The issue is that customopt (and all its children) assume the value type is already the correct one, and thus do not perform any conversion. Since we're parsing values from the config file, the conversion is desired to ensure they don't all end up as text - the config{bool,int,etc} methods called by configtyped perform the proper conversion. In most cases (all commands that declare default values) no conversio is needed since those already have the correct type.

You're right, I had the plain logic inverted in my head. Removed.

IIUC, an extension author may implement its own customopt subclasses, and
put them into the command table, so we can't make ui.configtyped to
support all of them.

Ah, makes sense. See if this addresses that case satisfactorily - it still has the caveat of not being able to "reset" container types, but that's true of the command line as well (if you have a list flag with a non-empty default, there's no way to remove that default item).

I think rdamazio's response is correct. The type information for an opt exists only in the 'defaultvalue' slot in the tuple, so it needs to be extracted from there before reading the config. I don't have a strong opinion as to where this code should go - 'ui' does seem marginally more appropriate than 'fancyopts', since it keeps 'fancyopts' low-level and not depending on 'ui'.

This doesn't handle callables properly. I wonder if the something like the following would work instead:

oldopt = fancyopts._defaultopt(olddefault)
newdefault = old.opt.newstate(olddefault, ui.config("commands", cfgitem)
c[idx] = (opt[0], opt[1], fancyopts._withnewdefault(oldopt, newdefault), opt[3])

Where '_withnewdefault' is a wrapper customopt that just changes the default.

This makes me nervous. What if someone re-uses a customopt instance in multiple commands? i.e.:

DATE_FLAG = mypkg.dateopt()
...
('b', 'before', DATE_FLAG, '')
('a', 'after', 'DATE_FLAG', '')

Now, setting commands.defaults.before=2018-03-05 also silently changes the default for 'after'. I suspect we need to introduce a wrapper class like what I suggest on lines 625-625, that delegates and leaves the original default unchanged. And either way, we should probably clarify in the docs on customopts what expected use of the class is (i.e., should we just forbid reuse, is 'oldstate' safe to mutate, etc.)

I thought callables were meant to be used to generate the default default, not with overridden values?

Nobody should use the same *instance* on multiple flags. Even with the current flags, if you use e.g. the same list on many, that'll cause problems with listopt.