argparse supports options with variable numbers of args. gh-53584 tracks a long-standing bug where options with a non-fixed number of args (nargs='?', nargs='*', or nargs='+') will greedily consume argument strings that should have been reserved for required positional arguments. For example:

parser.add_argument('--foo', nargs='?')
parser.add_argument('bar')
parser.parse_args(['--foo', 'abc'])  # bar got nothing

argparse works by pattern-matching, with the bulk of its logic defined in _parse_known_args(). That method encodes each argument string as a single character ('O' for a recognised option flag, 'A' for everything else, and '-' for strings following --) which gives us a resulting string pattern (e.g. 'OAOA'). We then "consume" arguments using this string pattern, handling both positionals (via the consume_positionals() closure) and optionals (via consume_optional()).

The bug arises in the latter. consume_optional() processes options by building a regex based on the option's nargs (e.g. '-*A?-*' for nargs='?') and then runs this regex against the remainder of the string pattern (i.e. anything following the option flag). This is done via _match_argument(). The regex will always stop at the next option flag ('O' token) but for non-fixed nargs values like '?' and '+' may greedily consume every positional ('A' token) up to that point.

This fix works by manipulating the string pattern as part of our optional consumption. Any 'A' tokens that are required by remaining positionals are masked to 'O' to prevent the regex consuming them. Masking will only consider tokens up to the next option flag ('O') and it both accounts for what future options can absorb (to avoid masking more than is necessary) and ensures that at least the minimum arguments required for the optional are actually consumed.

In addition, we also handle the parse_intermixed_args() case. In intermixed mode, positional-typed arguments already collected to the left of the current option are also accounted for, since they will satisfy positionals in the second-pass parse.

Note that this is a rather gnarly issue, and I've done my best to avoid changing the API behavior of the module without layering on too much additional complexity, in the hope that this might actually be backportable. Hopefully my proposed approach is sound but I'm happy to iterate on this if there's something I've missed or there is a better way to do this.