From: Luiz Eduardo Campos Date: Wed, 25 Mar 2026 18:38:50 GMT Subject: Re: [RFC PATCH 1/1] add -p: support discarding hunks with 'x' Message-ID: In-Reply-To: <2a0ccbfe-3d26-4146-89ed-3b942bdc861e@gmail.com> Hi Phillip, Thanks for the thorough review. > it is rather unexpected for "git add" to modify the working copy I share that concern. My motivation was allowing stage-and-discard in a single pass, which I find myself wanting regularly, and discussion in the linked thread suggested it could live in add -p. That said, I can see the argument that git add modifying the worktree breaks the mental model. I'm open to discussing whether a separate command or a flag on git restore -p would be a better home for this — the implementation would be largely the same either way. > It would be simpler just to take USE_HUNK or DISCARD_HUNK as an argument Agreed. I'll drop both the merge_for_discard boolean and the enum reassemble_mode, and instead pass the desired hunk type directly (USE_HUNK or DISCARD_HUNK) through reassemble_patch, merge_hunks, and should_merge_hunk. That removes the indirection entirely. > There's no need to add braces here Right, the project style omits braces for single-statement blocks. I'll drop them. > I'm not sure this is an improvement — it certainly makes the patch harder to > read because you end up changing the indentation of otherwise unchanged > lines that were in the else clause. > This is unchanged code that re-indented because of the addition of > "continue" above. Good point. With the simplified should_merge_hunk() taking a hunk type, I can keep the original if / else structure in reassemble_patch() and just swap out the condition. That avoids the continue rewrite and the indentation churn entirely. > We need to tell render_hunk whether a hunk is being applied in reverse or > not so that it knows whether to apply delta to the old offset or the new offset. You're right. I was wrong to think the delta direction would be the same as for staging. For the discard patch, skipped hunks' changes are still present in the worktree, so the new_offset (worktree side) is already correct and must not be adjusted. The adjustment needs to go to old_offset instead — exactly the is_reverse = 1 behaviour. This is consistent with patch_mode_checkout_index, which does the same operation (diff-files + apply -R to worktree) and sets is_reverse = 1. I'll temporarily flip is_reverse (or pass an explicit flag to render_hunk) when assembling the discard patch, and add the split-hunk tests below to confirm the offsets come out correct. > It would be nice to avoid showing this for 'checkout -p' etc. The help text display already filters lines from help_patch_remainder against the options present in s->buf (the prompt string). Since 'x' is only added to the prompt in patch_mode_add, the "x - discard..." line will already be suppressed for checkout -p, reset -p, etc. I'll add a comment in the next version to make that clearer, and add a test that verifies checkout -p help output does not mention 'x'. > Why do we need to run "git apply --check" here? We don't — the existing staging codepath in apply_patch does not run a separate --check before applying either; it just runs git apply directly and checks the exit code. I'll do the same and remove the redundant --check pass. I'll remove the redundant check and just run git apply -R directly. > Isn't this comment still relevant? Yes, I'll restore it. I removed it by accident when restructuring the function; it still applies to the USE_HUNK scan that follows. > What's the point of this change? There is none — it's an accidental whitespace change on the alignment of the (!file_diff->hunk_nr continuation line. I'll drop it. > So 'x' is only permitted if 'e' is — what's the reason for that? The nesting was unintentional, but the conditions guarding 'e' — hunk_index + 1 > file_diff->mode_change and !file_diff->deleted — do also apply to 'x', since reverse-applying a mode-change or deletion hunk is not meaningful. However, the ADD_P_DISALLOW_EDIT flag (used by git history split) should not gate 'x'. I'll give 'x' its own block that checks only mode_change and deleted, independent of ALLOW_EDIT. > I'm not sure why this test needs to be in a separate repository The sub-repo was a workaround: the test creates a commit (git commit -m base), and that extra commit shifted the history and caused later tests that depend on specific HEAD state to fail. Instead, I'll rework the test to use a file already tracked in the shared repo, dirty it, run git add -p with 'x', and clean up with test_when_finished. No extra commit needed. > It would be nice to see tests that split a hunk [...] and pathological context lines Agreed. I'll add tests that: (1) split a three-subhunk change (-a/+A, -c/+C, -e/+E), stage the first and third, and discard the second (2) the inverse — discard the first and third, stage the second (3) a pathological-context case along the lines of 2bd69b9024c Thanks again for the review — very helpful. Luiz Em qua., 25 de mar. de 2026 às 13:24, Phillip Wood escreveu: > > Hi Luiz > > On 25/03/2026 07:50, Luiz Campos wrote: > > When using `git add -p`, users can stage or skip hunks, > > but cannot discard unwanted changes from the working tree. > > > > Introduce a new 'x' action to discard the current hunk by > > reverse-applying it. > > > > This idea was suggested in a previous mailing list discussion: > > https://lore.kernel.org/git/X%2FiFCo0bXLR%2BLZXs@coredump.intra.peff.net/t/#m0576e6f3c6375e11cc4693b9dca3c1fc57baadd0 > > I tend to agree with peff's comments in that thread that it is rather > unexpected for "git add" to modify the working copy. I also think that a > command that lets you stage some changes and discard others could be > useful as I do both fairly frequently from my editor. Regardless of > whether we want a new command the implementation will be similar so I've > left some comments on the code below. > > > diff --git a/Documentation/git-add.adoc b/Documentation/git-add.adoc > > index 941135dc63..0ab81e5615 100644 > > --- a/Documentation/git-add.adoc > > +++ b/Documentation/git-add.adoc > > @@ -351,12 +351,15 @@ patch:: > > K - go to the previous hunk, roll over at the top > > s - split the current hunk into smaller hunks > > e - manually edit the current hunk > > + x - discard this hunk from the worktree > > p - print the current hunk > > P - print the current hunk using the pager > > ? - print help > > + > > -After deciding the fate for all hunks, if there is any hunk > > -that was chosen, the index is updated with the selected hunks. > > +After deciding the fate for all hunks, any hunks marked for > > +discard are removed from the working tree (reverted to the index > > +version for those lines). Then, if there is any hunk chosen for > > +staging, the index is updated with those hunks. > > Makes sense. > > > + > > You can omit having to type return here, by setting the configuration > > variable `interactive.singleKey` to `true`. > > diff --git a/add-patch.c b/add-patch.c > > index 4e28e5c187..ea38ab453e 100644 > > --- a/add-patch.c > > +++ b/add-patch.c > > @@ -259,7 +259,7 @@ struct hunk_header { > > struct hunk { > > size_t start, end, colored_start, colored_end, splittable_into; > > ssize_t delta; > > - enum { UNDECIDED_HUNK = 0, SKIP_HUNK, USE_HUNK } use; > > + enum { UNDECIDED_HUNK = 0, SKIP_HUNK, USE_HUNK, DISCARD_HUNK } use; > > struct hunk_header header; > > }; > > > > @@ -884,17 +884,35 @@ static void render_diff_header(struct add_p_state *s, > > } > > } > > > > +static bool should_merge_hunk(struct file_diff *file_diff, > > + size_t hunk_index, int use_all, > > + int merge_for_discard) > > +{ > > + if (use_all) > > + return true; > > If we're looking for hunks to discard then we want to return false if > all the hunks have been selected to be staged. > > > + return merge_for_discard > > + ? file_diff->hunk[hunk_index].use == DISCARD_HUNK > > + : file_diff->hunk[hunk_index].use == USE_HUNK; > > It would be simpler just to take USE_HUNK or DISCARD_HUNK as an argument > rather than a boolean here. > > > /* Coalesce hunks again that were split */ > > static int merge_hunks(struct add_p_state *s, struct file_diff *file_diff, > > - size_t *hunk_index, int use_all, struct hunk *merged) > > + size_t *hunk_index, int use_all, struct hunk *merged, > > + int merge_for_discard) > > Taking the type of hunk we want to retain (USE_HUNK or DISCARD_HUNK) > would