Re: Additional plumbing commands

From: Junio C Hamano <gitster@pobox.com>
Date: Jan 7, 2015, 16:31 UTC
Message-ID: <xmqqa91ulfhz.fsf@gitster.dls.corp.google.com>
In-Reply-To: <20150107080408.GA1301@peff.net>

Jeff King <peff@peff.net> writes:

Show 8 quoted lines

> ... Knowing what each state _means_
> and what operations are meaningful to perform is much trickier (e.g., if
> we are in a rebase, you probably do not want to start a new rebase. But
> is it wrong to cherry-pick?).
>
> It would be nice if we could find a way to generalize in-progress
> operations and what they mean for starting new operations, but that is
> a much harder problem (if it is even possible at all).

Very good points. Thinking aloud, to see if we can start from a few rules of thumb.

You can be "in the middle" for two largely different reasons.

One is that the command inherently wants to give control back to you. Think of a case where you said "exec false" in the "rebase -i" insn sheet, or "bisect" checked out a version for you to try.

The other is that the command wanted to continue in the ideal world, but couldn't and stopped asking for your help. Think of a case where "am" stopped due to corrupt patch, "cherry-pick A..B" or "revert" stopped due to conflicts.

In the former case, depending on the nature of the command, what are sensible things you can do are very different (during "bisect" you would typically not want to do anything that causes a commit created. During "rebase -i" you may want to run any command that creates a commit, to insert a new one into the series). But a good rule of thumb would be "If I should be able to edit the working tree file, I should also be able to do "cherry-pick --no-commit", "merge --no-commit", "apply", etc." "If I should be able to manually commit, I should also be able to "cherry-pick", "merge", etc.

In the latter, the _only_ reason you are given control back is to help the interrupted operation. So "cherry-pick --no-commit" or "apply" in lieu of editing files manually in order to fix conflicts should be allowed, but a random "git merge" (without --no-commit) would not be.

So after thinking aloud for a while, I very much agree with you that you cannot say "X is allowed but not Y" in many situations.

One thing we can say for sure is that in a middle of a multi-step operation (e.g. rebase, range pick) is stopped for either one of the two reasons, you cannot do another multi-step operation. This is not fundamental but a consequence of how the sequencing machinery is implemented. But other than that, it really is case-by-case and not even command-by-command.