It appears when adding auto-numbered positional fields in python 3.1 Formatter.parse was not updated to handle them and currently returns an empty string as the field name.

list(Formatter().parse('hello {}'))  # [('hello ', '', '', None)]

This does not align with Formatter.get_field which according to the docs: "Given field_name as returned by parse() (see above), convert it to an object to be formatted."

When supplying an empty string to .get_field() you get a KeyError

Formatter().get_field("", [1, 2, 3], {}). # raises KeyError

For reference, the documentation is at https://docs.python.org/3/library/string.html#custom-string-formatting

I guess in your example it should return:
[('hello ', '0', '', None)]

Yes it should return a string containing the index of the positional argument i.e. "0" so that it is compatible with .get_field(). Side note: It's a somewhat weird that .get_field expects a string while .get_value expects an int for positional arguments.

Side note: It's a somewhat weird that .get_field expects a string while .get_value expects an int for positional arguments.

.parse is just concerned with parsing, so it works on and returns strings. .get_field takes strings because it is the thing that's trying to determine whether or not a field name looks like an integer or not. At least that's how I remember it.

Another thing that occurred to me is the question of what .parse() should do when a mix of auto-numbered and manually numbered fields is supplied e.g. {}{1}. As of now .parse() happily processes such inputs and some other piece of code deals with this and ultimately raises an exception that mixing manual with automatic numbering is not allowed. If .parse() supported automatic numbering it would have to be aware of this too I guess?

The more I think about this, the more I think it's not .parse's job to fill in the field numbers, it's the job of whoever is calling it.

Just as it's not .parse's job to give you an error if you switch back and forth between numbered and un-numbered fields.

It's literally just telling you what's in the string as it breaks it apart, not assigning any further meaning to the parts. I guess I should have called it .lex, not .parse.

That definition of .parse() definitely makes sense. Do you then think this is out of scope for Formatter in general or just for .parse()?. Just for reference, this is what I currently use to get automatic numbering to work for my use case.

def parse_command_template(format_string):

    auto_numbering_error = ValueError(
        'cannot switch from automatic field numbering to manual field specification')

    index = 0
    auto_numbering = None

    for literal_text, field_name, spec, conversion in Formatter().parse(format_string):
        if field_name is not None:
            if field_name.isdigit():
                if auto_numbering is True:
                    raise auto_numbering_error
                auto_numbering = False

            if field_name == '':
                if auto_numbering is False:
                    raise auto_numbering_error
                auto_numbering = True
                field_name = str(index)
                index += 1

        yield literal_text, field_name, spec, conversion

I think your code is rational. But since string.Formatter gets such little use, I'm not sure it's worth adding this to the stdlib. On the other hand, it could be used internal to string.Formatter.

We'd need to pick a better name, though. And maybe it should return the field_name as an int.

That is, return field_name as an int if it's an int, otherwise as a string.

The more I think about this, the more I think it's not .parse's job to fill in the field numbers, it's the job of whoever is calling it.

I agree with this. .parse correctly returns empty string for empty field name. This is factorised well providing one-to-one relationship between input and output.

However, the main issue is that this is not possible for things with nesting to work.

If parsing only one level, then suggested parse change would return things correctly. E.g.:

list(string.Formatter().parse('{}{}'))
[('', '0', '', None), ('', '1', '', None)]

However, it would not return things correctly for:

list(string.Formatter().parse('{:.{}f}{}'))
[('', '0', '.{}f', None), ('', '1', '', None)]

While its equivalent of manually numbered fields would look like:

fmt = '{0:.{1}f}{2}'
# And correct parse output should be:
list(string.Formatter().parse('{:.{}f}{}'))
[('', '0', '.{}f', None), ('', '2', '', None)]

Auto-numbering is very easy to do outside it. Especially for a flat case. I suggest closing this as "Not planned".

That's reasonable! Looks like all that's left is to mention this in the parse/get_field docs.