avoid having to convert merge_for_discard back into the hunk type > in should_merge_hunk(). > > > { > > size_t i = *hunk_index, delta; > > struct hunk *hunk = file_diff->hunk + i; > > /* `header` corresponds to the merged hunk */ > > struct hunk_header *header = &merged->header, *next; > > > > - if (!use_all && hunk->use != USE_HUNK) > > + if (!should_merge_hunk(file_diff, *hunk_index, use_all, merge_for_discard)) { > > return 0; > > + } > > There's no need to add braces here > > > @@ -1014,11 +1032,13 @@ static int merge_hunks(struct add_p_state *s, struct file_diff *file_diff, > > > > static void reassemble_patch(struct add_p_state *s, > > struct file_diff *file_diff, int use_all, > > + enum reassemble_mode mode, > > struct strbuf *out) > > { > > struct hunk *hunk; > > size_t save_len = s->plain.len, i; > > ssize_t delta = 0; > > + int merge_for_discard = (mode == REASSEMBLE_DISCARD); > > It would by simpler just to take USE_HUNK or DISCARD_HUNK as a parameter > and pass that to reassemble_patch() rather than forcing the caller to > pass an enum that we then transform to a boolean. > > > > > render_diff_header(s, file_diff, 0, out); > > > > @@ -1026,25 +1046,26 @@ static void reassemble_patch(struct add_p_state *s, > > struct hunk merged = { 0 }; > > > > hunk = file_diff->hunk + i; > > - if (!use_all && hunk->use != USE_HUNK) > > + if (!should_merge_hunk(file_diff, i, use_all, merge_for_discard)) { > > delta += hunk->header.old_count > > - hunk->header.new_count; > > - else { > > - /* merge overlapping hunks into a temporary hunk */ > > - if (merge_hunks(s, file_diff, &i, use_all, &merged)) > > - hunk = &merged; > > + continue; > > I'm not sure this is an improvement - it certainly makes the patch > harder to read because you end up changing the indentation of otherwise > unchanged lines that were in the else clause. > > > + } > > > > - render_hunk(s, hunk, delta, 0, out); > > + if (merge_hunks(s, file_diff, &i, use_all, &merged, > > + merge_for_discard)) > > + hunk = &merged; > > > > - /* > > - * In case `merge_hunks()` used `plain` as a scratch > > - * pad (this happens when an edited hunk had to be > > - * coalesced with another hunk). > > - */ > > - strbuf_setlen(&s->plain, save_len); > > + render_hunk(s, hunk, delta, 0, out); > > We need to tell render_hunk whether a hunk is being applied in reverse > or not so that it knows whether to apply delta to the old offset or the > new offset. Currently it uses s->mode->reverse but that will not be > correct for the patch that discards changes from the working tree. I'm > not sure what the best way of doing that is, the simplest approach is to > invert s->mode->reverse when we want to keep hunks marked DISCARD_HUNK > and then restore the original value. > > > - delta += hunk->delta; > > - } > > + /* > > + * In case `merge_hunks()` used `plain` as a scratch > > + * pad (this happens when an edited hunk had to be > > + * coalesced with another hunk). > > + */ > > + strbuf_setlen(&s->plain, save_len); > > + > > + delta += hunk->delta; > > This is unchanged code that re-indented because of the addition of > "continue" above. > > > } > > } > > > @@ -1540,6 +1562,7 @@ N_("j - go to the next undecided hunk, roll over at the bottom\n" > > "/ - search for a hunk matching the given regex\n" > > "s - split the current hunk into smaller hunks\n" > > "e - manually edit the current hunk\n" > > + "x - discard this hunk from the worktree\n" > > It would be nice to avoid showing this for 'checkout -p' etc. > > > "p - print the current hunk\n" > > "P - print the current hunk using the pager\n" > > "> - go to the next file, roll over at the bottom\n" > > @@ -1547,21 +1570,57 @@ N_("j - go to the next undecided hunk, roll over at the bottom\n" > > "? - print help\n" > > "HUNKS SUMMARY - Hunks: %d, USE: %d, SKIP: %d\n"); > > > > +static int apply_discard_hunks(struct add_p_state *s, > > + struct file_diff *file_diff) > > +{ > > + struct child_process check_cp = CHILD_PROCESS_INIT; > > + struct child_process apply_cp = CHILD_PROCESS_INIT; > > + > > + strbuf_reset(&s->buf); > > + reassemble_patch(s, file_diff, 0, REASSEMBLE_DISCARD, &s->buf); > > + > > + discard_index(s->index); > > + > > + setup_child_process(s, &check_cp, "apply", "-R", "--check", NULL); > > + if (pipe_command(&check_cp, s->buf.buf, s->buf.len, NULL, 0, NULL, 0)) { > > + error(_("'git apply -R --check' failed")); > > + return -1; > > + } > > Why do we need to run "git apply --check" here? > > > + setup_child_process(s, &apply_cp, "apply", "-R", NULL); > > + if (pipe_command(&apply_cp, s->buf.buf, s->buf.len, NULL, 0, NULL, 0)) { > > + error(_("'git apply -R' failed")); > > + return -1; > > + } > > + > > + return 0; > > +} > > + > > static void apply_patch(struct add_p_state *s, struct file_diff *file_diff) > > { > > struct child_process cp = CHILD_PROCESS_INIT; > > size_t j; > > + int needs_refresh = 0; > > + > > + if (s->mode == &patch_mode_add) { > > + for (j = 0; j < file_diff->hunk_nr; j++) { > > + if (file_diff->hunk[j].use == DISCARD_HUNK) > > + break; > > + } > > + if (j < file_diff->hunk_nr && apply_discard_hunks(s, file_diff)) > > + return; > > + if (j < file_diff->hunk_nr) > > + needs_refresh = 1; > > + } > > > > - /* Any hunk to be used? */ > > Isn't this comment still relevant? > > > for (j = 0; j < file_diff->hunk_nr; j++) > > if (file_diff->hunk[j].use == USE_HUNK) > > break; > > > > if (j < file_diff->hunk_nr || > > - (!file_diff->hunk_nr && file_diff->head.use == USE_HUNK)) { > > - /* At least one hunk selected: apply */ > > + (!file_diff->hunk_nr && file_diff->head.use == USE_HUNK)) { > > What's the point of this change? > > > strbuf_reset(&s->buf); > > - reassemble_patch(s, file_diff, 0, &s->buf); > > + reassemble_patch(s, file_diff, 0, REASSEMBLE_STAGE, &s->buf); > > > > discard_index(s->index); > > if (s->mode->apply_for_checkout) > > @@ -1574,13 +1633,15 @@ static void apply_patch(struct add_p_state *s, struct file_diff *file_diff) > > NULL, 0, NULL, 0)) > > error(_("'git apply' failed")); > > } > > - if (read_index_from(s->index, s->index_file, s->r->gitdir) >= 0 && > > - s->index == s->r->index) { > > - repo_refresh_and_write_index(s->r, REFRESH_QUIET, 0, > > - 1, NULL, NULL, NULL); > > - } > > + needs_refresh = 1; > > } > > > > + if (needs_refresh && > > + read_index_from(s->index, s->index_file, s->r->gitdir) >= 0 && > > + s->index == s->r->index) { > > + repo_refresh_and_write_index(s->r, REFRESH_QUIET, 0, > > + 1, NULL, NULL, NULL); > > We now wait until we've applied both patches before refreshing the index > - sounds sensible. > > > + } > > } > > > > @@ -1722,6 +1784,10 @@ static size_t patch_update_file(struct add_p_state *s, > > !file_diff->deleted) { > > permitted |= ALLOW_EDIT; > > strbuf_addstr(&s->buf, ",e"); > > + if (s->mode == &patch_mode_add) { > > + permitted |= ALLOW_DISCARD; > > + strbuf_addstr(&s->buf, ",x"); > > + } > > So 'x' is only permitted if 'e' is what's the reason for that? > > diff --git a/t/t3701-add-interactive.sh b/t/t3701-add-interactive.sh > > [...] > > +test_expect_success 'add -p discard removes worktree change' ' > > + test_when_finished "rm -rf discard-testrepo" && > > + mkdir discard-testrepo && > > It's nice to see a test for 'x', but I'm not sure why this test needs to > be in a separate repository - why can't it use the same repository as > the other tests? > > > + ( > > + cd discard-testrepo && > > + git init -b main && > > + echo clean >discard-me && > > + git add discard-me && > > + git commit -m base && > > + echo extra >>discard-me && > > + test_write_lines x | git add -p discard-me && > > + printf "clean\n" >expect && > > + test_cmp expect discard-me && > > + git diff --cached >tmp && > > + test_must_be_empty tmp > > It would be nice to see tests that split a hunk like > > -a > +A > b > -c > +C > d > -e > +E > > and then (1) stage the first and third sub-hunks and discard the second, > (2) discard the first and third sub-hunks and stage the second. It would > also be nice to see a test that discards a hunk with pathological > context lines - see 2bd69b9024c (add -p: fix checkout -p with > pathological context, 2019-06-12) for an example. > > Thanks > > Phillip > > > > + ) > > +' > > + > > test_expect_success 'setup expected' ' > > cat >expected <<-\EOF > > diff --git a/non-empty b/non-empty > > @@ -521,13 +539,13 @@ test_expect_success 'split hunk setup' ' > > test_expect_success 'goto hunk 1 with "g 1"' ' > > test_when_finished "git reset" && > > tr _ " " >expect <<-EOF && > > - (2/2) Stage this hunk [y,n,q,a,d,K,J,g,/,e,p,P,?]