Skip to content

dev - #63

Open
probonopd wants to merge 149 commits into
mainfrom
dev
Open

dev#63
probonopd wants to merge 149 commits into
mainfrom
dev

Conversation

@probonopd

Copy link
Copy Markdown
Contributor

No description provided.

probonopd and others added 30 commits September 18, 2026 08:45
A UI test could not drag at all. The press was sent as a synthetic event, so
the X server never saw the button go down and the application received the
motion as plain mouse movement, never as a drag. Drags now press, move and
release through XTest, the way a real pointer does, so drag-and-drop can be
tested like any other gesture. drive_ui and run_uitest link libXtst for this.

A drop could only be aimed with a pixel offset, which breaks as soon as icon
size, grid spacing or window placement changes. drag "A" onto "B" presses on
the first widget and releases over the centre of the second, backed by a new
drive_ui drag_onto subcommand.

Tests had no reliable way to create the files they need. run types into the
application's Run dialog, cannot see what was launched, and was caught
dropping a command while reporting success. shell "cmd" runs the command in
the harness itself and fails the step on a non-zero exit, so it can build
fixtures and check what a UI action did to the file system.

A command naming a widget without a role took its first following word as
the role and removed it, so the "in" of "in window" was eaten and both the
window scope and any timeout after it were silently ignored. Only a word that
really names a role is consumed now, at every command that parses a role.

Executor.md and UITesting.md describe the new commands.
A drag onto a folder was always released the moment it arrived, so nothing that reacts to a pointer resting there, such as a folder springing open, could be tested. drag onto now takes hold <duration> in the uitest language and drive_ui drag_onto takes --hold <ms>; the pointer stays on the target for that long with the button down before it is released.
Starting a second isolated desktop truncated the logs of a test already running and could take over its display, because every session wrote to the same /tmp/uitest_* files and the script joined any display that was up. Logs and the dbus address now go to /tmp/uitest<N>_* named after the display number, and a display that is already up is refused unless /tmp/uitest-session.<N>.owner names the same test user, so each test user gets its own display without disturbing others. The session also takes the test user's home from the passwd database instead of assuming /home/uitest, so users other than uitest work.
…est run

A drag onto a target with a hold longer than about a second and a half was killed by the two second timeout meant for quick widget queries, leaving the mouse button down in the test session, and the error that followed crashed run_uitest because the message had been released with the query's autorelease pool before it was logged. The drag now gets the query timeout plus the hold, and error messages outlive the pool they are made in.
The Debian CI job stopped building DriveUI once drags went through XTest: libxtst-dev was never in the Debian package list, so X11/extensions/XTest.h was missing and the whole job failed before a single UI test ran. On Arch and FreeBSD the library only arrived by accident as a dependency of something else. The XTest development package is now listed explicitly for every distribution family that builds DriveUI.
…the pointer

activate application clicked the centre of the first widget the app reported. For Workspace that was often the Dock, so the click launched Terminal or System Preferences, which took the keyboard and swallowed what the test typed next. That made menu_follows_app time out on its Run dialog, let search_launch_processes pass on a stale window, and left windows covering the Desktop icons that dnd_drop_onto_desktop_folder drags. Activation now clicks nothing: the DriveUI bundle names the app's front window that can take the keyboard, drive_ui asks the window manager to activate it and waits until the app really has the keyboard.
gpbs started with --auto quits when its last client connection goes away, but it only counted clients that had put something on a pasteboard. When the cut/copy/paste test app quit, gpbs exited although Workspace was still connected, Workspace was left with a dead pasteboard proxy, and every Workspace UI test after that failed or the Workspace crashed. The libs-back patch counts every client connection, so gpbs stays up as long as any application uses it.
When a UI test failed in CI, only the last 40 lines of the Workspace, Menu and WindowManager logs were printed, which rarely reach back to the test that failed, so intermittent failures on FreeBSD could not be traced. The full logs are now copied next to the JUnit report, which the CI jobs already upload as an artifact on every platform.
UI tests could only drive widgets inside an application's own windows, so behaviour that belongs to the window manager, such as moving a window or snapping it to a screen edge, could only be tested with hand-written xdotool calls and geometry parsed in shell. Scripts can now drive window decorations the same way they drive window content: drag titlebar "T" by 200 60 or to left edge with an optional hold, and grab titlebar, move pointer and release pointer to check the screen while the button is still down. A script that stops between grab and release gets the button released for it, so no later test inherits a pressed mouse button.

