The classestyping.IO,typing.BinaryIO, andtyping.TextIO look like they want to be Protocols or ABCs, but in fact they are defined as regular generic classes, both at runtime and in typeshed. However, they are meant to encompass the concrete IO classes defined in theio module, and in typeshed we implement that by having these classes inherit fromtyping.*IO classes, even though there is no such inheritance at runtime.

This can easily lead to unsound behavior:

importio,typingdeff(x:int|io.BytesIO)->int:ifisinstance(x,typing.BinaryIO):returnx.fileno()returnxf(io.BytesIO())+1# boom

Type checkers thinkBytesIO is a subclass ofBinaryIO, because that's how it's defined in typeshed, but in fact it isn't at runtime.

I can see a few solutions:

1. Special-casetyping.*IO in the spec and say that type checkers should rejectisinstance()/issubclass() calls involving them.
2. Deprecate thetyping.*IO classes and eventually remove them, nudging people to use their own Protocols (or the newio.Reader/io.Writer) instead. This is conceptually clean but may be annoying for a lot of users; the typing classes are nice to use in simple application code.
3. Make these classes actually (runtime-checkable?) Protocols at runtime, though they would be unwieldily large.

Even if we do (2) or (3) type checkers might still want to do (1) since it will be a while before the relevant runtime changes take effect.

(Noticed this while looking intopython/cpython#133492 .)