Bug: Imports do not seems to be handled correctly in 0.720

[futursolo]

I recently bumped my Mypy version from 0.701 to 0.720. It seems like Mypy's starting to process imported modules residing in the same directory as the one that's being checked. But it does not handle them correctly.

Environment:

macOS 10.14.5
Python 3.7.4
Mypy 0.720/0.701

Directory Structure:

|
|-- mod_a
|  |-- __init__.py
|
|-- mod_b
   |-- __init__.py
   |-- b.py

Content of mod_a/__init__.py:

import mod_b

class A(mod_b.B):
    pass

Content of mod_b/__init__.py:

from .b import B

Content of mod_b/b.py:

class B:
    pass

Command:

mypy --python-version 3.7 \
    --strict --show-traceback \
    --ignore-missing-imports \
    --show-error-context \
    -p mod_a

Error message in 0.720(no error in 0.701):

mod_a/__init__.py:4: error: Name 'mod_b.B' is not defined
mod_a/__init__.py: note: In class "A":
mod_a/__init__.py:4: error: Class cannot subclass 'B' (has type 'Any')

[JelleZijlstra]

This is the expected behavior of the new --no-implicit-reexport flag, which is included in --strict. To make your code pass, use from .b import B as B in mod_b/__init__.py.

This is documented in https://mypy.readthedocs.io/en/latest/command_line.html#miscellaneous-strictness-flags.

[futursolo]

Is there any chance we can make the error message clearer?

[JukkaL]

Yes, the error message should be clearer. Created #7251 to track this.

[futursolo]

Also, upon further testing, --no-implicit-reexport does not seem to respect __all__either. Shouldn't variable names listed in __all__ also be considered as exported names?

[JelleZijlstra]

That's being fixed in #7099.