A window's on-screen frame can now be compared and recorded like its window count: assert and wait until accept xwindow "T" x, y, width or height with a comparison, and setcount records any of them into a variable.

drive_ui provides this through new xwindow_frame, titlebar_press, pointer_move and pointer_release commands. The titlebar is found through the X display by window title and pressed in its middle, whatever height the theme and scale factor give it. The pointer and button handling is shared with the existing drag. Executor.md and UITesting.md document the commands, and control/window_decorations.uitest exercises them.
UI tests could drag a window by its titlebar but not use its buttons, so closing, minimizing or zooming a window through the window manager could not be tested. Scripts can now write click titlebar "Title" close, minimize or zoom. The buttons are drawn by the theme and are no windows of their own; drive_ui reads their places from the _WINDOW_TITLEBAR_BUTTONS property the window manager puts on the titlebar and clicks the middle of the button through XTest, and fails with a clear message when the window has no such button or the window manager publishes none. drive_ui titlebar_buttons lists them for diagnosis. The control test now zooms, restores, minimizes and closes a window, and Executor.md and UITesting.md document the command.
While typing in Stickies, TextEdit or any other NSTextView, the mouse cursor turned into an arrow for a moment on every keystroke and then back into the I-beam, which looked like constant flicker. The text view invalidates its cursor rectangles after each edit, and libs-gui discarded them right away, which popped the I-beam off the cursor stack; the arrow reached the screen before the next event rebuilt the rectangles and pushed the I-beam again. The new libs-gui patch leaves the stale rectangles in place and lets resetCursorRects discard and rebuild them in one step before the next event is handled, so the cursor stays an I-beam while typing. Views that are removed or deallocated still discard their own rectangles as before.
A submenu with no room on the right of its menu opens on the left, where it covers the menu it came from or one of that menu's ancestors. Moving the pointer into it handed tracking to the covered menu instead, which closed the submenu, so its items could not be reached with the mouse. Menu tracking in libs-gui now gives the pointer to the most deeply nested open menu under it before it looks at the menu itself or its ancestors. The patch also adds an ObjectTesting test that proves the hand-over for a submenu covering an ancestor, a submenu covering its own parent, and an uncovered ancestor.
At a GSScaleFactor other than 1 the X11 backend refused to attach an OpenGL context to any view, as every view counts as scaled from window base coordinates then. Apps with an OpenGL view threw while building and drawing their window, so it never appeared (Player's cover carousel at 1.1). The GL subwindow is now placed from the view's rectangle in window base coordinates, which are device pixels, relative to the content view when the window manager draws the decorations; only rotated views are still refused.
…hat is already gone

