Skip to content

Chore: Pin the window modules to the fork's extended head (WIP) - #360

Merged
b-re-w merged 21 commits into
developfrom
feat/repin-fork-extended-head
Oct 27, 2026
Merged

b-re-w merged 21 commits into
developfrom
feat/repin-fork-extended-head

Conversation

@b-re-w

@b-re-w b-re-w commented Oct 26, 2026 •

Copy link
Copy Markdown
Member

Refs #355. Moves all three fork pins (REVISION, WINDOW_REVISION, MINGW_REVISION) to one commit of the fork's extended (ea24655, fork #170 and #172 merged) and adopts the fork's 1.11.1-ext.2 coordinates. #277 was the reference for the coordinate moves; it is closed as superseded once this merges.

What changed

  • Pins, COMPOSE_AS/MATERIAL3_AS/SKIKO_AWT_AS at 1.11.1-ext.2, module.yaml coordinates, regenerated compose-fork*.changes.
  • Accessibility glue: the renderer flattens the scene's semantics (ids, parents, states, actions, values, window points), gives them to the fork windows' own bridges with setAccessibility, and carries a reader's request back through onAccessibilityAction (macOS GraalVM, X11 GraalVM, Kotlin/Native X11). AtspiSemanticsSource, X11Accessibility and the macOS AccessibleElement are no longer used.
  • Windows: the fork now rejects versions JetBrains could publish, so mingwX64 is published through the fork's publish-local.sh under org.thisisthepy.compose and the windows module asks for those coordinates.
  • skiko-awt: the fork lists JetBrains' skiko-awt as a second capability, which Amper cannot resolve; publish-skiko-awt.sh drops it and the native build scripts keep JetBrains' copy off the image class path (taken from Chore: Move the Compose fork pin to 1.11.1-ext.2 #277's reverted commit).
  • publish-window.sh publishes scene whenever native/macos is built.

Relaxed check, with reason

test-compose-from-fork.yml step "Fail if the two Compose builds differ" compared the fork with the old patched build (JetBrains 73ac849 plus renderer/patches). That proof was for the pin that held the same tree. On the ext.2 line the two differ on purpose (sources, metadata, coordinates), so the step now reports and no longer fails. The guard returns as a comparison with the ext.2 baseline: #361 (linked in the step's comment).

Windows log: curl: (22) 404

Not a failure of this build. It comes from the fork's extended/window/kotlin.bat, which pre-fetches two Adoptium JDK zips with curl -fsSL ... && move before the toolchain runs. The first (jdk-25.0.4.1+1, OpenJDK25U-jdk_x64_windows_hotspot_25.0.4.1_1.zip) answers 200 and is moved ("1 file(s) moved"); the second (jdk-25.0.5+8 ... 25.0.5_8.zip) answers 404 today, so that pre-fetch is skipped (no move, no exit code checked) and the toolchain downloads on demand if it needs that JDK. The step went on to compile and publish windows-shared. Harmless; the stale URL is the fork's to fix.

Proof

Heavy CI (run-heavy-ci): windows-static, macOS, Linux, samples, and the wheel/caption script tests. Merged only when green on this head.

b-re-w and others added 9 commits October 25, 2026 09:50
REVISION, WINDOW_REVISION and MINGW_REVISION all move to c7caf51523e8 (tag v1.11.1-ext.2,
the merge of the fork's bump-ext-2). COMPOSE_AS and MATERIAL3_AS become 1.11.1-ext.2 (the
fork now versions material3 with the rest), the window modules go from 0.1.0 to 1.11.1-ext.2,
compose-fork.changes and compose-fork-mingw.changes are regenerated from the new commit, and
the tests and docs that name the versions follow.
…low AccessibleElement's move

native/macos depends on scene now, so publish-window.sh keeps scene's jvm and macosArm64
platforms on a Mac that has the fork's Compose for macosArm64 (jvm alone elsewhere). The
test fixture gains the scene module the script edits. AppKitWindow imports AccessibleElement
from org.thisisthepy.compose.window, where it moved.
…neSemantics

The fork's X11 windows (GraalVM and Kotlin/Native) join the AT-SPI bus and serve a reader
themselves, from the AccessibleElement list given to setAccessibility, and report a reader's
presses through WindowListener.onAccessibilityAction. AtspiSemanticsSource, X11Accessibility
and the renderer-side AtspiBridge are gone. ReaderSemantics reads the scene with the fork's
SceneSemantics, pushes the list through AccessibilityCache and carries a reader's request
back to the control by id, in both renderers. The renderer modules depend on the scene
module for it. The shell tests that read the fork's C and Kotlin follow the new layout:
the macOS reader record is dxc_ax_element in appkit_accessibility.h, and the X11 C record
is no longer written from Kotlin.
The fork's build-skiko-awt.sh publishes org.thisisthepy.compose.skiko:skiko-awt
as 1.11.1-ext.2. SKIKO_AWT_AS, the desktop module, the artifact check, the
native build scripts, workflow cache paths and shell tests follow it.
The previous change left broken YAML in both workflows (a stray path line, and the
comment landed in the NIK cache step). Restore both steps and add a marker file under
~/.m2/repository/org so upload-artifact keeps org as the root and the download restores
the org/thisisthepy/compose prefix.
…image class path

