I think there's quite a few scripts out there that expect sys.argv[0] to be the path to the pex package itself. It might be safer to just patch sys.modules by hand to avoid changing sys.argv[0].

I just ran into this myself where pkg_resources doesn't work for __main__, so I'm a big fan of this :)

Fair point about code expecting the existing behavior of sys.argv[0]. Alas the runpy.run_module function doesn't provide a way to override only sys.modules but not sys.argv[0]. So patching sys.modules by hand would amount to reinventing much of the logic in run_module.

It will essentially be inlining the run_module method. Something like

def run_module_with_sys_modules_override(mod_name):
  run_name = '__main__'
  _, loader, code, fname = runpy._get_module_details(mod_name)
  pkg_name = mod_name.rpartition('.')[0]
  with runpy._TempModule(run_name) as temp_module:
      mod_globals = temp_module.module.__dict__
      runpy._run_code(code, mod_globals, None,
                      run_name, fname, loader, pkg_name)

Calling the internal methods of runpy isn't ideal.

Definitely agree that depending on runpy internals isn't great. Before we go down too much into thinking how to implement this, I'd like to make a case that leaving sys.argv[0] alone is probably the more correct behavior in general. sys.argv[0] is how the program was started. You should be able to os.exec([sys.executable] + sys.argv) for essentially restarting the program. Does that seem reasonable?

From reading https://www.python.org/dev/peps/pep-0338/ , it looks like the argument for changing sys.argv[0] is that some scripts expect it. The example given is regrtest.py, and indeed, the script would not work due to an assert: https://github.com/python/cpython/blob/2.7/Lib/test/regrtest.py#L1588 . However, I can't figure out why that assert is there at all - nothing else seems to read sys.argv[0]. It seems to be a remnant of another time where the script used sys.argv[0] instead of __file__ and slowly transitioned to __file__ (which matches the pattern in other parts of the python world).

If the above argument makes sense, and others agree with it, the right thing to do here would probably be to propose changing runpy for python 3.6+. Assuming that's accepted, we can backport the functionality with the "hack" you posted for < 3.6. I'll send an mail to python-dev to see how folks feel about this.

Yeah totally agree sys.argv[0] is best left alone. Looks like I'll need to patch pex for my organization for the time being.

The rationale for changing sys.argv[0] is so tools like debuggers and coverage analysers can use run_module to invoke the script to be run, while still making the invoked script believe its running normally.

However, there are currently still some significant challenges in doing that in practice: https://bugs.python.org/issue9325