The previous patch made the NSMessagePort listener non-blocking but only logged when that failed, so a port could still come up with a blocking listener and hang its thread in accept() the first time the run loop reported a connection that a nested run loop had already accepted, which is how Workspace stopped responding and every app talking to its Dock hung on quit. Port creation now fails when the listener cannot be made non-blocking, like every other setup error there, and the patch carries a test that creates a port, checks its listener does not block and hands it a readable event with no pending connection, which hangs without the fix and returns with it.
Right after an app started, a quick click on a table row only highlighted it and the selection needed a second click, because libs-back read the mouse settings on the first button press and registered for their changes with the distributed notification center there; waiting for gdnc ran a nested run loop that queued the button release before the press. The first click now works like every other one: the X server reads the mouse settings and registers once when it starts, so no press waits for another process.
A window that cannot be resized where its grip sits, such as one gone full screen, had no way to be rid of it: GNUstep's -setShowsResizeIndicator: only logs that it is not implemented, and the getter answers from the style mask alone. The patch keeps the flag in a spare bit of the window flags, so themes drawing the grip can honour it and Player can drop its grip in full screen.
Clicking somewhere else to put a menu away could carry out whatever item the pointer had last been on, so a Dock folder shown as a list opened a folder that was never chosen. An item whose submenu has just been attached keeps its highlight while the pointer is off the menu, so that the pointer can reach that submenu, and the highlight was then carried out on the release. The menu now drops the highlight and chooses nothing when the mouse is released over none of its items, while a release on an item or in a submenu still works. Carried by a patch to libs-gui, proposed upstream as gnustep/libs-gui#948 with a test that presses on an item with a submenu, moves the pointer away and releases it there.
An application that handed a URL to the desktop stayed frozen for as long as the handler ran, and handlers such as xdg-open keep running until the browser they started quits, so clicking a link in a chat window locked that window for minutes. NSWorkspace now starts the handler and carries on instead of waiting for it to exit, as it does on other systems.
Workspace used two to five percent of a processor core permanently while
nothing happened on the desktop, which costs battery and keeps a laptop
warm for no reason. The desktop now sleeps between its timers, so an idle
session leaves the processor alone: Workspace dropped from 3.6 percent of a
core and 7,239 poll calls per second to 0.8 percent and 233, and every
other program with a repeating timer gains the same.

The run loop computed how long to wait with a truncating conversion to
milliseconds, so any remaining wait shorter than a millisecond became a
timeout of zero. The loop then polled without waiting, found nothing ready
and ran again at once, spinning at full speed through the last fraction of
a millisecond before each timer fire. A program with a single timer at 30
Hz made 7,466 poll calls a second instead of 30. This patch rounds the wait
up instead, which waits at most one millisecond longer than asked and lets
the loop block as it was meant to.
A UI test that needs a fixture file had to spell out an absolute path to it,
which only worked on a machine with the sources in /Developer and broke the
test in CI, where the checkout sits in the build agent's work directory. Every
script now gets ${testdir}, the folder it lives in, so it reaches its own
fixtures wherever the repository was cloned. The control flow test checks the
variable, and UITesting.md tells script authors to use it instead of an
absolute path.
…he default theme

An application launched while the GSTheme default was unset, or set to
GNUstep, kept that theme for the rest of its run: GSTheme registers its
observer of the user defaults only from +setTheme:, and +setTheme: is
not reached at start-up when the theme in the defaults is the one that
is already current.  Switching the desktop to another theme therefore
left every such application behind, and the user had to restart them.
The observer is now registered when GSTheme initialises, so the theme
default is followed from the first change onwards.
NSTableColumn returns its width, minimum width and maximum width as
CGFloat but kept the three in float ivars, so on a 64 bit system every
width was rounded on the way in: a column set to 123.456789 reported
123.45678710937501, and a layout computed from those widths drifted.  The
ivars are CGFloat now, while the archive format stays as it was, the three
still being written and read as float, so a column saved by an earlier
version is unchanged.

Submitted upstream as gnustep/libs-gui#951.
Laying out a toolbar item pushed the item's own image onto its back view
whenever the toolbar was not showing labels alone.  An item with a custom
view has no image of its own and the back view forwards -setImage: to that
view, so the view was sent nil and the image the application had put there
disappeared as soon as the toolbar laid itself out.  The image is now only
pushed when the item has no custom view or does have an image of its own.

