Re: [PATCH] config: add an includeIf.env{Exists,Bool,Is,Match}

["Ævar Arnfjörð Bjarmason" <avarab@gmail.com>]

On Mon, Sep 27 2021, Jeff King wrote:

> On Tue, Sep 28, 2021 at 01:52:26AM +0200, Ævar Arnfjörð Bjarmason wrote:
>
>> An implicit assumption of mine in the simpler positive-match-only
>> version (which I should have made clear) is that anyone who needs this
>> sort of complexity can just arrange to wrap their "git" in a function,
>> or do this sort of thing in their ~/.bashrc, i.e. just:
>> 
>>     if code_of_arbitrary_complexity
>>     then
>>         export GIT_DO_XYZ_INCLUDES=1
>>     fi
>> 
>> Then in your config:
>> 
>>     includeIf.envBool:GIT_DO_XYZ_INCLUDES.path=~/.gitconfig.d/xyz.cfg
>> 
>> And having written that out I think the best thing to do is probably to
>> have a version that only does the envExists and envBool version (or just
>> envBool), and skip envIs and envMatch entirely.
>
> I'm not sure I agree. If you are willing to wrap git, then you can just
> add:
>
>   git -c include.path=~/.gitconfig.d/xyz.cfg
>
> to the command-line in the first place. Or if you're willing to use our
> undocumented interface, you can even do it in your .bashrc:
>
>   if code_of_arbitrary_complexity
>   then
>           GIT_CONFIG_PARAMETERS="'include.path'='~/.gitconfig.d/xyz.cfg'"
>   fi

Sort of, that'll give you unconditional inclusion, but won't e.g. handle
a case where the env include only runs in some .git/config, or depending
on other inclusion (e.g. in ~/dev/git, but only if XYZ env var).

But yeah, it won't handle all potential cases. I figured for this sort
of thing it was better to start small and see if the provided interface
was enough..

> The value of this env matching is that it is done at run-time without
> wrapping, and can meaningfully inspect the state of the world. E.g., the
> $TERM thing that started this thread.

Yeah, maybe we should have at least an ifStrEQ, whatever we call it...

>> In the case of env:PATH we're just setting users up for some buggy or
>> unexpected interaction with something that would be better done either
>> via a gitdir include, or if they really need $PATH they can just wrap
>> "git" in a function that sets a boolean inclusion variable.
>
> Yes, I have trouble imagining why any matching on env:PATH would be
> useful (or $PWD, since we have the much less confusing gitdir
> conditional). Which isn't to say I want to forbid it, but just because
> people can shoot themselves in the foot with complexity doesn't mean
> that "envIs" is a bad thing when it's not misused.

I'm biased by past on-list discussions where existing behavior, no
matter if unintentional or emergent can be really hard to fix once
established.

>> > I think it's just the mashed-up colons that I find ugly in the first
>> > one. But I agree the latter isn't that nice either, and introduces the
>> > ambiguity you describe.
>> 
>> FWIW I hacked up a --config-key --config-value pairing so you could set
>> keys with "=" in them on the command-line, I'm not sure I like the
>> interface, but it gets rid of that ":" v.s. "=" edge case:
>> https://github.com/avar/git/commit/a86053df48b
>
> Yeah, we talked about that a while ago, but nobody liked the interface
> enough to actually code it (and as far as I know, it's really
> theoretical; nobody has actually wanted to set such an option from the
> command-line yet, and we have the --config-env stuff for people who want
> to robustly pass along arbitrary keys).
>
> A perhaps more subtle but less awkward to type version is to just
> require two arguments, like:
>
>   git --config <key> <value> ...

I suppose --config would work like that, you can'd to it with "-c". I
think it's more confusing to have a "-c" and "--config" which unlike
most other things don't follow the obvious long and short option names
working the same way.

> but I'd just as soon continue to leave it un-implemented if nobody has
> actually needed it in practice.

*nod*. I do think it's bad design to introduce an "env" inclusion
feature that relies on "=" though while we don't have something like
that, i.e.

I think we should probably not add that --config-{key,value}, but
avoiding the arbitrary limitation of not being able to specify certain
config keys seems prudent in that case, and since the "=" v.s. ":" is
only an aesthetic preference I think being able to compose things
without limitations wins out.

We do have the "=" key limitation now, but I don't think it's there for
any key we currently define, except things like "url.<base>.insteadOf"
if the "<base> has a "=" in it (and maybe just that one).