On Fri, Oct 18, 2019 at 05:01:42PM -0400, Eric Sunshine wrote:

On Fri, Oct 18, 2019 at 10:37 AM SZEDER Gábor [off-list ref] wrote:

quoted

On Thu, Oct 17, 2019 at 01:52:27PM -0400, Eric Sunshine wrote:

quoted

quoted

+               case "$1" in
+               --show-idx)     show_idx=y ;;
+               *)              return 1 ;;

Should this emit an error message to aid a person debugging a test
which fails on a call to __git_find_on_cmdline()? [...]

And printing anything to standard error during completion is
inherently bad: it disrupts the command line, can't be deleted [...]
Remaining silent about the unrecognized option
is in my opinion better, because then the completion script usually
does nothing, and Bash falls back to filename completion.  Yeah,
that's not ideal, but at least the user can easily correct it and
finish entering the command.

I had tunnel-vision and was thinking about this only in the context of
the tests. However, while I agree that spewing errors during
completion is not ideal, aren't there two different classes of errors
to consider? Some errors can crop up via normal usage of Git commands
in Real World situations; those errors should be suppressed since they
are expected and can be tolerated. However, the second class of error
(such as passing a bogus option to this internal function) is an
outright programming mistake by a maintainer of the completion script
itself, and it would be helpful to let the programmer know as early as
possible about the mistake.

Or, are there backward-compatibility or other concerns which would
make emitting error messages undesirable even for outright programmer
mistakes?

It's not necessarily an outright programming mistake, and that error
could be triggered by ordinary users as well.

Let's suppose that a user has a custom 'git-foo' command in $PATH with
a custom '_git_foo' completion function in '~/.my-git-completions',
which the helper function '__git_bar --option' from our completion
script.  Let's also suppose that the user sources this completion
function from '~/.bashrc', but otherwise uses the system-wide git
completion script, and that $HOME is shared across multiple computers.

In this (arguably somewhat convoluted) scenario it might happen that
on a not quote up-to-date computer the system-wide git completion
script already has the '__git_bar' helper function, but it doesn't yet
support '--option'.  If '__git_bar' then prints an error to stderr,
then the command line will get badly messed up, and the user will have
to ctrl-C and start over.

However, if '__git_bar' silently ignores the unknown option, then the
worst that can happen is that completion doesn't work, and e.g. it
falls back to Bash's filename completion or offers something
nonsensical.  In either case, after a brief "Huh?!" moment the user
can correct it by hitting backspace a couple of times and then enter
the rest of the command by hand.