Hello,
I wonder if there is a reason why the xtensor_type_caster_base uses xt::xtensor for the signature

instead of using numpy.ndarray[...]

In my opinion, using the Numpy name is more descriptive to a Python user, and actually matches the type.

I made a quick test using

_("numpy.ndarray[") + npy_format_descriptor<typename Type::value_type>::name + _("]");

which seems to work.

I could send a PR if you are interested.

Thank you