TypeVarTuple and ParamSpec allow too few arguments to be specified when used together

[Viicos]

Bug report

Bug description:

Following #99344, it seems to me that the following should raise a TypeError:

class A[*Ts, **P]: ...


A[int]  # Should raise
A[int, str]  # Should raise (parameters for **P should be specified as [str], unless P was the only parameter).
 
class B[T, **P]: ...

B[int, str]  # Should also raise

I understand that these use cases may be really hard to support at runtime, if so I don't mind closing the issue.

CPython versions tested on:

3.14

Operating systems tested on:

No response

[sobolevn]

Right now

class A[*Ts, **P]: ...

A[int]

substitutes *Ts as () and P as [int], which looks kinda correct?

See

cpython/Lib/typing.py

Lines 1085 to 1108 in 6fb5f7f

	def _paramspec_subst(self, arg):
	if isinstance(arg, (list, tuple)):
	arg = tuple(_type_check(a, "Expected a type.") for a in arg)
	elif not _is_param_expr(arg):
	raise TypeError(f"Expected a list of types, an ellipsis, "
	f"ParamSpec, or Concatenate. Got {arg}")
	return arg
	def _paramspec_prepare_subst(self, alias, args):
	params = alias.__parameters__
	i = params.index(self)
	if i == len(args) and self.has_default():
	args = [*args, self.__default__]
	if i >= len(args):
	raise TypeError(f"Too few arguments for {alias}")
	# Special case where Z[[int, str, bool]] == Z[int, str, bool] in PEP 612.
	if len(params) == 1 and not _is_param_expr(args[0]):
	assert i == 0
	args = (args,)
	# Convert lists to tuples to help other libraries cache the results.
	elif isinstance(args[i], list):
	args = (*args[:i], tuple(args[i]), *args[i+1:])
	return args

_paramspec_prepare_subst args=((), <class 'int'>)
__main__.A[int]

But,

class B[T, **P]: ...

B[int, str]  # Should also raise

clearly looks like a bug to me.

cpython/Lib/typing.py

Lines 1035 to 1041 in 6fb5f7f

	def _typevar_subst(self, arg):
	msg = "Parameters to generic types must be types."
	arg = _type_check(arg, msg, is_argument=True)
	if ((isinstance(arg, _GenericAlias) and arg.__origin__ is Unpack) or
	(isinstance(arg, GenericAlias) and getattr(arg, '__unpacked__', False))):
	raise TypeError(f"{arg} is not valid as type argument")
	return arg

is never called for some reason.

It also substitutes as:

_paramspec_prepare_subst args=(<class 'int'>, <class 'str'>)
__main__.B[int, str]

However, B[[int, str]] does not work and raises: TypeError: Too few arguments for <class '__main__.B'>

[Viicos]

Right now

class A[*Ts, **P]: ...

A[int]

ParamSpec should be parameterized with a list/tuple of types, unless the ParamSpec is the only parameter present. This is documented at the end of this section:

Another difference between TypeVar and ParamSpec is that a generic with only one parameter specification variable will accept parameter lists in the forms X[[Type1, Type2, ...]] and also X[Type1, Type2, ...] for aesthetic reasons. Internally, the latter is converted to the former, so the following are equivalent:

>>> class X[**P]: ...
...
>>> X[int, str]
__main__.X[[int, str]]
>>> X[[int, str]]
__main__.X[[int, str]]

So I think in:

class A[*Ts, **P]: ...

A[int]

*Ts indeed substitutes as (), and P as int but this should raise because the convenience of parameterizing with int instead of [int] doesn't apply here.

Valid forms would be:

A[int, [str, int]]
# Ts -> (int,), P -> [str, int]
A[[str, int]]
# TS -> (), P -> [str, int]

[Viicos]

is never called for some reason.

I think this is coming from the _generic_class_getitem() function. It is called in two cases:

• When you define parameters on your class, i.e. when doing Generic[T, ...] or Protocol[T, ...]. This is covered by the first part of the function:

  cpython/Lib/typing.py

  Lines 1128 to 1135 in 90b82f2

  	if is_generic_or_protocol:
  	# Generic and Protocol can only be subscripted with unique type variables.
  	if not args:
  	raise TypeError(
  	f"Parameter list to {cls.__qualname__}[...] cannot be empty"
  	)
  	if not all(_is_typevar_like(p) for p in args):
  	raise TypeError(

• When you parameterize your generic class, i.e. when doing MyClass[int, ...]. This is covered by the second part of the function:

  cpython/Lib/typing.py

  Lines 1141 to 1157 in 90b82f2

  	else:
  	# Subscripting a regular Generic subclass.
  	for param in cls.__parameters__:
  	prepare = getattr(param, '__typing_prepare_subst__', None)
  	if prepare is not None:
  	args = prepare(cls, args)
  	_check_generic_specialization(cls, args)
  	new_args = []
  	for param, new_arg in zip(cls.__parameters__, args):
  	if isinstance(param, TypeVarTuple):
  	new_args.extend(new_arg)
  	else:
  	new_args.append(new_arg)
  	args = tuple(new_args)
  	return _GenericAlias(cls, args)

As we can see, the __typing_prepare_subst__ functions are being called for each parameter, but we don't ever call __typing_subst__ on them.

However, if you parameterize the generic class with the same parameters, this code path will run and you will end up with a _GenericAlias instance:

T = TypeVar('T')
P = ParamSpec('P')

class MyClass(Generic[T, P]): pass

alias = MyClass[T, P]
alias.__class__
#> typing._GenericAlias

And when doing alias[int, str], we will run _GenericAlias.__getitem__() this time, which does implement the necessary checks and calls out __typing_subst__ (_GenericAlias.__getitem__() calls _determine_new_args(), which does similar things as _generic_class_getitem() but also calls _make_substitution(), which has a lot of additional logic).

Things seem to work as expected when parameterizing the alias:

class MyClass(Generic[T, P]): pass

alias = MyClass[T, P]
alias[int]
#> TypeError: Too few arguments for __main__.MyClass[~T, ~P]
alias[int, str]
#> TypeError: Expected a list of types, an ellipsis, ParamSpec, or Concatenate. Got <class 'str'>

class MyClass(Generic[*Ts, P]): pass

alias = MyClass[*Ts, P]
alias[int]
#> TypeError: Expected a list of types, an ellipsis, ParamSpec, or Concatenate. Got <class 'int'>
alias[str, int]
#> TypeError: Expected a list of types, an ellipsis, ParamSpec, or Concatenate. Got <class 'int'>

Before working on fix, I'd like to know if this the omission _generic_class_getitem() is actually not an omission but expected behavior. cc @JelleZijlstra, as you worked PEP 695/696 implementation on this function.

[Viicos]

Actually, that might be too disruptive to add. It might even have been intentional, but this would break use cases where forward references are used when parameterizing classes:

class B[T, **P]: ...

B[int, "ForwardP"]  # If we were to check the class arguments here, this would fail as str is not a list of types,an ellipsis, ParamSpec or Concatenate.

ForwardP = ParamSpec("ForwardP")

Maybe we could try to do the necessary checks, but silently accept strings and forwardrefs?

[Viicos]

I'm closing this one in favor of #132100, which gives more context.