Submitted upstream as gnustep/libs-gui#952.
…menu bar

-[NSScreen visibleFrame] took the whole screen and subtracted the menu
bar, so it ignored whatever else reserves space on the screen: an
application that placed a window in the visible frame put it under the
Dock.  GSDisplayServer now answers -workAreaForScreen: with the area the
display server leaves to application windows, which the X11 backend takes
from _NET_WORKAREA, and the visible frame starts from that area.  A
display server that knows of no such area answers an empty rect and the
visible frame is what it was, so nothing that lacks the backend side
changes behaviour or raises.

Submitted upstream as gnustep/libs-gui#953.
Loading a nib brought every window listed as visible to the front.  Nib
conversion tools leave that list empty or mark every top level window in
it, so an application either showed nothing at load or covered the screen
with all of its dialogs at once.  The window the File's Owner's window
outlet points at is brought to the front instead, and the old behaviour
is kept for a nib that has no such outlet.

This one is ours to carry: the list an application ends up showing is a
matter of what our nib conversion produces, so it is not proposed
upstream.
The X11 backend had no way to pass on the area a window manager leaves to
application windows, and it used that area as the screen frame instead
whenever there was a single monitor.  A frame that is not the whole screen
disagrees with the coordinate flip the backend does for windows, so a
window asked for at the top of the screen landed lower down by whatever
the panels reserve, while an application that wanted to avoid the Dock had
nothing to ask.

The screen frame is now always the whole monitor, the work area is read
from _NET_WORKAREA with its origin flipped into screen coordinates rather
than clamped to zero, and -workAreaForScreen: answers it; the other
backends answer their full screen, which changes nothing for them.
Together with the libs-gui side this is what -[NSScreen visibleFrame]
returns.

This replaces screen-frame-not-workarea.patch, which took the work area
out of the frame but stopped there, and which cannot be applied beside
this one.  Upstream has the first half open as gnustep/libs-back#217 and
the origin flip as #226; the accessor itself is not upstream yet.
gpbs re-launches itself to become a daemon and then dropped the NSTask
without waiting for it, so the intermediate process stayed a zombie for
as long as the session lived, in both the normal and the failing path.
It is waited for now, which is immediate: the child has already forked
the daemon and exited by then.
The patch that built the visible frame from the work area went on to
subtract the menu bar height from it, so where the window manager had
already reserved the menu bar's strip, as it does when the menu bar is a
window of its own, those pixels were taken off twice and applications were
told the screen was shorter than it is.  Only as much of the menu bar
height as the work area has not already been reduced by is taken off now.

Measured on an 800x600 screen with a 22 pixel menu bar at the top and a
dock at the bottom: the visible frame was 0 128 800 428 and is now
0 128 800 450, which is the work area itself.

Reviewers of gnustep/libs-gui#953 asked for exactly this; the pull request
carries the same change.
probonopd and others added 30 commits October 24, 2026 13:44
networkmidid, the network MIDI daemon in gershwin-components, gives each
network MIDI session a port in the ALSA sequencer through libasound, and
none of the package lists installed its headers, so it could not be
built anywhere, and on FreeBSD and NextBSD there was no sequencer at
all. The Linux lists now install the libasound development package
(libasound2-dev, alsa-lib, alsa-lib-devel). FreeBSD, GhostBSD and
NextBSD get alsa-lib, alsa-seq-server, the daemon that provides
/dev/snd/seq through cuse and bridges USB MIDI devices into it, and
alsa-utils for aconnect and aseqdump. OpenBSD has no ALSA; networkmidid will
use sndio there.
A toolbar that shows icons only padded every item with a view to at least 60 points and to the width of its label, so Workspace's Look Inside and Action buttons came out at different widths and could push the search field out of the window. nstoolbaritem-icon-only-view-width.patch sizes such an item by its view, as gnustep/libs-gui#1016 proposes upstream, so both buttons are 44 points wide.

The toolbar customization palette showed every item without a palette label of its own without any label, and offered no default set to drag back once a toolbar had been customized. toolbar-palette-default-set.patch fixes both, as gnustep/libs-gui#1014 and #1015 propose upstream: items keep their label, and the palette has the default set as one row that restores the default items when dropped on the toolbar.

