Skip to content

Feat: Windows drop targets, live Resizable, per-window color scheme and zoom (FR-53, #398 #400 #403) - #598

Merged
b-re-w merged 8 commits into
developfrom
feat/fr53-windows-398-400-403
Nov 13, 2026
Merged

b-re-w merged 8 commits into
developfrom
feat/fr53-windows-398-400-403

Conversation

@b-re-w

@b-re-w b-re-w commented Nov 12, 2026 •

Copy link
Copy Markdown
Member

Closes #398
Closes #400
Closes #403

Not built locally; verified by the session's builder.

What changed

#398, drop targets and cancelCapture (FR-53.7). MultiWindowLoop builds the window manager with drop targets answered from the bounds the scenes register (DropTargets.Registered, with the point taken from dp to the pixels of the window's monitor) and with cancelCapture. A window torn off under a held pointer cancels the pointer input of the scene the button went down in, between two events, and keeps the rest of that gesture (moves with the button down, the release) from the scene. No common renderer file changed, so this does not meet the macOS and Linux branches.

#400, Resizable on an open window (FR-53.3). WindowsPlatformWindow.setResizable calls dxc_native_window_set_resizable on the selected window.

#403, color scheme and zoom per window (FR-53.3, FR-53.4). The scene half was already per window, because each window's content is composed under its node id: it resolves the theme the table holds for that node and holds a zoom of its own that takes the node's ZoomLevel and reports ZoomChanged with the node id. The checks now prove that on real windows. New here is what the system draws: setColorScheme calls dxc_native_window_set_color_scheme, at open and on change.

Needs the Compose fork first

The Win32 style and the DWM dark caption are window code, which lives in thisisthepy/compose-multiplatform-core-extended (extended/window/windows-shared/c/win32_window.c). The pinned WINDOW_REVISION (e3088cc0) defines none of these three:

int32_t dxc_native_window_set_resizable(int32_t resizable);   // 1 when applied
int32_t dxc_native_window_set_color_scheme(int32_t scheme);   // 0 follow system, 1 light, 2 dark; 1 when applied
int32_t dxc_native_window_caption_dark(int32_t *dark);        // 1 when it wrote *dark

So that this branch links against either pin, renderer/windows/native/window_layer_fallbacks.c holds a stand-in for each that the MSVC linker takes only where the fork's definition is missing (/alternatename). A stand-in returns 0 and changes nothing about a window.

Until the pin moves to a fork commit that defines them, two checks fail on purpose and name the missing call:

  • fr53_windows_resizable_change_toggles_the_frame_of_an_open_window
  • fr53_windows_window_color_scheme_darkens_its_own_caption

The other four new checks do not depend on the fork.

Checks

Run by the renderer itself against real windows (renderer/windows/src/Fr53WindowsCheck.kt), in "Native renderer test", job "Windows x64 Kotlin/Native static renderer and an application that depends on the crate", step "Build, check, move and run an application" (scripts/check-windows-consumer.sh). It is a heavy workflow, so it needs the run-heavy-ci label.

  • fr53_windows_dragged_window_enters_moves_and_drops_on_a_subscribed_node
  • fr53_windows_tear_off_ends_the_source_pointer_capture
  • fr53_windows_resizable_change_toggles_the_frame_of_an_open_window
  • fr53_windows_window_theme_sets_the_color_scheme_of_one_window
  • fr53_windows_window_color_scheme_darkens_its_own_caption
  • fr53_windows_zoom_level_reaches_its_window_and_reports_its_node

scripts/tests/windows-kotlin-native.test.sh now accepts a called C name with no definition when a stand-in for it is defined once.

Verified

The link stand-ins are gone: the window pin from #615 (fork #194) defines dxc_native_window_set_resizable, set_color_scheme and caption_dark. On feat/p1T-ci-batch2 (this PR with #615 and #614), native renderer run 43345202790, Windows x64 Kotlin/Native static renderer: all 13 fr53_windows checks ok (incl. tear_off_ends_the_source_pointer_capture, resizable_change_toggles_the_frame_of_an_open_window, window_color_scheme_darkens_its_own_caption, zoom_level_reaches_its_window_and_reports_its_node) and the four fr53_windows_app steps.

…ource capture (FR-53, #398)

Covers FR-53.7 on Windows: a window dragged over a node of another window that
registered OnWindowDragOver sends it WindowDragOver Enter, Move and Drop in the
node's own coordinates, and a window opened with DragFromPointer ends the
gesture of the node that held the pointer.

