Is this not something you can add together from wrapper.pytestmark and func.pytestmark?

Is that actually part of the public API?

Internally, pytest uses these three functions to do the mark manipulation:

def get_unpacked_marks(obj):
    """
    obtain the unpacked marks that are stored on an object
    """
    mark_list = getattr(obj, "pytestmark", [])
    if not isinstance(mark_list, list):
        mark_list = [mark_list]
    return normalize_mark_list(mark_list)


def normalize_mark_list(mark_list):
    """
    normalizes marker decorating helpers to mark objects

    :type mark_list: List[Union[Mark, Markdecorator]]
    :rtype: List[Mark]
    """
    extracted = [
        getattr(mark, "mark", mark) for mark in mark_list
    ]  # unpack MarkDecorator
    for mark in extracted:
        if not isinstance(mark, Mark):
            raise TypeError("got {!r} instead of Mark".format(mark))
    return [x for x in extracted if isinstance(x, Mark)]


def store_mark(obj, mark):
    """store a Mark on an object
    this is used to implement the Mark declarations/decorators correctly
    """
    assert isinstance(mark, Mark), mark
    # always reassign name to avoid updating pytestmark
    # in a reference that was only borrowed
    obj.pytestmark = get_unpacked_marks(obj) + [mark]

I was between just vendoring them all, doing the semi-private import, or adding the obj.pytestmark list together. I took the middle route, I'll change it to just adding the pytestmark lists together.

As a side note, I find your code formatting style extremely hard to read.

I agree this just looks like the worst of black, things like

wrapper.pytestmark = get_unpacked_marks(
    wrapper
) + get_unpacked_marks(func)

just completely obfuscate the structure of the code -- compare with

wrapper.pytestmark = (get_unpacked_marks(wrapper)
                      + get_unpacked_marks(func))

(or with an additional newline after the opening parenthesis).

This is indeed black's doing.

Ironically, the biggest selling point of black for me is avoiding exactly these conversations.

The fact that it's automated doesn't absolve it from completely violating basic esthetics.

fwiw @NelleV and I have been trying to come up with a better automated scheme and it's certainly not trivial... Nevertheless, since afaict matplotlib hasn't added black as a dependency, it's a bit weird to add badly black-formatted code...

(And @tacaswell it looks like you've set it to 71c wide? Which is unconventional to say the least...)

(I personally just don't buy into the idea that code can be properly formatted without understanding the underlying semantics.)

(Juan is considering doing constrained reinforcement learning to learn the rules…)

I thought I set it to 80c (to keep our pyflakes happy). Black does a slightly better job if you give it a bit more width to work with.