fix(stdio): bound EOF drain wait

Author: Epochex

src/mcp/server/lowlevel/server.py

@@ -69,6 +69,8 @@ async def main():
 
 logger = logging.getLogger(__name__)
 
+DEFAULT_READ_EOF_DRAIN_TIMEOUT_SECONDS = 1.0
+
 LifespanResultT = TypeVar("LifespanResultT", default=Any)
 
 _ParamsT = TypeVar("_ParamsT", bound=BaseModel, default=BaseModel)
@@ -410,6 +412,9 @@ async def run(
         # to drain their responses via the still-open write stream (e.g. stdio
         # with bash-redirected stdin).
         drain_on_read_close: bool = False,
+        # Maximum time to wait for in-flight handlers to drain after read EOF.
+        # None means wait indefinitely.
+        read_eof_drain_timeout_seconds: float | None = DEFAULT_READ_EOF_DRAIN_TIMEOUT_SECONDS,
     ) -> None:
         async with self.lifespan(self) as lifespan_context:
             dispatcher: JSONRPCDispatcher[TransportContext] = JSONRPCDispatcher(
@@ -421,6 +426,7 @@ async def run(
                 # the initialized state instead of failing the init-gate.
                 inline_methods=frozenset({"initialize"}),
                 close_write_stream_on_read_close=not drain_on_read_close,
+                read_eof_drain_timeout_seconds=read_eof_drain_timeout_seconds,
             )
             runner = ServerRunner(
                 server=self,

src/mcp/shared/jsonrpc_dispatcher.py

@@ -228,6 +228,7 @@ def __init__(
         raise_handler_exceptions: bool = False,
         inline_methods: frozenset[str] = frozenset(),
         close_write_stream_on_read_close: bool = True,
+        read_eof_drain_timeout_seconds: float | None = None,
     ) -> None: ...
     @overload
     def __init__(
@@ -240,6 +241,7 @@ def __init__(
         raise_handler_exceptions: bool = False,
         inline_methods: frozenset[str] = frozenset(),
         close_write_stream_on_read_close: bool = True,
+        read_eof_drain_timeout_seconds: float | None = None,
     ) -> None: ...
     def __init__(
         self,
@@ -251,6 +253,7 @@ def __init__(
         raise_handler_exceptions: bool = False,
         inline_methods: frozenset[str] = frozenset(),
         close_write_stream_on_read_close: bool = True,
+        read_eof_drain_timeout_seconds: float | None = None,
     ) -> None:
         self._read_stream = read_stream
         self._write_stream = write_stream
@@ -264,6 +267,7 @@ def __init__(
         self._peer_cancel_mode: PeerCancelMode = peer_cancel_mode
         self._raise_handler_exceptions = raise_handler_exceptions
         self._close_write_stream_on_read_close = close_write_stream_on_read_close
+        self._read_eof_drain_timeout_seconds = read_eof_drain_timeout_seconds
         # Request methods handled inline in the read loop (awaited before the
         # next message is dequeued) instead of spawned concurrently. Use for
         # methods whose side effects must be observable to the next message,
@@ -436,17 +440,22 @@ async def run(
                     self._fan_out_closed()
                     normal_eof = True
                 finally:
-                    if not normal_eof:
+                    if not normal_eof or self._close_write_stream_on_read_close:
                         # Transport closed abnormally: cancel in-flight handlers.
                         # On normal EOF, let already-received handlers drain
                         # their responses before the task group exits.
                         tg.cancel_scope.cancel()
+                    elif self._read_eof_drain_timeout_seconds is not None:
+                        tg.cancel_scope.deadline = anyio.current_time() + self._read_eof_drain_timeout_seconds
         finally:
             # Covers the cancel/crash paths where the inline fan-out above is
             # never reached. Idempotent.
             self._running = False
             self._tg = None
             self._fan_out_closed()
+            if not self._close_write_stream_on_read_close:
+                with anyio.CancelScope(shield=True):
+                    await self._write_stream.aclose()
 
     async def _dispatch(
         self,

tests/server/test_cancel_handling.py

@@ -120,7 +120,13 @@ async def handle_call_tool(ctx: ServerRequestContext, params: CallToolRequestPar
     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(), drain_on_read_close=True)
+        await server.run(
+            server_read,
+            server_write,
+            server.create_initialization_options(),
+            drain_on_read_close=True,
+            read_eof_drain_timeout_seconds=None,
+        )
         server_run_returned.set()
 
     init_req = JSONRPCRequest(
@@ -166,6 +172,70 @@ async def run_server():
             await server_run_returned.wait()
 
 
+@pytest.mark.anyio
+async def test_server_bounds_drain_on_read_eof_when_handler_never_finishes():
+    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()
+        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(),
+            drain_on_read_close=True,
+            read_eof_drain_timeout_seconds=0.05,
+        )
+        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(2):
+        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()
+            await to_server.aclose()
+
+            await server_run_returned.wait()
+
+    assert handler_cancelled.is_set()
+
+
 @pytest.mark.anyio
 async def test_server_reraises_handler_cancellation_when_server_is_cancelled():
     """If the server task is cancelled (e.g. KeyboardInterrupt), in-flight