Should http.client.HTTPMessage be subtype of Mapping[str, str]?

[Unrud]

Example:

URLopener.http_error has the argument headers: Mapping[str, str]. It gets called with the value of HTTPResponse.msg which has the type HTTPMessage.

[srittau]

Since HTTPMessage does not derive from Mapping at runtime and is also not registered as implementing Mapping, it shouldn't be one in stubs either. See some related discussion in #3576. In my opinion, the correct fix is to add a MappingLike protocol to _typeshed and use that for the headers argument.

[Akuli]

I don't think MappingLike should go to _typeshed, because the meaning of "mapping-like" depends on the code that uses it. For example, it could mean "has .items() method", "gives keys when iterating and values from .__getitem__()", or "has .keys() and .__getitem__(). A single protocol can't cover it all. We already have SupportsItems, SupportsKeysAndGetItem and a couple other classes, so we should just use those.

[Unrud]

Fixed in 9bba8a4