I don’t have a strong opinion about changing __all__ in these cases. I only noticed the potential problem when I went to add a new class to the module, and thought this was common practice. If we leave it as it is, it would be good to add comment in the source code explaining this decision. Also the test case could still be useful to catch future bugs.

The currently situation means the status constants are be missing from pydoc help output, and are not available when you do “from http.client import *” in the interactive interpreter.

HTTPMessage is a strange class indeed; see Issue 5053. But it is referenced a few times by the documentation, so I originally assumed it should be in __all__.