[From the old argparse tracker: http://code.google.com/p/argparse/issues/detail?id=20]

You can't follow a nargs='+' optional argument with a positional argument:

>>> import argparse
>>> parser = argparse.ArgumentParser(prog='PROG')
>>> parser.add_argument('--badger', nargs='+')
>>> parser.add_argument('spam')
>>> parser.parse_args('--badger A B C D'.split())
usage: PROG [-h] [--badger BADGER [BADGER ...]] spam
PROG: error: too few arguments

Ideally, this should produce:

>>> parser.parse_args('--badger A B C D'.split())
Namespace(badger=['A', 'B', 'C'], spam='D')

The problem is that the nargs='+' causes the optional to consume all the arguments following it, even though we should know that we need to save one for the final positional argument.

A workaround is to specify '--', e.g.:

>>> parser.parse_args('--badger A B C -- D'.split())
Namespace(badger=['A', 'B', 'C'], spam='D')

The problem arises from the fact that argparse uses regular-expression style matching for positional arguments, but it does that separately from what it does for optional arguments.

One solution might be to build a regular expression of the possible things a parser could match. So given a parser like::

  parser = argparse.ArgumentParser()
  parser.add_argument('-w')
  parser.add_argument('-x', nargs='+')
  parser.add_argument('y')
  parser.add_argument('z', nargs='*')

the regular expression might look something like (where positionals have been replaced by the character A)::

  (-w A)? (-x A+)? A (-w A)? (-x A+)? A* (-w A)? (-x A+)?

Note that the optionals can appear between any positionals, so I have to repeat their regular expressions multiple times. Because of this, I worry about how big the regular expression might grow to be for large parsers. But maybe this is the right way to solve the problem.

This is definitely a different bug from the one you just marked it as a duplicate of.

I read too fast, I’m sorry for that.

Here's a unit test for the simplest cases.

Looks good to me.  Do you want to propose a code patch too, and/or more tests for non-simple cases?

My attached patch adds the "--" between the optionals and the arguments, if there are optionals which have variable length and at least some positional argument can be provided.

Patch is for python 3.2 svn revision 86553.

Patch looks good, thanks!  Does it integrate Catherine’s tests?

is there a chance to fix this issue?

There is.  Someone wanting to help could reply to the question I asked :)

So Kotan's patch doesn't actually solve the original problem. Instead, it inserts the workaround into the help message of the parser. I think this is probably not the right fix. We should probably do two things:

(1) Right now: create a documentation patch which at least explains the current limitations of argparse parsing, and describes the '--' workaround. Probably this patch should add a separate section about '--', give an example like the one in this issue, and then cross-reference this section from nargs='?', nargs='*', nargs='+' and the "Arguments containing -" section.

(2) Longer term: create a code patch that implements the changes to the regular expression-based parsing like I've suggested.

And I guess Issue 9182 is the right place for (1).

I've played a bit the idea that barthard sketched.  I don't have all the details worked out, but I believe this is what will happen:

With

  parser = argparse.ArgumentParser()
  parser.add_argument('-w')
  parser.add_argument('-x', nargs='+')
  parser.add_argument('y')
  parser.add_argument('z', nargs='*')

some possible parses are

    '-w 1 -x 2 3 4 5',     # w:1, x:[2,3,4], y:5, z:[] -
                           # fail +
    '-w 1 2 -x 3 4 5',     # w:1, y:2, x:[3 4 5], z:[] +

    '-w 1 -x 2 3',         # w:1, x:[2], y:3, z:[]     -
                           # fail +
    '-x 1 2 -w 3 4 5 6',   # w:3, x:[1,2], y:4, z:[5,6] +
                           # w:3, x:[1], y:2, z:[4,5,6] -

    '-x 1 2 3 4 -w 5 6 7', # w:5, x:[1,2,3,4], y:5, z:[7] +
                           # w:5, x:[1,2,3], y:4, z:[6,7] -

    '1 2 3 -x 4 5 -w 6',   # w:6, x:[4,5], y:1, z:[2,3] +
      
'+' lines are those currently produced
'-' lines are ones that would be produced by these ideas

'-w 1 -x 2 3 4 5' is the protypical problem case.  The current parser allocates all [2,3,4,5] to -x, leaving none for y, thus failing.  So desired solution is to give 5 to y, leaving -x with the rest.

'-x 1 2 -w 3 4 5 6' is a potentially ambiguous case.  The current parser lets -x grab [1,2]; y then gets 4, and z the remainder.  But the alternative is to give 2 to y, leaving -x with just [1].

In this case 
arg_strings_pattern = 'OAAOAAAA'

replacing the Os with the option flags: '-xAA-wAAAA'

I match this with a refined version of bethard's regex:

    pat1='((?:-wA)|(?:-xA+)|(?:-wA-xA+)|(?:-xA+-wA))'
    pat = _re.compile('%s?(?P<y>A)%s?(?P<z>A*)%s?'%(pat1,pat1,pat1))

groups (without the Nones) and groupdict are

    ['-xA', 'A', '-wA', 'AAA']
    {'z': 'AAA', 'y': 'A'}

So this does effectively give y the 2nd argument, leaving -x with just the 1st.

The current parser effectively groups the arguments as

    ['-xAA, '-wA', 'A', 'AA']

In the real world, generating and apply a global pattern like this could get complicated.  For example there are long option names ('--water'), and combined argument strings ('-x1', '-x=1').

I need to make one correction to my last post:

    '-x 1 2 -w 3 4 5 6',   # w:3, x:[1,2], y:4, z:[5,6] +
                           # w:3, x:[1], y:2, z:[4,5,6] -

The second solution is only possible if 'z' is not consumed when 'y' is being processed.  In current version, if consume_positionals() is called with a 'AOAAAA' pattern, 'y' will match the first 'A', and 'z' will match ''.  That means '4 5 6' will be left over.

It's only when I use the patch in http://bugs.python.org/issue14191#msg187051
(argparse doesn't allow optionals within positionals)
that the processing 'z' is delayed, so it can get [4,5,6].

So at least with the 4 arguments in this example, bethard's idea only seems to make a difference in the case of '-w 1 -x 2 3 4 5', where 'y' lays claim to the last string, and '-x' gets the rest.

This patch implements, I think, the ideas bethard proposed.  It is test patch, not intended for production.

Most of work is in ArgumentParser._get_alt_length() which
- generates a pattern along the lines bethard proposed
- generates a string like arg_strings_pattern, but with optionals strings ('-x') instead of 'O'.
- runs a match
- from groups like '-xAAA', creates dict entries like: 
    alt_opt_length['x'] = 3

Later, in consume_optionals(), this alternative count replaces arg_count if it is lower.  The next consume_positionals() then takes care of consuming the unconsumed arguments.

If _get_alt_length() has any problems, it logs an error, and returns an otherwise empty dict.  So it 'fails' quietly without affecting regular parsing.

Reasons for failing include (for now) the use of subparsers, optionals with explicit args, and special prefix_chars.  With exclusions like this, test_argparse.py runs without errors or failures.

Since this is still a testing vehicle, it writes an issue9338.log file with debugging entries.

This version works, but is both not sufficiently general and too general.  As bethard notes, the testing pattern could get very large if there are many optionals.  Ideally the pattern will allow the optionals in any order and combination between positionals.  The ambiguities that I discussed in the previous 2 posts disappear if the patching pattern is sufficiently general.

But I also suspect it is too general.  It does not need to match every case, just those where an optional is consuming arguments that should go to a positional.  But if we come up with something more specific, this could still be a useful testing tool.

This is a test file for the patch I just submitted.

It is not a formal unitttest, but uses print output as much as assert.

Cases include the example bethard used, as well as ones from test_argparse.py that initially caused problems.

Here's another approach to the problem, using an iterative localized search. For simple cases it produces the same thing, but in complex cases it is more robust.

It is based on two ideas:

- if the action in consume_optional() is being 'greedy', use

    slots = self._match_arguments_partial([action]+positionals, selected_patterns)

  to determine if this action can share arguments with any of the remaining positionals.  This is similar to how consume_positionals() allocates arguments to the set of positionals.

