git add --interactive "Your edited hunk does not apply"

Question

I'm trying to use git add --interactive to selectively add some changes to my index, but I continually receive the "Your edited hunk does not apply. Edit again..." message. I get this message even if I choose the e option, and immediately save/close my editor. In other words, without editing the hunk at all, the patch doesn't apply.

Here's the exact example I'm using (I'm trying to put together a small demo):

Original file:

first change
second change off branch
third change off branch
second change
third change
fourth change

New file:

Change supporting feature 1
first change
second change off branch
third change off branch
second change
third change
fourth change
bug fix 1
change supporting feature 1

I'm trying to show how to use git add --interactive to only add the "bug fix 1" line to the index. Running interactive add on the file, I choose the patch mode. It presents me with

diff --git a/newfile b/newfile
index 6d501a3..8b81ae9 100644
--- a/newfile
+++ b/newfile
@@ -1,6 +1,9 @@
+Change supporting feature 1
 first change
 second change off branch
 third change off branch
 second change
 third change
 fourth change
+bug fix 1
+change supporting feature 1

I respond with split, followed by "no" to apply the first hunk. The second hunk, I try to edit. I originally tried deleting the bottom line - that didn't work. Leaving the hunk alone completely doesn't work either, and I can't figure out why.

Answer 1

Is this like in this git-add post?

Manually editing the hunk is immensely powerful, but also a bit complicated if you've never done it before.
The most important thing to keep in mind: The diff is always indented with one character in addition to whatever other indentation is there.
The character can either be:

• a space (indicates an unchanged line),
• a - indicating that the line was removed,
• or a + indicating that the line was added.

Nothing else. It must be either a space, a - or a +. Anything else, and you'll get errors
(there's no character for a changed line, since those are handled by removing the old line, and adding the changed one as new).

Since you've got the diff open in your favorite text editor (you did configure Git to use your favorite text editor, right?), you can do whatever you want - as long as you make sure the resulting diff applies cleanly.

And therein lies the trick. If you've never done this before, Git will tell you "Your edited hunk does not apply. Edit again?" so often, you'll start to hate yourself for your inability to figure this out, even though it seems so easy (or Git because it can't figure out what you want).

One thing that tripped me up quite often was that I forgot the one character indent.
I'd mark a line with a - to be removed, but in most text editors that inserts a -, it doesn't overwrite the space that was there before. This means you're adding an additional space to the whole line, which in turn means the diff algorithm can't find/match the line in the original file, which in turn means Git will yell at you.

The other thing is that the diff still has to make sense. "Sense" means that it can be applied cleanly. Exactly how you create a sensible diff seems to be a bit of an dark art (at least to me right now), but you should always keep in mind how the original file looked like, and then plan your -s and +s accordingly. If you edit your hunks often enough you'll eventually get the hang of it.

See also this commit on git add -p.

Ortomala Lokni's answer refers to Joaquín Windmüller blog post "Selectively select changes to commit with git (or Imma edit your hunk)"

Instead of counting lines, what Git would like to do is to coalesce overlapping hunks (when one is edited) before applying said edited hunk.
That was discussed mid-2018, and would avoid scenario like:

if you split a hunk, edit the first subhunk, transforming a trailing context line to a deletion then if you try to stage the second subhunk, it will fail.

Answer 2

Of course I'm late to this but nevertheless wanted to mention for the record that this issue was discussed last year on the git mailing list and it looks like not much has changed since.

This particular issue stems from splitting and attempting to edit the same hunk. The analysis, as originally posted by Jeff King, of the underlying problem essentially is:

Hm. OK, I see. The "does this diff apply" check feeds both parts of the split patch to git-apply. But of course the second part will never correctly apply, because its context overlaps with the first part, but doesn't take it into account.

Doing the check with just the edited patch would work. But that doesn't take into account that your edited patch will potentially fail to apply in the long run, depending on whether or not you accept the other half of the split patch. And we can't know that yet, because the user may not have told us (they could have skipped the first half, and then come back to it later after the edit step).

Jeff concludes his post with a very pragmatic workaround that always succeeds, and is thus highly recommended:

So in general, I think splitting and editing the same hunk is inherently dangerous and is going to lead to these sorts of problems. And because editing provides a superset of the functionality, I think you should just edit and either allow the first part of the hunk to be applied or not depending on your preference.

By only choosing to edit a hunk not previously split, you will not have to deal with the line numbers.

Answer 3

For this particular example, you need to tweak the line numbers in the hunk. Change the line:

@@ -1,6 +2,8 @@

so that it instead reads:

@@ -2,7 +2,8 @@