On Mon, Sep 23, 2024 at 10:07:08AM +0100, phillip.wood123@gmail.com wrote:

Thanks for the re-roll. I'm still not convinced that changing this without
keeping an easy way to get the current behavior is a good idea.

Thank you for maintaining an open attitude and helping make the change
reasonable.

On 18/09/2024 18:51, Rubén Justo wrote:

quoted

The "edit" option allows the user to directly modify the hunk to be
applied.

If the modified hunk returned by the user is not an applicable patch,
they will be given the opportunity to try again.

For this new attempt we give them the original hunk;  they have to
repeat the modification from scratch.

Instead, let's give them the modified patch back, so they can identify
and fix the problem.

It's still not clear how an inexperienced user is meant to do that.

If inexperience refers to the user not being familiar with the patch
format, I don't think it's something we should worry about.  It is
very likely that few users will attempt to (e)dit, and those who do,
inexperienced users experimenting, the normal (and sensible) path they
will follow at the slightest difficulty will probably be to abandon
patch editing and try a more accessible option, such as: cancel the
`add -p` session, edit normally, and restart the session.

If inexperience refers to users unfamiliar with the (e)dit option but
knowledgeable about the patch format, I believe they will have enough
experience to know that reconstructing a patch is a tricky task.
Again, the most likely path they will follow when encountering
difficulties will be to restart the edit (Junio's message in the
thread is an example of this).

At any rate, although I understand the concern, this series doesn't
aim to improve the mechanisms that help identify the problems in a
faulty patch.  The goal of this series is to offer (actually to
recover) the possibility for the user to make corrections.

Something I haven't mentioned before in this thread is that currently,
if the user makes a mistake while editing a patch, we don't give them
the opportunity to review their error.  We leave them in doubt.  This
happened recently to me editing a patch with context lines containing
whitespace errors and "whitespace=fix".  But that's another story that
I still need to work on [1].

So, to me, it seems sensible to let the user review the faulty patch,
even if it's only to discard it.

quoted

If they really want to start over with a fresh patch they still can
say "no" to cancel the "edit" and start anew [*].

This is not very obvious to the user,

It has been so for a decade...

Keep in mind that this message will probably only be shown very _very_
rarely to users who are most likely very familiar with (e)dit.

it would be much better to give them
the choice when we prompt them about editing the hunk again.

That's an option I explored but abandoned.

I didn't come up with any message that I liked that wasn't literally a
long paragraph.

In the end, I gave up in favor of what I believe is a better option,
which is recovering the original intention of "no".

I'm not against the option you propose, I'm just not convinced that
what we already have, since ac083c47ea (git-add--interactive: manual
hunk editing mode, 2008-07-03), isn't intuitive enough for the
users of (e)dit.

We've been
giving the user the original hunk for the last six and a half years so I
think it's a bit late to unilaterally change that now.

For me, this isn't a reason not to make the change.

quoted

diff --git a/add-patch.c b/add-patch.c
index 557903310d..75b5129281 100644
--- a/add-patch.c
+++ b/add-patch.c

@@ -1111,7 +1111,8 @@ static void recolor_hunk(struct add_p_state *s, struct hunk *hunk)
  	hunk->colored_end = s->colored.len;
  }
-static int edit_hunk_manually(struct add_p_state *s, struct hunk *hunk)
+static int edit_hunk_manually(struct add_p_state *s, struct hunk *hunk,

I would add
				const struct hunk *backup,

here

quoted

+			      size_t plain_len, size_t colored_len)
  {
  	size_t i;

@@ -1146,6 +1147,10 @@ static int edit_hunk_manually(struct add_p_state *s, struct hunk *hunk)
  				      "addp-hunk-edit.diff", NULL) < 0)
  		return -1;
+	/* Drop possible previous edits */
+	strbuf_setlen(&s->plain, plain_len);
+	strbuf_setlen(&s->colored, colored_len);

then we can restore the back up here with

	*hunk = *backup;

That would make it clear that we're resetting the hunk and would continue to
work if we change struct hunk in the future.

quoted

  	/* strip out commented lines */
  	hunk->start = s->plain.len;
  	for (i = 0; i < s->buf.len; ) {

@@ -1157,12 +1162,13 @@ static int edit_hunk_manually(struct add_p_state *s, struct hunk *hunk)
  	}
  	hunk->end = s->plain.len;
+
+	recolor_hunk(s, hunk);
+

This means we're now forking an external process when there is no hunk to
color. It would be better to avoid that by leaving this code where it was
and restoring the backup hunk above.

I don't see that external process. ¿?

After reviewing the code in the previous iteration, based on your
comments, I concluded that `recolor_hunk()` makes more sense before
the "return 0".  Even if we end up discarding this series.  I think.

quoted

  	if (hunk->end == hunk->start)
  		/* The user aborted editing by deleting everything */
  		return 0;
-	recolor_hunk(s, hunk);
-

  		/*
  		 * TRANSLATORS: do not translate [y/n]

@@ -1289,8 +1290,14 @@ static int edit_hunk_loop(struct add_p_state *s,
  					"Edit again (saying \"no\" discards!) "
  					"[y/n]? "));

I think we should make this a three-way choice so the user can choose to
keep their changes or start from a valid hunk.

As I said above, I don't object to this, but I'm not convinced it's
necessary.

quoted

  		if (res < 1)
-			return -1;
+			break;
  	}
+
+	/* Drop a possible edit */
+	strbuf_setlen(&s->plain, plain_len);
+	strbuf_setlen(&s->colored, colored_len);
+	*hunk = backup;
+	return -1;
  }
  static int apply_for_checkout(struct add_p_state *s, struct strbuf *diff,

diff --git a/t/t3701-add-interactive.sh b/t/t3701-add-interactive.sh
index 718438ffc7..f3206a317b 100755
--- a/t/t3701-add-interactive.sh
+++ b/t/t3701-add-interactive.sh

@@ -165,6 +165,19 @@ test_expect_success 'dummy edit works' '
  	diff_cmp expected diff
  '
+test_expect_success 'editing again works' '
+	git reset &&
+	write_script "fake_editor.sh" <<-\EOF &&
+	grep been-here "$1" >output
+	echo been-here >"$1"
+	EOF
+	(
+		test_set_editor "$(pwd)/fake_editor.sh" &&
+		test_write_lines e y | GIT_TRACE=1 git add -p

This is still missing "n q". Apart from that the test is looking good.

I've been resisting the idea of "completeness", because I think "e y"
should also be fine.  But I'm not going to resist anymore here :-),
since I don't think the test has much more value without "n q".  So
I'll add it.