From 90c8d7ba850ca7ba14d453de80cb06940471d29c Mon Sep 17 00:00:00 2001
From: michelligabriele <gabriele.michelli@icloud.com>
Date: Thu, 4 Jun 2026 22:22:21 +0200
Subject: [PATCH] fix(mcp): re-raise cancellation on py3.10 where
 Task.cancelling() is absent

---
 .../mcp_server/mcp_server_manager.py          | 31 +++++++++++-----
 .../test_mcp_oauth_passthrough_tools.py       | 35 +++++++++++++++++++
 2 files changed, 57 insertions(+), 9 deletions(-)

diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
index 1db1c9d5dbe..6a1f7833b6c 100644
--- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
+++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
@@ -2540,16 +2540,29 @@ class MCPServerManager:
                     status_code=status_code, server_name=server_name
                 ) from e
 
-            # No recoverable HTTP status. Respect a genuine *outer* cancellation
-            # (shutdown / client disconnect) so cooperative cancellation still
-            # works; only treat an internally-originated cancel as a failure.
+            # No recoverable HTTP status. We must not launder this into an empty
+            # success, but we also must not swallow a genuine *outer*
+            # cancellation (shutdown / client disconnect) — re-raising those
+            # keeps cooperative cancellation working.
+            #
+            # On Python 3.11+ `asyncio.Task.cancelling()` lets us tell an
+            # internally-originated cancel-scope cancellation (the upstream
+            # failure we want to surface) from a real outer cancel (which we
+            # re-raise). On Python 3.10 that API does not exist, so we cannot
+            # classify the cancellation; there we conservatively re-raise to
+            # preserve cancellation semantics rather than risk converting a real
+            # shutdown cancel into a server error. (The trade-off: an
+            # upstream-induced bare cancellation on 3.10 propagates as an error
+            # instead of a clean per-server failure — still never a false
+            # success.)
             task = asyncio.current_task()
-            is_outer_cancel = bool(
-                task is not None
-                and getattr(task, "cancelling", None) is not None
-                and task.cancelling() > 0
-            )
-            if is_outer_cancel:
+            cancelling = getattr(task, "cancelling", None) if task is not None else None
+            if cancelling is None:
+                # Python 3.10 (or no running task): cannot classify — cooperate
+                # with the cancellation.
+                raise
+            if cancelling() > 0:
+                # A real outer cancellation was requested on this task.
                 raise
             verbose_logger.warning(
                 f"Tool listing from {server_name} was interrupted before "
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_oauth_passthrough_tools.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_oauth_passthrough_tools.py
index 497301d447e..f1c5a86d270 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_oauth_passthrough_tools.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_oauth_passthrough_tools.py
@@ -278,3 +278,38 @@ async def test_fetch_tools_surfaces_error_on_bare_cancellation():
         await manager._fetch_tools_with_timeout(mock_client, server.name, server=server)
     assert exc_info.value.status_code is None
     mock_client.list_tools.assert_awaited_with(raise_on_error=False)
+
+
+@pytest.mark.asyncio
+async def test_fetch_tools_reraises_genuine_outer_cancellation():
+    """A real outer cancellation (shutdown / client disconnect) must propagate,
+    not be converted into MCPUpstreamError — otherwise cooperative cancellation
+    breaks. On 3.11+ this is detected via Task.cancelling(); on 3.10 the
+    classifier is unavailable and we re-raise unconditionally, so this contract
+    holds on every supported runtime."""
+    manager = MCPServerManager()
+    server = MCPServer(
+        server_id="o1",
+        name="mock_databricks",
+        url="https://upstream/mcp",
+        transport=MCPTransport.http,
+        auth_type=MCPAuth.oauth2,
+    )
+
+    started = asyncio.Event()
+
+    async def _hang(*args, **kwargs):
+        started.set()
+        await asyncio.sleep(3600)
+
+    mock_client = MagicMock()
+    mock_client.list_tools = _hang
+
+    task = asyncio.create_task(
+        manager._fetch_tools_with_timeout(mock_client, server.name, server=server)
+    )
+    await started.wait()
+    task.cancel()
+
+    with pytest.raises(asyncio.CancelledError):
+        await task