Hack: unbreak instantiation of pydantic models

Author: dandavison

temporalio/worker/workflow_sandbox/_restrictions.py

@@ -951,21 +951,32 @@ def _is_restrictable(v: Any) -> bool:
 
 
 class _RestrictedProxy:
+    # When we instantiate this class, we have the signature of:
+    #   __init__(
+    #       self,
+    #       name: str,
+    #       obj: Any,
+    #       context: RestrictionContext,
+    #       matcher: SandboxMatcher
+    #   )
+    # However there are some edge cases; see comments below.
+
+    def __new__(cls, *args, **kwargs) -> Any:
+        # When pydantic instantiates a model containing a field whose type
+        # annotation is proxied, it attempts to call the field type constructor
+        # with the field value but instead calls the _RestrictedProxy
+        # constructor. Return the field value.
+        if len(args) == 1:
+            return args[0]
+        return super().__new__(cls)
+
     def __init__(self, *args, **kwargs) -> None:
-        # When we instantiate this class, we have the signature of:
-        #   __init__(
-        #       self,
-        #       name: str,
-        #       obj: Any,
-        #       context: RestrictionContext,
-        #       matcher: SandboxMatcher
-        #   )
-        # However when Python subclasses a class, it calls metaclass() on the
+        # When Python subclasses a class, it calls metaclass() on the
         # class object which doesn't match these args. For now, we'll just
         # ignore inits on these metadata classes.
         # TODO(cretz): Properly support subclassing restricted classes in
         # sandbox
-        if isinstance(args[2], RestrictionContext):
+        if len(args) == 4 and isinstance(args[2], RestrictionContext):
             _trace("__init__ on %s", args[0])
             _RestrictionState(
                 name=args[0], obj=args[1], context=args[2], matcher=args[3]