Re: [PATCH/RFC] commit: new option to abort -a something is already staged

From: Jonathan Nieder <jrnieder@gmail.com>
Date: Aug 20, 2018, 19:30 UTC
Message-ID: <20180820193007.GB31020@aiede.svl.corp.google.com>
In-Reply-To: <20180820154120.19297-1-pclouds@gmail.com>

Hi,

Nguyễn Thái Ngọc Duy wrote:

Show 7 quoted lines

> So many times I have carefully cherry picked changes to the index with
> `git add -p` then accidentally did `git commit -am ....` (usually by
> retrieving a command from history and pressing Enter too quickly)
> which destroyed beautiful index.
>
> One way to deal with this is some form of `git undo` that allows me to
> retrieve the old index.

Yes, I would love such an undo feature!

How would you imagine that this information would get stored? We can start with adding that and a low-level (plumbing) interface to it, to avoid being blocked on getting the user-facing (porcelain) "git undo" interface right. (Or we can go straight for the porcelain. It's fine for it to start minimal and gain switches later.)

[...]

Show 5 quoted lines

> Instead, let's handle just this problem for now. This new option
> commit.preciousDirtyIndex, if set to false, will not allow `commit -a`
> to continue if the final index is different from the existing one. I
> don't think this can be achieved with hooks because the hooks don't
> know about "-a" or different commit modes.

I frequently use "git commit -a" this way intentionally, so I would be unlikely to turn this config on. That leads me to suspect it's not a good candidate for configuration:

- it's not configuration for the sake of a transition period, since some
  people would keep it on forever

- it's not configuration based on different project needs, either

So configuration doesn't feel like a good fit.

It might be that I can switch my muscle memory to "git add -u && git commit", in which case this could work as a new default without configuration (or with configuration for a transition period). A separate commandline option "git commit -a --no-clobber-index" might make sense as well, since then I could pass --clobber-index to keep my current workflow. That would also follow the usual progression: introduce commandline option first, then config, then change default.

That said, I lean toward your initial thought, that this is papering over a missing undo feature. Can you say more about how you'd imagine undo working?

Thanks, Jonathan