? + 1: -1,2 +1,3 +15 > > + (2/2) Stage this hunk [y,n,q,a,d,K,J,g,/,e,x,p,P,?]? + 1: -1,2 +1,3 +15 > > _ 2: -2,4 +3,8 +21 > > go to which hunk? @@ -1,2 +1,3 @@ > > _10 > > +15 > > _20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]?_ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]?_ > > EOF > > test_write_lines s y g 1 | git add -p >actual && > > tail -n 7 actual.trimmed && > > @@ -540,7 +558,7 @@ test_expect_success 'goto hunk 1 with "g1"' ' > > _10 > > +15 > > _20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]?_ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]?_ > > EOF > > test_write_lines s y g1 | git add -p >actual && > > tail -n 4 actual.trimmed && > > @@ -550,11 +568,11 @@ test_expect_success 'goto hunk 1 with "g1"' ' > > test_expect_success 'navigate to hunk via regex /pattern' ' > > test_when_finished "git reset" && > > tr _ " " >expect <<-EOF && > > - (2/2) Stage this hunk [y,n,q,a,d,K,J,g,/,e,p,P,?]? @@ -1,2 +1,3 @@ > > + (2/2) Stage this hunk [y,n,q,a,d,K,J,g,/,e,x,p,P,?]? @@ -1,2 +1,3 @@ > > _10 > > +15 > > _20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]?_ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]?_ > > EOF > > test_write_lines s y /1,2 | git add -p >actual && > > tail -n 5 actual.trimmed && > > @@ -567,7 +585,7 @@ test_expect_success 'navigate to hunk via regex / pattern' ' > > _10 > > +15 > > _20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]?_ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]?_ > > EOF > > test_write_lines s y / 1,2 | git add -p >actual && > > tail -n 4 actual.trimmed && > > @@ -579,11 +597,11 @@ test_expect_success 'print again the hunk' ' > > tr _ " " >expect <<-EOF && > > +15 > > 20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? @@ -1,2 +1,3 @@ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? @@ -1,2 +1,3 @@ > > 10 > > +15 > > 20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]?_ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]?_ > > EOF > > test_write_lines s y g 1 p | git add -p >actual && > > tail -n 7 actual.trimmed && > > @@ -595,11 +613,11 @@ test_expect_success TTY 'print again the hunk (PAGER)' ' > > cat >expect <<-EOF && > > +15 > > 20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? PAGER @@ -1,2 +1,3 @@ > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? PAGER @@ -1,2 +1,3 @@ > > PAGER 10 > > PAGER +15 > > PAGER 20 > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? > > EOF > > test_write_lines s y g 1 P | > > ( > > @@ -796,21 +814,21 @@ test_expect_success 'colors can be overridden' ' > > +new > > more-context > > +another-one > > - (1/1) Stage this hunk [y,n,q,a,d,s,e,p,P,?]? Split into 2 hunks. > > + (1/1) Stage this hunk [y,n,q,a,d,s,e,x,p,P,?]? Split into 2 hunks. > > @@ -1,3 +1,3 @@ > > context > > -old > > +new > > more-context > > - (1/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? @@ -3 +3,2 @@ > > + (1/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? @@ -3 +3,2 @@ > > more-context > > +another-one > > - (2/2) Stage this hunk [y,n,q,a,d,K,J,g,/,e,p,P,?]? @@ -1,3 +1,3 @@ > > + (2/2) Stage this hunk [y,n,q,a,d,K,J,g,/,e,x,p,P,?]? @@ -1,3 +1,3 @@ > > context > > -old > > +new > > more-context > > - (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? > > + (1/2) Stage this hunk (was: y) [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? > > EOF > > test_cmp expect actual > > ' > > @@ -1424,9 +1442,9 @@ test_expect_success 'invalid option s is rejected' ' > > test_write_lines j s q | git add -p >out && > > sed -ne "s/ @@.*//" -e "s/ \$//" -e "/^(/p" actual && > > cat >expect <<-EOF && > > - (1/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,s,e,p,P,?]? > > - (2/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? Sorry, cannot split this hunk > > - (2/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,e,p,P,?]? > > + (1/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,s,e,x,p,P,?]? > > + (2/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? Sorry, cannot split this hunk > > + (2/2) Stage this hunk [y,n,q,a,d,k,K,j,J,g,/,e,x,p,P,?]? > > EOF > > test_cmp expect actual > > ' >