The fork lists JetBrains' skiko-awt as a second capability of every variant of its
own skiko-awt. Amper takes a variant only when its capabilities are exactly the
module's own coordinate, so it found none and stopped with "Platform is not
supported by the library" at the GraalVM Linux build. publish-skiko-awt.sh drops
the extra capability from the published copy. The fork's Compose modules still ask
for org.jetbrains.skiko:skiko, so the native build scripts now filter JetBrains'
skiko-awt off the class path and the fork's is the only one the image takes.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Moves WINDOW_REVISION from the backport branch (aed8307) to extended
7dd0d4b, which has the caption fix (fork #172). The wheel fix (fork #170) is
not merged yet, so win32-wheel-direction fails until the pin moves once more.

The window layout tests follow the fork: macOS describes elements in its own
dxc_ax_element (checked by the fork's appkit_window_test.m), X11 answers
readers over AT-SPI from Kotlin, and the AppKit accessibility functions live
in appkit_accessibility.h.

Still to port: the renderer's Linux and GraalVM accessibility glue, which
uses AtspiSemanticsSource, X11Accessibility and graalvm.macos.AccessibleElement,
all removed by the fork (d03420c57).

Refs #355
…ty bridges

The fork removed AtspiSemanticsSource, X11Accessibility and graalvm.macos.AccessibleElement.
The renderer now flattens the scene's semantics with ids, parents, states, actions and values in
window points, gives them to the window with setAccessibility, and carries a reader's request
back from onAccessibilityAction to the control (macOS GraalVM, X11 GraalVM, Kotlin/Native X11).
@b-re-w
b-re-w marked this pull request as ready for review October 26, 2026 08:41
@b-re-w b-re-w added the run-heavy-ci Run the heavy CI workflows on this pull request's head label Oct 26, 2026
@b-re-w

b-re-w commented Oct 26, 2026

Copy link
Copy Markdown
Member Author

Heavy CI finding: the fork head publishes skiko-awt as org.thisisthepy.compose.skiko:skiko-awt:1.11.1-ext.2 and the window modules as 1.11.1-ext.2 (fork 53dd7b0be), while the renderer still asks for skiko-awt 0.144.6.1, window modules 0.1.0 and Compose 1.11.1-ext.1. The 'Build the Compose fork's desktop modules' job fails on that mismatch. Fixing it means adopting the ext.2 coordinates for all three pins, which is the scope of #277 (red). Waiting for the owner's decision on #277 versus a narrower path.

@b-re-w b-re-w removed the run-heavy-ci Run the heavy CI workflows on this pull request's head label Oct 26, 2026
…at one extended commit

Merges the coordinate moves of #277 (Compose, material3, window modules, skiko-awt) and keeps this
branch's accessibility glue. REVISION, WINDOW_REVISION and MINGW_REVISION are now ea24655; the change
lists are regenerated from it.
@b-re-w b-re-w added the run-heavy-ci Run the heavy CI workflows on this pull request's head label Oct 26, 2026
…ork's coordinates

The fork rejects versions JetBrains could publish, so the Windows build can no longer publish Compose
as org.jetbrains.compose 1.11.1. The windows module now asks for org.thisisthepy.compose mingwx64
artifacts at the pin's version, and the build script publishes them like every other target.
… native/macos, retire the old-build equivalence failure

The fork lists JetBrains' skiko-awt as a second capability of its skiko-awt, which Amper cannot
resolve; publish-skiko-awt.sh drops it and the native build scripts keep JetBrains' copy off the
image class path (from #277). publish-window.sh publishes scene whenever native/macos is built, for
the platforms whose Compose is present. The equivalence comparison with the old patched build no
longer fails the job now that the pin is on the ext.2 line.
@b-re-w b-re-w removed the run-heavy-ci Run the heavy CI workflows on this pull request's head label Oct 26, 2026
@b-re-w b-re-w added the run-heavy-ci Run the heavy CI workflows on this pull request's head label Oct 26, 2026
@b-re-w b-re-w added run-heavy-ci Run the heavy CI workflows on this pull request's head and removed run-heavy-ci Run the heavy CI workflows on this pull request's head labels Oct 27, 2026
@b-re-w
b-re-w merged commit 9f1ce58 into develop Oct 27, 2026
82 of 93 checks passed
@b-re-w
b-re-w deleted the feat/repin-fork-extended-head branch November 12, 2026 16:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

run-heavy-ci Run the heavy CI workflows on this pull request's head

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant