Hopefully I'm not being obtuse, but I seem to be getting incorrect/unexpected help output when using mutually_exclusive_group under an add_argument_group if this layering is happening in a parent parser.

Here's an example of the triggering usage:

import argparse

global_options = argparse.ArgumentParser(add_help=False)
global_options.add_argument('--summary', action='store_true', help='summarize information')
global_options.add_argument('--verbose', action='store_true', help='tell us more')

output_format = global_options.add_argument_group("Output format", "ways to foo")
styles = output_format.add_mutually_exclusive_group()
styles.add_argument('--plain', dest='style')
styles.add_argument('--green', dest='style')
styles.add_argument('--blue', dest='style')

parser = argparse.ArgumentParser()
commands = parser.add_subparsers()
hit = commands.add_parser('hit', parents=[global_options])
miss = commands.add_parser('miss', parents=[global_options])

print(parser.parse_args(['hit', '-h']))

which produces:

usage: bar.py hit [-h] [--summary] [--verbose]
                  [--plain STYLE | --green STYLE | --blue STYLE]

optional arguments:
  -h, --help     show this help message and exit
  --summary      summarize information
  --verbose      tell us more
  --plain STYLE
  --green STYLE
  --blue STYLE

Output format:
  ways to foo


"--plain" "--green" and "--blue" *should* be under "Output format"