jsonrpc_dispatcher.py

@@ -2514,3 +2514,90 @@ async def call() -> None:
         for stream in (c2s_send, c2s_recv, s2c_send, s2c_recv):
             stream.close()
     assert result_box == [{"ok": True}]
+
+
+@pytest.mark.anyio
+async def test_transport_exception_fails_pending_request_without_hanging():
+    """A read-stream fault wakes an in-flight `send_raw_request` with `CONNECTION_CLOSED`.
+
+    Regression for the streamable-http hang: before this, a transport error (e.g. an SSE
+    read timeout) reached only the observer, so a request already waiting on its response
+    sat until its own `opts["timeout"]` elapsed. Now the dispatcher fails the waiter at once,
+    and the error message carries the transport exception so the caller can see the cause.
+    """
+    c2s_send, c2s_recv = anyio.create_memory_object_stream[SessionMessage | Exception](4)
+    s2c_send, s2c_recv = anyio.create_memory_object_stream[SessionMessage | Exception](4)
+
+    client: JSONRPCDispatcher[TransportContext] = JSONRPCDispatcher(c2s_recv, s2c_send)
+    on_request, on_notify = echo_handlers(Recorder())
+    outcome: dict[str, BaseException] = {}
+    boom = TimeoutError("sse read timed out")
+    try:
+        async with anyio.create_task_group() as tg:
+            await tg.start(client.run, on_request, on_notify)
+
+            async def call() -> None:
+                # No `timeout` opt: without the fix this would block forever.
+                try:
+                    await client.send_raw_request("tools/call", {"name": "slow"})
+                except BaseException as exc:  # noqa: BLE001 - capture whatever the waiter raises
+                    outcome["exc"] = exc
+
+            tg.start_soon(call)
+            # Let the request register in `_pending` and park on its response.
+            with anyio.fail_after(5):
+                sent = await s2c_recv.receive()
+            assert isinstance(sent, SessionMessage)
+            assert isinstance(sent.message, JSONRPCRequest)
+
+            # The transport now yields an exception instead of a response.
+            await c2s_send.send(boom)
+            with anyio.fail_after(5):
+                while "exc" not in outcome:
+                    await anyio.sleep(0)
+            tg.cancel_scope.cancel()
+    finally:
+        for s in (c2s_send, c2s_recv, s2c_send, s2c_recv):
+            s.close()
+
+    raised = outcome["exc"]
+    assert isinstance(raised, MCPError)
+    assert raised.error.code == CONNECTION_CLOSED
+    # The original transport exception is preserved in the message for debugging.
+    assert "sse read timed out" in raised.error.message
+
+
+def test_fail_pending_reports_transport_exception_and_clears_pending():
+    """White-box: `_fail_pending` wakes waiters with the exception detail, then empties `_pending`."""
+    c2s_send, c2s_recv = anyio.create_memory_object_stream[SessionMessage | Exception](1)
+    s2c_send, s2c_recv = anyio.create_memory_object_stream[SessionMessage | Exception](1)
+    d: JSONRPCDispatcher[TransportContext] = JSONRPCDispatcher(s2c_recv, c2s_send)
+    send, recv = anyio.create_memory_object_stream[dict[str, Any] | ErrorData](1)
+    d._pending[1] = _Pending(send=send, receive=recv)  # pyright: ignore[reportPrivateUsage]
+
+    d._fail_pending(TimeoutError("sse read timed out"))  # pyright: ignore[reportPrivateUsage]
+
+    signalled = recv.receive_nowait()
+    assert isinstance(signalled, ErrorData)
+    assert signalled.code == CONNECTION_CLOSED
+    assert "sse read timed out" in signalled.message
+    assert d._pending == {}  # pyright: ignore[reportPrivateUsage]
+    for s in (c2s_send, c2s_recv, s2c_send, s2c_recv, send, recv):
+        s.close()
+
+
+def test_fail_pending_keeps_existing_outcome_when_waiter_already_resolved():
+    """White-box: a waiter that already holds a real result is not clobbered by the fault signal."""
+    c2s_send, c2s_recv = anyio.create_memory_object_stream[SessionMessage | Exception](1)
+    s2c_send, s2c_recv = anyio.create_memory_object_stream[SessionMessage | Exception](1)
+    d: JSONRPCDispatcher[TransportContext] = JSONRPCDispatcher(s2c_recv, c2s_send)
+    send, recv = anyio.create_memory_object_stream[dict[str, Any] | ErrorData](1)
+    d._pending[1] = _Pending(send=send, receive=recv)  # pyright: ignore[reportPrivateUsage]
+    send.send_nowait({"real": "result"})
+
+    d._fail_pending(TimeoutError("sse read timed out"))  # pyright: ignore[reportPrivateUsage]
+
+    assert recv.receive_nowait() == {"real": "result"}
+    assert d._pending == {}  # pyright: ignore[reportPrivateUsage]
+    for s in (c2s_send, c2s_recv, s2c_send, s2c_recv, send, recv):
+        s.close()