fix(tools): Prevent session drop on MCP tool error

Enable MCP graceful error handling by default and retrieve background session context task exceptions. This prevents unhandled AnyIO TaskGroup transport failures from bubbling through the event loop and abruptly terminating live streaming sessions with 1006 abnormal closure.

Change-Id: Ib4669fb57cddfc6ce61f9a3991b80db3f2b3e9ab

Author: sasha-gitg

src/google/adk/features/_feature_registry.py

@@ -138,7 +138,7 @@ class FeatureConfig:
         FeatureStage.EXPERIMENTAL, default_on=True
     ),
     FeatureName._MCP_GRACEFUL_ERROR_HANDLING: FeatureConfig(
-        FeatureStage.EXPERIMENTAL, default_on=False
+        FeatureStage.EXPERIMENTAL, default_on=True
     ),
     FeatureName.PROGRESSIVE_SSE_STREAMING: FeatureConfig(
         FeatureStage.EXPERIMENTAL, default_on=True

src/google/adk/tools/mcp_tool/session_context.py

@@ -130,6 +130,12 @@ async def start(self) -> ClientSession:
       if not self._task:
         self._task = asyncio.create_task(self._run())
 
+        def _retrieve_exception(t: asyncio.Task):
+          if not t.cancelled():
+            t.exception()
+
+        self._task.add_done_callback(_retrieve_exception)
+
     await self._ready_event.wait()
 
     if self._task.cancelled():

tests/unittests/tools/mcp_tool/test_mcp_session_manager.py

@@ -588,6 +588,9 @@ async def test_create_session_cleans_up_without_aclose_if_loop_is_different(
       self,
   ):
     """Verify that sessions from different loops are cleaned up without calling aclose()."""
+    from google.adk.features import FeatureName
+    from google.adk.features._feature_registry import temporary_feature_override
+
     manager = MCPSessionManager(self.mock_stdio_connection_params)
 
     # 1. Simulate a session created in a "different" loop
@@ -617,8 +620,11 @@ async def test_create_session_cleans_up_without_aclose_if_loop_is_different(
           mock_wait_for.return_value = new_session
           mock_session_context_class.return_value = AsyncMock()
 
-          # 3. Call create_session
-          session = await manager.create_session()
+          # 3. Call create_session with flag off to hit wait_for branch
+          with temporary_feature_override(
+              FeatureName._MCP_GRACEFUL_ERROR_HANDLING, False
+          ):
+            session = await manager.create_session()
 
           # 4. Verify results
           assert session == new_session
@@ -969,8 +975,8 @@ class TestMCPGracefulErrorHandlingFlagContract:
   loudly so we don't silently break GE's rollout.
   """
 
-  def test_default_state_is_off_so_cl_is_a_noop(self):
-    """The CL must be a no-op until GE explicitly enables it."""
+  def test_default_state_is_on(self):
+    """The fix must be enabled by default."""
     import os
 
     from google.adk.features import FeatureName
@@ -981,34 +987,34 @@ def test_default_state_is_off_so_cl_is_a_noop(self):
     saved = {k: os.environ.pop(k) for k in (enable, disable) if k in os.environ}
     try:
       assert (
-          is_feature_enabled(FeatureName._MCP_GRACEFUL_ERROR_HANDLING) is False
+          is_feature_enabled(FeatureName._MCP_GRACEFUL_ERROR_HANDLING) is True
       )
     finally:
       os.environ.update(saved)
 
-  def test_env_var_enable_flips_flag_on_at_runtime(self):
-    """The env var GE will set must turn the fix on without a rebuild."""
+  def test_env_var_disable_flips_flag_off_at_runtime(self):
+    """The env var must turn the fix off without a rebuild."""
     import os
 
     from google.adk.features import FeatureName
     from google.adk.features import is_feature_enabled
 
-    enable = "ADK_ENABLE_MCP_GRACEFUL_ERROR_HANDLING"
-    saved = os.environ.pop(enable, None)
+    disable = "ADK_DISABLE_MCP_GRACEFUL_ERROR_HANDLING"
+    saved = os.environ.pop(disable, None)
     try:
-      os.environ[enable] = "1"
+      os.environ[disable] = "1"
       assert (
-          is_feature_enabled(FeatureName._MCP_GRACEFUL_ERROR_HANDLING) is True
+          is_feature_enabled(FeatureName._MCP_GRACEFUL_ERROR_HANDLING) is False
       )
       # And once it's removed, we revert. Confirms the value is read
       # live from os.environ on every call (no caching, no binary push).
-      del os.environ[enable]
+      del os.environ[disable]
       assert (
-          is_feature_enabled(FeatureName._MCP_GRACEFUL_ERROR_HANDLING) is False
+          is_feature_enabled(FeatureName._MCP_GRACEFUL_ERROR_HANDLING) is True
       )
     finally:
       if saved is not None:
-        os.environ[enable] = saved
+        os.environ[disable] = saved
 
   def test_env_var_disable_acts_as_kill_switch(self):
     """The disable env var lets consumers turn off without a rebuild."""