Callback-Protocol for the Union of all Callable-types that only take keyword arguments.

[randolf-scholz]

Is it possible to express the type of the union of all Callable-types that only take keyword arguments (*)? I tried the following, but it gives both false positives and false negatives:

from typing import Protocol, ParamSpec

P = ParamSpec("P")

class KWOnlyFunc(Protocol[P]):
    """Protocol for callables that accept only keyword-arguments."""
    def __call__(self, **kwargs: P.kwargs) -> None: ...

def false0(x: int) -> None: ...
def false1(x: int, **kwargs: int) -> None: ...
def false2(*args: int, **kwargs: int) -> None: ...
def false3(*args: int, z: int, **kwargs: int) -> None: ...
def false4(*args: int, z: int=0, **kwargs: int) -> None: ...
def true0(**kwargs: int) -> None: ...
def true1(*, z: int) -> None: ...
def true2(*, z: int, **kwargs: int) -> None: ...
def true3(*, z: int=0) -> None: ...
def true4(*, z: int=0, **kwargs: int) -> None: ...

f0: KWOnlyFunc = false0 # ✔  true negative
f1: KWOnlyFunc = false1 # ✔  true negative
f2: KWOnlyFunc = false2 # ✘  false postive
f3: KWOnlyFunc = false3 # ✔  true negative
f4: KWOnlyFunc = false4 # ✘  false postive

t0: KWOnlyFunc = true0  # ✔  true positive
t1: KWOnlyFunc = true1  # ✘  false negative
t2: KWOnlyFunc = true2  # ✘  false negative
t3: KWOnlyFunc = true3  # ✘  false negative
t4: KWOnlyFunc = true4  # ✔  true positive

https://mypy-play.net/?mypy=latest&python=3.11&gist=6307ab826f8a3bf3ee6274999f15dde6

(*) I.e. They satisfy the following TypeGuard

def is_kw_only_func(f: Callable) -> TypeGuard[KWOnlyFunc]:
    sig = inspect.signature(f)
    return all(p.kind in (p.KW_ONLY, p.VAR_KEYWORD) for p in sig.parameters.values())

[erictraut]

It's not legal to specify a **kwargs: P.kwargs parameter without a corresponding *args: P.args. PEP 612 is very clear on this point. Mypy should generate an error for your invalid protocol definition, but it doesn't.

[randolf-scholz]

Do you know if there are any plans for adding a map-type variadic (as a counterpart to TypeVarTuple?), this would solve my problem.

[randolf-scholz]

It's also noteworthy that the rationale given in the PEP for disallowing P.kwargs without P.args is not really applicable here. Seems like this option was overlooked.

[Daverball]

PEP612 specifically mentions bound as not having been specified/implemented yet for ParamSpec, even though for all intents and purposes it is a TypeVar, so it should be added in the future.

So in a future PEP bound on ParamSpec could be used to provide an upper bound on the shape of the ParamSpec which would allow you to restrict the shape to contain no positional parameters. But I think what needs to happen before that is to introduce a notation for all the different kinds of arguments in a ParamSpec, since right now you can only bind unnamed parameters and can't specify optional parameters or variadic arguments either.

Something like PEP695 but for function signatures, rather than generics.