add --interactive: allow `update` to stage deleted files

From: Johannes Schindelin <[email protected]>

The scripted version of git add -i used git update-index --add --remove, but the built-in version implemented only the --add part.

This fixes https://github.com/msys2/MSYS2-packages/issues/3066

---
add --interactive: fix bug in built-in variant

This fixes a bug when using the built-in version of git add -i to update a file that has been deleted (in order to stage its deletion), where it fails with:

fatal: unable to stat 'myfile': No such file or directory

Since the built-in version of git add -i has been made the default in v2.37.0, from the users' point of view this is a regression, and this patch fixes it. I therefore consider this v2.37.1 material.

This addresses https://github.com/msys2/MSYS2-packages/issues/3066

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1273%2Fdscho%2Fadd-i-update-deleted-file-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1273/dscho/add-i-update-deleted-file-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1273

 add-interactive.c          | 12 ++++++++++--
 t/t3701-add-interactive.sh |  9 +++++++++
 2 files changed, 19 insertions(+), 2 deletions(-)

index 6047e8f6489..22fcd3412ca 100644
@@ -697,8 +697,16 @@ static int run_update(struct add_i_state *s, const struct pathspec *ps,
 
 	for (i = 0; i < files->items.nr; i++) {
 		const char *name = files->items.items[i].string;
-		if (files->selected[i] &&
-		    add_file_to_index(s->r->index, name, 0) < 0) {
+		struct stat st;
+
+		if (!files->selected[i])
+			continue;
+		if (lstat(name, &st) && is_missing_file_error(errno)) {
+			if (remove_file_from_index(s->r->index, name) < 0) {
+				res = error(_("could not stage '%s'"), name);
+				break;
+			}
+		} else if (add_file_to_index(s->r->index, name, 0) < 0) {
 			res = error(_("could not stage '%s'"), name);
 			break;
 		}

index fc26cb8bae8..b354fb39de8 100755
@@ -103,6 +103,15 @@ test_expect_success 'status works (commit)' '
 	grep "+1/-0 *+2/-0 file" output
 '
 
+test_expect_success 'update can stage deletions' '
+	>to-delete &&
+	git add to-delete &&
+	rm to-delete &&
+	test_write_lines u t "" | git add -i &&
+	git ls-files to-delete >output &&
+	test_must_be_empty output
+'
+
 test_expect_success 'setup expected' '
 	cat >expected <<-\EOF
 	index 180b47c..b6f2c08 100644

base-commit: e4a4b31577c7419497ac30cebe30d755b97752c5

gitgitgadget

"Johannes Schindelin via GitGitGadget" <[email protected]>

I had the same thought, but I think that it points to something I've suspected privately for a while which is that we don't actually get many users to opt-in to features early when they are hidden behind a configuration.

I had hoped that things like feature.experimental would cause more users to try features that weren't quite ready for prime-time more often, and that we'd get more and better feedback as a result.

But I think that this proves that is generally not so. I think we should probably err on the side of enabling new features by default so long as there is sufficiently low risk, rather than hide them behind config. Or at least, not hide them behind a config variable for so long (though I am guilty of this myself with the pack.writeReverseIndex, which I have been meaning to flip the default on for a little while now).

In either case...

Yes, indeed. This fix looks good to me.

Thanks,
Taylor

I think we could get quite far by piggy-backing on the advice we emit when you haven't set user.{email,name}, or otherwise find some minimally annoying way to ask users to opt-in.

But while an opt-out being the default isn't quite the same as just changing the code (as you can still toggle it off without downgrading), that also has the disadvantage that we'd be more conservative about adding such features, which is partially why we have feature.* in the first place.

The X-Y problem of how we get early code to users who'd be able to tell us that it breaks has been discussed several times, and all the solutions are trade-offs.

One thing I thought of now but I don't think has been brought up before:
I wonder if doing "more stable" (or "LTS"?) releases would be a good compromise between "early bird" and long-term stability, where we'd just make every Nth release (say those divisible by 5) have feature.experimental=false.

But any such scheme also quickly runs into the problem that we're just not adding a lot of these, and for e.g. the built-in conversion we might do the whole feature cycle in 3-4 releases, so such "stable" releases would force us to keep the code around longer.