Tables with alternating row colours left the area below their last row plain after part of it was redrawn, because the stripes restarted at row 0 for the redrawn area. table-stripes-below-last-row.patch keeps the stripes in place by position, as gnustep/libs-gui#1017 proposes upstream, so every table under the Eau theme shows them all the way down, not only Workspace's list view, which paints its own. The upstream diffs and PR texts are kept next to the patches, and the whole libs-gui patch set still applies in order on the pinned commit.
The Windows build ran in the MSYS2 MINGW64 environment, which MSYS2 has
deprecated; its packages are degrading on the mirrors and CI runs started
failing in pacman with corrupted or stalled downloads. The build now runs
in UCRT64, the environment MSYS2 recommends, so the Windows CI keeps
working. UCRT64 was chosen over CLANG64 because it keeps the same clang and
lld toolchain with libstdc++ and libgcc, the runtime libobjc2's exception
handling was brought up on, and only replaces the legacy msvcrt with the
Universal CRT; every package the stack needs exists there.

windows.txt now names the mingw-w64-ucrt-x86_64 packages, which also gives
the core system a new windows-system-id. windows-bundle-runtime.sh no longer
matches a hard-coded mingw64 directory when picking the toolchain DLLs to
bundle but the active environment's MINGW_PREFIX, and the comments and the
README name the new environment.
Text in table and outline rows sat at the top of each row instead of in its middle, 2 pixels high at scale 1 and 4 to 5 pixels at scale 2, in the Keyboard pane and every other list whose rows are taller than one line of text, because NSTextFieldCell draws its title at the top of its title rect and in a table that rect is the whole cell. table-text-vertical-center.patch gives a non-wrapping text cell in a table or outline view a one-line title rect centered in the cell and rounded to whole pixels, used for drawing and for the field editor alike, so text and in-place editing line up in the middle of the row for every theme. Wrapping cells, labels, text fields and headers are unchanged. It is proposed upstream as gnustep/libs-gui#1018; the upstream diff and PR text are kept next to the patch, and the whole libs-gui patch set still applies in order on the pinned commit.
The Windows workflow published the prebuilt core system only on push, but
it could only be started by hand, so nothing was ever published and every
Windows run of the other repositories built the whole GNUstep stack again.
It now also runs on pushes to dev that touch the files windows-system-id.sh
hashes, and publishes on manual runs too. Since the zip is named by that
hash, a zip published from any branch only serves checkouts that would
build the identical system. A concurrency group keeps two runs from
uploading the same asset at once.

The MSYS2 mirrors often stall or serve a corrupted package, which failed
whole runs in Bootstrap. bootstrap.sh now retries the Windows pacman
transaction up to three times with a growing pause, dropping partial and
uninstalled downloads from the cache in between so a bad file is not
reused. setup-msys2 no longer installs git and make itself, since it has
no retry; windows.txt already lists both.
A system install started from a gershwin-developer checkout on dev, or from a worktree made from a dev commit, still built every Gershwin repository from its default branch unless BRANCH=dev was given. That put an old Eau main next to a dev WindowManager, so windows showed Eau's orbs and the window manager's own buttons at once and alert buttons came out in a tiny font. checkout.sh now prefers the dev branch when the build system itself is on dev, or detached at a commit that is on origin/dev but not yet on origin/main; a checkout on main keeps building the default branches, and an explicit BRANCH= still wins.
…veloper into fix/shallow-clone-update

# Conflicts:
#	Library/Repositories.csv
#	Library/Scripts/checkout.sh
Check out shallow clones for the live images
* More will break after next pkg update until this whole list is reworked.
checkout.sh cloned every repository in Repositories.csv into
Library/Sources, including gershwin-developer itself, although the
checkout running the script already is gershwin-developer. Software
Update then updated that second clone while the package lists, scripts
and Repositories.csv kept being read from the first, so new
prerequisites such as alsa-lib never arrived and builds failed. The
script now skips its own entry; Software Update treats /Developer as the
gershwin-developer checkout. Existing second clones are left alone and
are no longer used.
With monitors of different sizes everything was drawn at one scale
factor: on the larger or denser monitor windows were too small or too
big, and the Display pane could not give each monitor its own scale.

