stubgen vs. inspect.signature

[smurfix]

How difficult would it be to teach nanobind to create the attributes required for inspect.signature to do a passable job?

stubgen doesn't help with real-time introspection.

[wjakob]

I haven't used inspect.signature much. Can you explain what things trip it up?

Please note that heavy changes here might be impossible due to SemVer versioning.

[smurfix]

Anything, basically.
void take(int) results in def take(arg: int, /) -> None: ... from nanobind.stubgen, but an unhelpful <Signature (*args, **kwargs)> from inspect.signature.

[wjakob]

It seems that the implementation is only designed to work for functions declared by Python itself.

[smurfix]

Mmh. The implementation of inspect.signature seems to merely access some by-now-moderately-standard Python attributes and syntesizes its reply from them.
While it'd be great for nanobind to simply(?!?) add them, my knowledge of C++ templates is (unfortunately) way not up to that task.

[enerjazzer]

I vote for this feature too.

I need proper reflection of a function signature as I need this information to generate C++ code from Python.
inspect.signature(...).parameters gives me an OrderedDict of parameters, exactly what I need.

The existing __nb_signature__ is essentially a string, so I need to parse it somehow (like using undocumented sphinx functions).

[wjakob]

It's easy to vote, but if nobody does the work it is not going to happen. FWIW this feature is not on my todo list.