From: Luiz Eduardo Campos Date: Wed, 25 Mar 2026 19:22:12 GMT Subject: Re: [RFC PATCH 0/1] add -p: support discarding hunks Message-ID: In-Reply-To: Hi Junio, Thanks a lot for the detailed feedback. > "add -p" that touches the working tree feels quite weird. I understand the concern, and I agree that having `git add -p` perform destructive changes on the working tree is a significant departure from its current mental model. > people trigger "discard" by mistake ... no way to undo This is a very good point. I had been thinking of this primarily as a convenience for workflows where users repeatedly skip hunks, but you are right that making such an operation easy to trigger in an interactive session could lead to accidental data loss, and that would be problematic. > What you want is probably a single command ... This is a very interesting direction. My original motivation was exactly to avoid having to switch between `git add -p`, `git restore -p`, and `git stash -p` when reviewing changes, but I had not considered approaching it as a single interface with modes over the different views (HEAD, index, worktree) and non-destructive flows like stashing. That does seem like a more coherent model. In the meantime I will rely on the workflows you mentioned (`stash -p` or `restore -p` followed by `git add -p`) rather than pursuing discard within `add -p` alone. I will step back from the narrow "discard in add -p" RFC while I think more about this broader direction. One question I would like guidance on is whether something like what you describe would be more appropriate as a new top-level command, or as an extension of an existing one (and if so, which entry point would be the least surprising). If a unified tool ever includes a destructive "revert hunk in the worktree" mode, I agree it would need to be very clearly separated and hard to trigger by mistake. Thanks again for the guidance; it is very helpful. Luiz Em qua., 25 de mar. de 2026 às 15:03, Junio C Hamano escreveu: > > Luiz Campos writes: > > > Hi, > > > > This is an RFC for adding a 'discard hunk' action to `git add -p`. > > > > Currently, when using `git add -p`, users can stage or skip hunks, > > but cannot discard unwanted changes directly from the working tree. > > This often leads to repeatedly skipping the same hunks across > > multiple passes. > > > > This patch introduces a new 'x' action to discard the current hunk > > by reverse-applying it to the working tree. > > > > This idea was previously discussed on the mailing list: > > https://lore.kernel.org/git/X%2FiFCo0bXLR%2BLZXs@coredump.intra.peff.net/t/#m0576e6f3c6375e11cc4693b9dca3c1fc57baadd0 > > > > Open questions: > > - Should discard happen immediately or be deferred until patch application? > > - Are there edge cases involving overlapping hunks or edited hunks? > > After reading the discussion (by the way, I do not recall seeing it, > so thank you very much for having a link to it), I agree with what > Peff said back then. "add -p" that touches the working tree feels > quite weird. > > In addition to that, letting it make destructive change makes the > idea even less appetizing. Once you remove the changes introduced > by the hunk, it is forever gone. A "discard" in "add -p" would not > solve your problem without adding many unhappy users who lost their > work by mistake. I do not want to see people trigger "discard" by > mistake in "add -p" session _and_ find that there is no way to undo > that mistaken discard. > > "stash -p" followed by "add -p" is probably the best we can do that > is safe. When the unwanted change is truly unwanted garbage that > you would never ever want to see again, "restore -p" followed by > "add -p" would be an alternative. > > One reason why they are not satisfying is because during the later > "add -p" session, we will notice that some unwanted things we failed > to notice and get rid of (either by sending them to stash or restoring > it away) are still there, reminding us that we are imperfect human, > and at that point, it is not easy to switch back to the "stash -p" > or "restore -p" from there. > > What you want is probably a _single_ command that lets you inspect > the differences among the HEAD, the index, and the working tree, and > allows you to move things hunk-by-hunk in different directions. > > * You can go through the "git diff --cached" (i.e., changes already in > the index), and selectively undo/revert the changes to the index, > similar to "git reset -p". > > * You can go through the "git diff" (i.e., changes between the > index and the working tree), and selectively apply the changes to > the index, similar to "git add -p". > > * You can go through the "git diff HEAD" (i.e. changes since your > last commit), and selectively send the changes to a stash entry, > similar to "git stash -p". This is not destructive. > > And if the single command lets you switch among working with these > modes, you no longer need to worry about forgetting to send > some changes to stash to concentrate on working on the rest. > > In addition, optionally you can also have this in the same command: > > * You can go through the "git diff", and selectively revert the > changes to the working tree, similar to "git restore -p". > > This additional mode *is* destructive, but if you know from the hunk > that you will never need the change in it, it would be a right tool > for it.