- try this 'sharing' with the last optional.  If that is not enough, try the penultimate optional as well, and continue working toward the start as needed.

Since the normal parsing is from left to right, figuring out when to start 'sharing' requires some sort of search strategy.  I have moved the start_index loop into a consume_loop() function, and added a switch so it can parse the arguments without invoking take_action() (so arguments are not evaluated, and namespace is not changed).

If there is a suspected 'greed' problem, consume_loop() is called (in test mode) one or more times to determine the right-most optionals to use, and once more (with take_action) to parse and evaluate the arguments.

As in the previous patch this writes a log file for debugging purposes. test_argparse.py now has a number of tests for this issue.  

It is more robust than the previous patch, and does not need special handling for things like subparsers and explicit arguments.

Oops, I attached the wrong file.  Here's the correct one.

I ran into this bug the first time I needed nargs + in a tool.  

I found of course that if the option with the nargs is followed by another option before the positional arguments it will work as expected.  But then the help would have to point this out, and it still could be used incorrectly (so then I get a mail about a bug in my tool.) ;-)>

My workaround was to use action=append instead of nargs, then user would just have to give the option for each nargs desired.  Since my use would be short this was OK.  But the usage message does not reflect the multiple use nature of this option

But what I expected to find in the doc was a way to specify the use of a separator char between the nargs option arguments.  For example specify that ',' is the separator arg (currently a space is the separator.)  So if option is -foo the cli could be:

myprog.py -foo bar1,bar2,bar3 pos1 pos2

(Of course I could just have the tool take a comma delimited single argument and parse it in the tool's logic, but again then a custom usage message would be needed.)

Has this solution been considered?

parse_args() would see ['-foo', 'bar1,bar2,bar3', 'pos1', 'pos2'].  The splitting on space is done by the shell.  So having your own code split 'bar1,bar2,bar3' is simplest.  But that would be messed up if the user entered 'bar1, bar2, bar3...'.  You could also ask the user to use quotes - "bar1, bar2, bar3".

http://bugs.python.org/issue15112
breaks one test that I added to issue

    +class TestPositionalsAfterOptionalsPlus(ParserTestCase):
    +    """Tests specifying a positional that follows an arg with nargs=+
    +    http://bugs.python.org/issue9338#msg111270
    +    prototypical problem"""
    +
    +    argument_signatures = [
    +        Sig('-w'),
    +        Sig('-x', nargs='+'),
    +        Sig('y', type=int),
    +        Sig('z', nargs='*', type=int)]
    +    failures = ['1 -x 2 3 -w 4 5 6' # error: unrecognized arguments: 5 6
    +                # z consumed in 1st argument group '1'
    +    ]

This no longer fails.  Due to 15112, z=[5,6].  That is, it is no longer consumed by the 1st argument group.

Recent StackOverFlow question related to this issue - where the following positional is a subparsers.


http://stackoverflow.com/questions/41742205/how-to-argparse-with-nargs-and-subcommands

Is there any chance this will ever get fixed?  Patches have been available for 5 years with no progress.

I'm a newcomer and thought about trying to follow up on this and potentially update existing patches, would this be a good idea?

Paul, do you have any thoughts on this one.

ISTM that this issue isn't well suited for a newcomer because it's somewhat complex and it isn't clear what if anything should be done.

Reproduced on 3.11:

>>> import argparse
>>> parser = argparse.ArgumentParser(prog='PROG')
>>> 
>>> parser.add_argument('--badger', nargs='+')
_StoreAction(option_strings=['--badger'], dest='badger', nargs='+', const=None, default=None, type=None, choices=None, help=None, metavar=None)
>>> parser.add_argument('spam')
_StoreAction(option_strings=[], dest='spam', nargs=None, const=None, default=None, type=None, choices=None, help=None, metavar=None)
>>> parser.parse_args('--badger A B C D'.split())
usage: PROG [-h] [--badger BADGER [BADGER ...]] spam
PROG: error: the following arguments are required: spam