There is one test failing because we are using format_simple everywhere. Should we skip expected types when dealing with FunctionLike or use format?

I'd start by not adding the "; expected blah" part to the error if format_simple() returns something empty.

Do we really want to do that? Even without this patch mypy prints malformed messages, e.g.

(mypy-dev) ~/src/mypy [master] λ cat x.py 
from typing import Dict, Callable
def f() -> str:
    return 'foo'
d: Dict[str, Callable[..., int]] = {'bar': f}
(mypy-dev) ~/src/mypy [master] λ mypy x.py 
x.py:4: error: Dict entry 0 has incompatible type "str":

Notice the : at the end. And this doesn't happen with Callables only (which I understand may get long), but even with Type.

I'd go for using format method, just like MessageBuilder.invalid_index_type does.

Using format seems reasonable to me.

@miedzinski After some thinking, it looks like your initial design was actually right and it is OK to use is_subtype in the context of inference for dict literals. Also the added tests don't trigger the new error message (this is related). So that I would restore the is_subtype and rewrite the added tests like you proposed on gitter.

Sorry for troubles.