Re: Can we do better than "git checkout/add -p"
Patrick Steinhardt <[email protected]>
| Newsgroups | org.kernel.vger.git |
|---|---|
| Message-ID | <[email protected]> |
On Fri, Aug 07, 2026 at 08:50:39AM -0700, Junio C Hamano wrote: > "D. Ben Knoble" <[email protected]> writes: > > > I raise this as the kind of interface we could learn from: emulating > > it might be a bit heavier (a full TUI?), but is certainly more > > convenient to use than the prompt-loop over hunks. > > Yes, the 'one hunk at a time' model was easy to implement and start > using, but its limitations are apparent. Users want to be able to > jump around, starting in the middle and returning to the top later, > for example. > > It is more or less orthogonal to the reason I started this > discussion, though, which is that limiting the direction in which > modifications flow restricts the workflow, burdens the user, and > makes the process error-prone. > > When I see a hunk, I can immediately tell if it is one of three > kinds (i.e., those we want to add, those we want to leave in the > working tree, and those we want to discard from the working tree). > But with 'git add -p' (especially with the original version of the > feature, before the 'e' (edit) command was introduced), the third > kind must be treated the same way as the second. Then, after I am > done with 'git add -p', I must go through the remaining hunks, sift > them into two categories (those we want to keep in the working tree > and those we want to discard), and run 'git checkout -p' to deal > with the latter. > > We should be able to improve this workflow without deviating from > the 'one hunk at a time' model. I wonder whether we can take JJ as inspiration. For commands like jj-split(1) it has the ability to interactively select specific hunks via `jj split --interactive`, too. But instead of looping through stuff, it uses a full TUI that: - Gives you a list of all files that have changed. On this level you can select/deselect the complete file. - Also allows you to expand files and then select/deselect individual hunks and lines. I found that model to be quite a bit superior to Git's own interactive mode. I've been playing around with the thought of introducing ncurses-based interfaces into Git. I've been mostly thinking about git-history(1) here so that you can just move commits around, squash them together, drop them and so on. But I think fancy stuff like TUIs can also be applied to other parts of Git, as well, to make things a bit more visual to our users and, as a consequence, easier to use. Patrick