fix(mcp): preserve loop during probes

tools/mcp_tool.py

@@ -3946,7 +3946,7 @@ def probe_mcp_server_tools() -> Dict[str, List[tuple]]:
     except Exception as exc:
         logger.debug("MCP probe failed: %s", exc)
     finally:
-        _stop_mcp_loop()
+        _stop_mcp_loop_if_idle()
 
     return result
 
@@ -4084,10 +4084,25 @@ def _kill_orphaned_mcp_children(include_active: bool = False) -> None:
         )
 
 
-def _stop_mcp_loop():
+def _stop_mcp_loop_if_idle() -> bool:
+    """Stop the MCP loop only when no registered server still owns it.
+
+    Probe paths create temporary MCPServerTask instances that are not placed in
+    ``_servers``.  They should clean up an otherwise-idle loop, but must not
+    tear down the process-global loop when live agent tools are registered on
+    it.  Otherwise a dashboard/CLI probe can make later MCP tool calls fail
+    with ``MCP event loop is not running``.
+    """
+    return _stop_mcp_loop(only_if_idle=True)
+
+
+def _stop_mcp_loop(*, only_if_idle: bool = False) -> bool:
     """Stop the background event loop and join its thread."""
     global _mcp_loop, _mcp_thread
     with _lock:
+        if only_if_idle and (_servers or _server_connecting):
+            logger.debug("Leaving MCP event loop running; active servers are registered or connecting")
+            return False
         loop = _mcp_loop
         thread = _mcp_thread
         _mcp_loop = None
@@ -4104,3 +4119,4 @@ def _stop_mcp_loop():
         # graceful shutdown are now orphaned — include active PIDs too
         # since the loop is gone and no session can still be in flight.
         _kill_orphaned_mcp_children(include_active=True)
+    return True