MultiWindowLoop builds the window manager with drop targets answered from the
bounds the scenes register (taken from dp to the pixels of the window's
monitor), and with cancelCapture. A window cancels the scene's pointer input
between two events and keeps the rest of that gesture from the scene; the Win32
window keeps the mouse captured until the button comes up, so the release
still arrives and is dropped.

Checks, run by the renderer on a Windows runner:
fr53_windows_dragged_window_enters_moves_and_drops_on_a_subscribed_node,
fr53_windows_tear_off_ends_the_source_pointer_capture.
Covers FR-53.3 on Windows: changing Resizable on a live Window node gives the
window the thick frame and the maximize box or takes them away.

WindowsPlatformWindow.setResizable calls dxc_native_window_set_resizable on the
selected window. The style is the window layer's to change, and the Compose
fork commit the build pins does not define that call yet, so
windows/native/window_layer_fallbacks.c carries a stand-in that the MSVC linker
takes only where the fork's definition is missing (/alternatename). The
stand-in returns 0, the renderer says once that the change did not apply, and
the check fails with the name of the missing call until the pin moves.

windows-kotlin-native.test.sh accepts a called name with no definition when a
stand-in for it is defined once.

Check, run by the renderer on a Windows runner:
fr53_windows_resizable_change_toggles_the_frame_of_an_open_window flips
Resizable both ways on two windows and reads WS_THICKFRAME and WS_MAXIMIZEBOX
back from the system.
…R-53, #403)

Covers FR-53.3 on Windows: a window's theme override (SetWindowTheme) sets its
color scheme, at open and after, and ZoomLevel on a Window node reaches that
window's zoom, with ZoomChanged carrying the window's node id (FR-53.4).

The scene half was already per window: each window's content is composed under
its node id, so it resolves the theme the table holds for that node and holds
a zoom of its own that takes the node's level. The checks now prove it on real
windows. What was missing is what the system draws: WindowsPlatformWindow
.setColorScheme calls dxc_native_window_set_color_scheme on the selected
window, and the backend applies the scheme a window opens with.

The Compose fork commit the build pins does not define that call or
dxc_native_window_caption_dark yet, so both have stand-ins in
window_layer_fallbacks.c that link only where the fork's definition is
missing. Until the pin moves the caption check fails with the name of the
missing call; the theme and zoom checks do not depend on it.

Checks, run by the renderer on a Windows runner:
fr53_windows_window_theme_sets_the_color_scheme_of_one_window,
fr53_windows_window_color_scheme_darkens_its_own_caption,
fr53_windows_zoom_level_reaches_its_window_and_reports_its_node.
…ear-off (FR-53, #398)

The Windows loop passed its own cancelCapture, which replaced the shared one that
tells the capturing node its Cancel through the pointer router. It now does both: the
router sends the Cancel and drops the rest of the press, and the source window ends
the gesture in its scene and withholds the release.

fr53_windows_tear_off_ends_the_source_pointer_capture failed on CI because of the
check, not the tear-off. The window layer reads the buttons from the real key state,
so a posted press holds nothing, and the first click's release lets go of the mouse
capture, which makes the system send a move to the window under the real pointer.
That move arrived after the second press, said no button was down, and the scene took
it as the missing release: the click the check blamed on the tear-off. The check now
lets that move through before it presses, proves the press is still held before the
window is torn off, and also asserts the node's Cancel and that nothing more of the
press reaches it.

Covers FR-53.7 (the capturing node gets a Cancel) and the Windows half of FR-53.11
criterion 10.
…-53, #398)

A dispatched run held the Windows static job for 95 minutes in section 3c and, once cancelled, kept no log. 3b and 3c now run under timeout, print their last lines on failure, and each fr53_windows check says it starts before it runs. The backend grep test follows the tear-off cancel, which now sends the node its Cancel as well.
Takes the Windows files as verified on feat/p1T-ci-batch2 (run 43345202790, every fr53_windows check ok): the fork's real dxc_native_window_set_resizable / set_color_scheme / caption_dark (pin from #615) replace the link-time stand-ins, which are deleted, and the symbol test is back to each @SymbolName defined exactly once.
@b-re-w
b-re-w marked this pull request as ready for review November 13, 2026 03:35
@b-re-w
b-re-w merged commit b38913f into develop Nov 13, 2026
7 of 12 checks passed
@b-re-w
b-re-w deleted the feat/fr53-windows-398-400-403 branch November 13, 2026 03:36
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.

1 participant