On 18/02/16 21:26, Leo Famulari wrote:
> On Thu, Feb 18, 2016 at 09:01:00PM +0100, Andreas Enge wrote:
>> On Wed, Feb 17, 2016 at 07:11:44PM -0500, Leo Famulari wrote:
>>> I missed this earlier. We don't usually propagate this sort of "end-user
>>> program" dependencies. For example rsync does not propagate ssh; the
>>> user must install ssh. [0]
>>
>> I think the cleanest approach would be to patch all calls to "git" in
>> the vcsh shell script by calls to "/gnu/store/xxxx-git-.../bin/git",
>> where the base name is taken as "(assoc-ref %build-inputs "git")" or the like.
>>
>> What do you think?
> 
> Indeed, this sounds like a clean solution. In that case, git would be a
> plain input.
> 
>>
>> It might be as simple as a call in a phase to "substitute*", replacing
>> "git " by the path+" ". The script is a bit annoying in that it scatters
>> calls to git all over the place. It would be cleaner to define "GIT=git"
>> at the start of the file, and then always call "$(GIT)". Maybe that is
>> something to suggest to the upstream author?
> 
> Since 'git' is such a short string, the possibility of "mis-patching"
> when performing the string substitution seems a little too high. What do
> you think?

A patch could be written as Andreas describes (to use a variable), and
then this could be used for Guix. Once written, the patch could also be
sent to the upstream maintainer to see if they want to include it.