Re: [PATCH 0/1] add git-splice subcommand for non-interactive branch splicing

[Junio C Hamano <gitster@pobox.com>]

Adam Spiers <git@adamspiers.org> writes:

> Therefore there is a risk that each new UI for higher-level workflows
> will end up re-implementing these mid-level operations.  This
> undesirable situation could be avoided if git itself provided those
> mid-level operations.

Let me make sure if I get your general idea right, first.

Is your aim is to give a single unified mid-layer that these other
tools can build on instead of rolling their own "cherry-pick these
ranges, then squash that in, and then merge the other one in, ..."
sequencing machinery?

If so, I think that is a very good goal.

>     # Remove commits A..B (i.e. excluding A) from the current branch.
>     git splice A..B
>     # Remove commit A from the current branch.
>     git splice A^!
>     # Remove commits A..B from the current branch, and cherry-pick
>     # commits C..D at the same point.
>     git splice A..B C..D

We need to make sure that the mid-layer tool offers a good set of
primitive operations that serve all of these other tools' needs.  I
do not know offhand if what you implemented that are illustrated by
these examples is or isn't that "good set".

Assuming that there is such a "good set of primitives" surfaced at
the UI level so that these other tools can express what they want to
perform with, I'd personally prefer to see a solution that extends
and uses the common "sequencer" machinery we have been using to
drive cherry-picks, reverts and interactive rebases that work on
multiple commits.  IOW, it would be nice to see that the only thing
"git splice A..B" does is to prepare a series of instructions in a
file, e.g. .git/sequencer/todo, just like "git cherry-pick A..B"
would, and let the sequencer machinery to handle the sequencing.

E.g. In a history like

    ---o---A---o---B---X---Y---Z   HEAD

"git splice A..B" command would write something like this:

    reset to A
    pick X
    pick Y
    pick Z

to the todo file and drive the sequencer.  As you notice, you would
need to extend the vocabulary of the sequencer a bit to allow
various things that the current users of the sequencer machinery do
not need, like resetting the HEAD to a specific commit, merging a
side branch, remembering the result of an operation, and referring
to such a commit in later operation.  For example, if you tell "git
splice" to expunge A from this sample history (I am not sure how you
express that operation in your UI):

         B---C---D
        /         \
    ---o---A---E---F---G   HEAD

it might create a "todo" list like this to rebuild the history:

    reset to A^
    pick B
    pick C
    pick D
    mark :1
    reset to A^
    pick E
    merge :1 using F's log message and conflict resolution as reference
    pick G

to result in:

         B---C---D
        /         \
    ---o-------E---F---G   HEAD

Do not pay too much attention to how the hypothetical "extended todo
instruction set" is spelled in the above illustration (e.g. I am not
advocating for multi-word command like "reset to"); these are only
to illustrate what kind of features would be needed for the job.  In
the final shape of the system, "merge" in the illustration above may
be a more succinct "merge F :1", for example (i.e. the first
parameter would name an existing merge to use as reference, the
remainder is a list of commits to be merged to the current HEAD),
just like "pick X" is a succinct way to say "cherry-pick the change
introduced by existing commit X to HEAD, reusing X's log message
and author information".

Something like that may have a place in the git-core, I would think. 

I am not sure if a bash script that calls rebase/cherry-pick/commit
manually can serve as a good "universal mid-layer" or just adding
another random command to the set of existing third-party commands
for "higher-level workflows".