New libs-back and libs-gui patches read each monitor's scale from the
RandR output property DPI (96 per 100%) and lay out each window at the
factor of the monitor it is on. A window that moves to a monitor of
another scale keeps its content size in points and changes its size in
pixels, with hysteresis so it switches once; desktop windows keep their
factor, hidden windows switch when they are shown, and a guard stops a
window that keeps switching. The backend delivers resize events in the
order of the X events, ignores window sizes that are out of date,
collapses bursts of RandR events onto the newest one, and gives the
pointer the size of the monitor it is on. With all monitors at one
scale nothing changes.
With monitors of different pixel density everything was drawn at one scale factor, so windows were too small or too big on one of them. New libs-back and libs-gui patches read each monitor's scale from the RandR output property DPI and lay out each window at the factor of the monitor it is on; a backport of libs-back #251 collapses bursts of RandR events onto the newest one. With all monitors at one scale nothing changes.
checkout.sh also cloned gershwin-developer into Library/Sources, and Software Update then updated that second clone while package lists and scripts were read from the first, so new prerequisites never arrived and builds failed. The script now skips its own entry.
The WindowManager now loads theme cursors at the size of each monitor through libXcursor, and its build failed on Debian, where the Xcursor headers were not installed by any other package. The Xcursor development package is now listed for every system, so the build no longer depends on it arriving by chance.
The WindowManager build failed on Debian for want of the Xcursor headers; they are now in the package list of every system.
The libs-gui patch that keeps the alternating row colors below the last row counted the stripes as the clip rect's position divided by the row height, which only works when all rows have the same height, as Fred Kiefer pointed out in gnustep/libs-gui#1017. Tables and outline views with rows of different heights now get each row's stripe in that row's own rect, and below the last row the stripes continue at the standard row height from the end of the last row, so a partial redraw paints the same stripes as a full one. The patch notes, the upstream diff and the PR text follow the reworked PR, and the whole libs-gui patch set still applies in order on the pinned commit.
NextBSD's base system already ships Apple's libxml2, libxslt and ICU
(libicucore), so drop the FreeBSD ports of them from nextbsd.txt.

ffmpeg, freerdp and xdotool still pull in the ports libxml2 and ICU, and
pkgconf searches /usr/local before /usr/libdata/pkgconfig, so libs-base
linked the ports copies. At run time ld-elf.so.1 loaded base's
/usr/lib/libxslt.so.1 instead (same soname) and libs-gui's GSspell
Info-gnustep.plist step failed:

  /usr/lib/libxslt.so.1: version LIBXML2_1.0.11 required by
  libgnustep-base.so.1.31 not defined

Configure libs-base, libs-corebase and libs-gui with
PKG_CONFIG_PATH=/usr/libdata/pkgconfig so they pick the base .pc files.
For libs-gui, also skip the ports icu-config and pass icu-i18n's cflags
so its pkg-config check finds libicucore instead of falling back to
-licui18n -licuuc -licudata.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NextBSD: build against base libxml2, libxslt and ICU
Alt-Tab could crash the window manager: libs-gui kept the default bold system font name without retaining it, so with the monospace default fixed-pitch patch the name was freed and the switcher title label read dangling memory.

The new libs-gui patch backports upstream 8a092cfa7 so that the default font names are retained when stored.
GNUstep maintainers asked for curly brackets around every if, else, for and while body, even one-liners, and the upstream pull requests were amended accordingly. The carried patches and the .upstream.diff copies still had the brace-less form, so they no longer matched what is under review.

The patches and diffs of the table stripes, icon-only toolbar item width, toolbar palette default set, per-screen scale (libs-gui and libs-back) and table text centering changes are rebuilt from the amended branches: the patches by replaying the whole set on the pins and re-diffing, the upstream diffs from the branch commits. Only braces and indentation differ; the full set still applies in glob order with patch -p1.
The libs-gui pull requests for icon-only toolbar item width and table text centering were rebased onto current master because their ChangeLog entries conflicted. The .upstream.diff copies carried the old ChangeLog hunks, so they are rebuilt from the rebased commits. The .patch files are unchanged and still apply in glob order on the pin.
A view is clipped to its own bounds, so a theme could only mark the default button inside the button's frame and the ring George has for it was never drawn. With this patch a theme reports how far outside the frame it draws and libs-gui has it paint in the button's superview after the subviews, and redraws that area whenever the ring would go stale. The default outset of 0 changes nothing for other themes. Proposed upstream as gnustep/libs-gui#1021.
Copying a folder that contains a symbolic link, in Workspace or any other application, gave the file the link points to the link's own mode, usually 0777, together with the link's owner and dates, because libs-base applied the copied link's attributes through calls that follow links. A folder copy could so make an unrelated file, even one outside the folder, writable by everyone. copy-symlink-attributes-not-target.patch makes the copy, link and move code of NSFileManager set the owner of a recreated link with lchown(), its dates with utimensat() and AT_SYMLINK_NOFOLLOW, and leave its mode alone, while changeFileAttributes:atPath: and regular files and folders behave as before. The patch adds Tests/base/NSFileManager/copySymlinkAttributes.m, which fails 6 of its 16 checks on upstream master and passes with the patch; the upstream copy and the PR text are kept next to the patch.
The new copySymlinkAttributes test of the libs-base symbolic link patch failed on the Android job of the upstream CI, because linkPath:toPath:handler: makes hard links and the Android sandbox does not allow them, so the test reported a failure that has nothing to do with the fix. The test now tries one hard link first and skips the linkPath checks when that is refused; on this machine it still fails 6 of 16 checks on upstream master and passes all 16 with the patch, and the whole patch set still applies in order on the pin.
The copySymlinkAttributes test of the libs-base symbolic link patch broke the Windows builds of the upstream CI, because it uses sys/time.h, utimes() and link(), which MSVC and MinGW do not have, so the whole test run reported a failed build. The test now only reports a skip on Windows, where POSIX symbolic links are not available anyway; on this machine it still fails 6 of 16 checks on upstream master and passes all 16 with the patch, and the patch set still applies in order on the pin.
The libs-gui maintainer asked to implement the logic behind
-[NSView wantsDefaultClipping] instead of a superview hook that only
served the default button ring (gnustep/libs-gui#1021). Themes with a
ring such as George need the patch in its new form to keep showing it,
and any view that does not clip can now draw outside its frame without
leaving stale pixels behind.

A view that returns NO from -wantsDefaultClipping may now draw anywhere
in the visible part of its superview: display, focus and invalidation no
longer cut its rectangles at its bounds, the superview redisplays it
wherever it redraws, and the area is redrawn when it moves, resizes or
leaves its superview. With a ring outset from the theme, the default
button returns NO and the theme draws the ring together with the button;
-drawDefaultButtonRingForCell:inView: is gone, so themes must be updated
together with this patch. The upstream diff and PR text follow the
reworked PR.
Our cursor-rects-no-flicker patch (gnustep/libs-gui#942) kept the I-beam
steady while typing in a text view. Upstream fixed the same bug in
NSTextView instead (gnustep/libs-gui#834, merged 2026-08-01): on current
libs-gui master, typing ten characters causes no cursor change at all,
and #942 has been closed as obsolete.

Our pin predates #834, so dropping our patch alone would bring the
flicker back. The patch is replaced with a backport of #834, which gives
no cursor change while typing on the pin as well; drop it once the pin
includes that commit.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants