Repository navigation
Conversation
…gth (PR-2.1) The measure request record grew from 72 to 80 bytes when the font features went into it (#788, FR-83). The smoke host writes its requests by hand in C and kept 72, so the renderer read the second, third and fourth request of the round trip eight, sixteen and twenty-four bytes early, and the smoke test failed on every platform with "the measure round trip did not agree, or never ran" (develop CI run 43617787436: iOS simulator, GraalVM Linux and macOS, Windows). MEASURE_RECORD is 80. The records are zeroed before they are filled, so the new font features slot reads as none. scripts/tests/smoke-host-measure-record.test.sh holds the C length equal to MEASURE_RECORD_LEN; it fails on the old value.
Develop made ui::Modifier own its values, so it is no longer Copy and NativeViewScope::compose could not move it out of a borrowed scope. It now takes it with std::mem::take, the way every other widget scope does.
…p (FR-81) Runs codegen and stage-android-kotlin.sh over the merged schema: develop's Presentation 206, MaxWidth 207, Separator 208, the Sticky modifier and the larger SetOverlayTokens record, together with this branch's NativeView widget, NativeVisible 190, OnNativeViewFocus 191, AttachNativeView 37, FocusNativeView 38 and the NativeViewParent and NativeViewFocus events. The schema hash moves to 0x524d7cce28ecd4cc. The mutation vector holds 147 records (counted by walking the batch), which is the recordCount codegen states; events.bin is unchanged, so the event offsets the tests use stay where they are.
Since 1.11.1-ext.4 (fork #167) the fork's desktop modules (ui-desktop, foundation-desktop and the rest) name org.thisisthepy.compose.skiko: skiko-awt at the fork version. skiko-awt is a JVM jar published as a root alone, with no per-target publications, and publish-skiko-awt.sh puts it in ~/.m2. fork-dependencies.sh still asked every root sibling for a publication per served target (skiko-awt-desktop), so require-fork-artifacts.sh desktop refused a complete repository and every renderer build script stopped before the toolchain ran. A root whose metadata has no variant pointing at a target publication ("available-at") is now published when its own files are. A multiplatform root without the target's publication is still refused. fork-dependencies.test.sh covers both; the first case failed before this change with skiko-awt named as missing.
Fix: Write the smoke host's measure requests at the record's real length (PR-2.1)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
…t (FR-81) The picture checks read the window with CGWindowListCreateImage. For this window, whose content view is a vibrant NSVisualEffectView, that call draws the window again when asked instead of returning what the window server composited. The fork measured it putting a native view a step ahead of the frame under it in 61 of 96 images while none of 93 ScreenCaptureKit pictures of the same turn had the two apart, so the scroll test was failing on the capture rather than on the screen. Every check that judges pixels (the animated scroll, the clip by the ancestors, the popover's rounded hole, the later view on top) now reads a ScreenCaptureKit stream of the window's content, the way the fork's NativeViewSameFrameTest does. The scroll test keeps one column of every picture the window server composited during the animation and checks each, as it checked each image before; it now also requires the view to have been seen at three places or more, so pictures of a page standing still cannot pass it, and on failure it lists every picture's column. Without Screen Recording permission, or with no picture of the window, each picture check prints that nothing on the screen was looked at and what was not checked, and the checks that do not look at the screen still run. The pointer test compared NSCursor.currentCursor with assertSame, which failed with both references printed at the same address: Kotlin/Native gave two references to one Objective-C object. It compares the objects' addresses now.
Every view's onGloballyPositioned worked out the region of every view in the window again, so with four views a scrolling frame ran the position, clip and hole computation about ten times (2.5 callbacks a frame, four views each) to send at most one or two changes. Measured in the macOS budget test (Kotlin/Native debug, four views, animated scroll): the coordinator spent 200 to 250 us a frame, of which the window's own setFrame and setVisibleRegion were 73 to 95 us, so 130 to 155 us was this repository's glue redoing views that had not moved. A view's callback now places that view only. Compose calls the callback for every view whose place may have changed, an ancestor's move or resize included, so each view still hears about its own change in the same pass. An overlay's callback and a scale change still work out every view, because they change every view's holes. After the change the coordinator spends 98 to 117 us a frame including the probe's own clock reads, of which the window's calls are 57 to 68 us: the glue's share is about 30 to 50 us. The budget test's frame medians are too noisy run to run to show it alone (before: +0.37, +0.56, -0.04 ms; after: +0.22, +0.28, +0.15, +0.13 ms over a 1.2 to 1.5 ms frame); the rest of the cost is Compose dispatching onGloballyPositioned (about 0.07 ms against plain boxes) and the window's place(), which rebuilds the container frames and mask on every call. fr81_view_is_clipped_when_an_ancestor_shrinks_around_it checks the assumption this rests on: a view that does not move but whose ancestor shrinks around it is cut to the new clip from its own callback.
…-layout.test.sh Since the fork's event record gained the event time (fork #214, in the pin since 1.11.1-ext.3), its Kotlin gives the record's size as EVENT_TIME_OFFSET + 8. The test still looked for the old form, the text offset plus the text length, so its Linux half failed with "fail: event size" on every run (develop CI run 43617787436, ubuntu). macOS skips that half, which is why no local gate saw it. The probe now also prints the time field's offset. The test checks that the Kotlin names that offset, reads the time there, and that C's size of the record is the offset plus eight. Run in a Linux container against the pinned fork: ok.
Test: Check the X11 event's size through its time field in x11-window-layout.test.sh
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
…-windows-edit-menu
The build job built Compose from the fork on every run, about 63 minutes on macOS. Its output is the local Maven repository the build publishes, which the step copies to FORK_REPO; every later step (the old build, the comparison, the renderer build) reads the fork's publication from FORK_REPO and never from the work directory. FORK_REPO is now restored from a cache before the build, the build step is skipped on a hit, and it is saved after a successful build. Key: compose-fork-<os>-<arch>-<matrix.target>-jdk21-<hash of renderer/scripts/build-compose.sh>-legacy-<LEGACY_REF>-equivalent-<EQUIVALENT_REVISION>. build-compose.sh holds every pin that decides the publication as literals (FORK, REVISION, COMPOSE_AS, MATERIAL3_AS) and reads no .changes file, so its hash changes whenever the fork pin does. The two env pins only affect the old build and the comparison but are in the key so a change to them never trusts an older run. The save never runs on a pull request, and runs only from develop or main, since a cache saved on another branch is not visible to develop. scripts/tests/compose-fork-build-cache.test.sh pins all of this.
…fork test With the fork's publication cached, a pin-unchanged run still built Compose twice more: the old patched build and, on Linux, the full Linux set the renderer resolves. Both are now cached the same way. The old build publishes into LEGACY_REPO, which is what the comparison reads. Its script and patches come from the commit LEGACY_REF and nothing else, so the key is compose-legacy-<os>-<arch>-<target>-jdk21-<LEGACY_REF>. Putting the fork's publication back into ~/.m2 moves to a step of its own that always runs, so a restored old build leaves ~/.m2 as a full run does. The Linux set is published over that put-back repository, and its --clean wipes only COMPOSE_WORK, which nothing later reads. ~/.m2/repository is cached as that step leaves it, under compose-fork-linux-<os>-<arch>-linux-jdk21-<hash of build-compose.sh>, and restored after the fork's publication is put back, so the result after a hit matches the result after a full run. Both saves run only off pull requests, from develop or main, after their step succeeded. scripts/tests/compose-fork-build-cache.test.sh pins both.
…142) The menu builds on the fork's 1.11.1-ext.4 window modules, which the merge of feat/compose-ext4 brings: dxc_native_set_edit_menu and win32EditMenuFlags at window pin a13a6289029.
Covers the five missing acceptance criteria tests tracked in #29: - FR-13.1-3 c2: fr13_1_3_highlighted_rust_text_carries_no_literal_colour_on_the_wire - FR-13.1-3 c4: fr13_1_3_code_roles_leave_widget_property_and_modifier_schemas_unchanged - FR-26 c1b: fr26_bold_and_link_runs_encode_distinct_styles_and_only_the_link_carries_a_handler and fr26_bold_and_link_runs_draw_differently_and_only_the_link_reports_a_click - FR-26 c1c: fr26_a_text_with_no_spans_keeps_its_record_size_and_path - FR-36 c6: fr36_a_worker_notification_sender_queues_requests_a_frame_and_rides_the_next_batch
A dispatch ran every job, about five hours, even to check one branch. ci.yml now takes a suites input: all (the default, so a release check is still complete), fast (the Rust, JVM and script tests, the Windows C compile and the GraalVM renderer smoke), or a comma list of job names. A plan job runs scripts/ci-suites.sh on the input and every job runs only on its output; a job that takes the Windows kit brings windows_static_kit with it, and an unknown name fails the plan rather than checking less. Test: scripts/tests/ci-suites.test.sh pins the input and its default, that every job runs on the plan and is named in the script, and the mapping.
test_windows_resize_metrics and test_windows_resize_probe report resize timings and real drags at two scales; they measure and do not gate a merge, yet every dispatch spent hours on them. They are now the measure suite: not in all (still the default and every job that gates), run when asked for, and all,measure runs everything. Suite names can be mixed with job names in a comma list. Test: ci-suites.test.sh pins that all leaves out exactly the measurement jobs, measure is them, and all,measure is every job.
The fork's X11 window numbers F1 to F12 since fork #219, which the ext.4 pin brings. a_key_with_no_number_claims_no_meaning used F1 as its example of a key the window does not number, so on Linux it failed with "Expected <-1>, actual <122>" (batch CI run 43639736378, Test the linux renderer; the only failed test of that job). The example is now Scroll Lock, which the window still leaves unnumbered, and a new test checks both ends of the function row reach the scene as F1 and F12. Not compiled here: the Linux module builds only on Linux.
…43.1) The Renderer reports how many device pixels one dp of a window covers. The event is tag 53, a 20 byte record with one f32 after the header, appended to the event table and to the events vector. A scale of zero, below zero or NaN is a protocol error. The Host keeps the value in a new density module: density() reads it and is 1.0 before the first report, subscribe() calls back on a change and not on a repeat, and a multi-window application has the same per Window node (density_of, subscribe_window, WindowHandle::density). Covers FR-43.1 criteria 1 to 3: fr43_density_event_round_trips, fr43_density_defaults_to_one_before_the_first_event and fr43_density_change_reaches_readers. The generated files are not in this commit: Protocol.gen.kt, schema-hash.txt and tests/vectors are produced by the codegen binary. The schema hash moves and events.bin grows from 1755 to 1775 bytes. Needed by DarkPyonix/dioxus-compose#240.
The content tells the Host the display's scale once before the first frame and again when it changes, under the window's node like the zoom. The value is the screen's own scale with the zoom left out: whoever makes the zoomed Density provides the number it was made from (LocalDisplayScale), so a zoom step is never reported as a density. Covers FR-43.1 criterion 4 (fr43_density_reported_at_start) and an automated stand-in for criterion 5 (fr43_density_reported_when_window_changes_screen); the move between two real monitors stays a check by hand. The Kotlin here needs HostEvent.DensityChanged from a codegen run, and compose-rust/android-kotlin needs restaging.
…roid copy (FR-43.1)
Feat: Report the window's density to the Host (DensityChanged, FR-43.1)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
…r (FR-24.1) FR-24.1 put the handler on the containers and TextField. The other builders could carry Modifier::motion and had no way to hear it settle, though the rsx path and the Renderer take OnMotionFinished on any node (#707). A motion_handler! macro adds the method to a builder with an on_motion_finished field, the way click_handler! adds on_click, and each compose hands the handler to the existing motion_input. LazyList writes the method by hand, as TextField does, because the macro is not in scope there. Tests cover FR-24.1 per builder family: the handler property is sent for the node, and a MotionFinished report reaches that builder's closure once with the size it carries.
… criteria (#310) Leader-approved, provisional (2026-11-15).
…(FR-56) (#789) Leader-approved, provisional (2026-11-15).
Chore: Build the Windows renderer kit in release mode
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
Fix: Run the Windows UI on a thread the Host makes with a 64 MB stack (PR-3)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
The Compose, window and Windows pins all move to 96ae925f7fb6a730d7af31c2c29c856f35778d25, the tip of the fork's extended branch and the commit its v1.11.1-ext.5 tag goes on. It publishes every line as 1.11.1-ext.5, so COMPOSE_AS, MATERIAL3_AS and SKIKO_AWT_AS move with it, and every live coordinate follows: the renderer module.yaml files, COMPOSE_VERSION in Transform3DLayers.kt and its staged Android copy (stage-android-kotlin.sh), the troubleshooting guides in English and Korean, the scripts, the test comments and the workflow comment. ext.5 is the first tag since the ext.2 build that changes Compose itself, in foundation (fork #240). The Kotlin/Native Windows and Linux scroll configs answer the wheel with the sign and the line desktop Compose uses on the same system; before, the wheel scrolled the wrong way there, and on Windows three times too far. The macOS config works out a scroll with no NSEvent from its travel, 10 dp a line, where it came to zero. The scene sends a scroll where the pointer last was (ScenePointer) instead of the top left corner, and the macOS window places the pointer before a scroll that came before any move. Together they make the wheel and the trackpad scroll on the Kotlin/Native macOS window, where nothing scrolled. The tag also holds what the window pin carried past ext.4 (fork #232, #235, #239, #236, #234, #237) and the AT-SPI readback tool matching roles by number (fork #241). No glue changes: the renderer calls none of the scene's input functions (its Linux and Windows windows go through its own WindowEvents.kt, which already scrolls at the last pointer place, and the macOS renderer goes through the fork's scene), and the C symbols the desktops link are unchanged (desktop-c-symbols.test.sh). compose-fork.changes and compose-fork-mingw.changes are regenerated from 73ac849 to the new pin. No path is dropped; the new paths are the foundation scroll configs and their tests, the fork's wheel and native view tests, two fork workflows and its AGENTS.md. The historical CHANGELOG entry for ext.4 keeps its version and ext.5 gets its own. The tag and the Maven Central artifacts do not exist yet; this was built and checked against 1.11.1-ext.5 published into ~/.m2 from the same commit.
…trackpad WheelScrollWindowTest opens a Kotlin/Native macOS window the way the renderer does (MacosWindow with ComposeRustContent over a scripted Host), with a box that does not scroll at the top left corner and a ScrollColumn and a LazyColumn beside it, each started in the middle of its content. It moves the pointer over one container and hands the view under it AppKit scroll events made with CGEventCreateScrollWheelEvent2: -3 and +3 lines, and -40 and +40 pixels. It reads where each container is from its ScrollChanged reports and asserts that -3 lines moves the one under the pointer 3 lines of 10 points toward its end, +3 brings it back, 40 trackpad points move it 40 points, and the other container does not move. It would have caught the bug ext.5 fixes (fork #240): against the 1.11.1-ext.4 pins all three tests fail with the container moved by 0.0 pixels, because every scroll reached the scene at the window's corner and its distance came to zero. Covers FR-13.6 (scroll containers) for pointer input on the macOS window.
LinuxWheelScrollTest opens a window through the Linux backend with the same page as the macOS wheel test: a box at the corner, a ScrollColumn and a LazyColumn started in the middle of their content. Another display client sends the window a MotionNotify over a container, then a button 5 press and release (a notch toward the reader) and a button 4 one, with XSendEvent. The container under the pointer must move toward its end by three lines of the square root of its height in pixels and then come back, and the other one must stay. Against 1.11.1-ext.4 the notch down would move the container toward the top, by 64 dp a line, which is the wrong-way wheel fork #240 fixes. Runs under Xvfb in the Kotlin/Native desktop renderer test (linux); it cannot run or compile on a Mac. Covers FR-13.6.
The Windows renderer carries its real-window checks itself (Fr53WindowsCheck.kt), because a Kotlin/Native test executable for mingwX64 cannot link the MSVC-built window. fr13_windows_a_wheel_turned_forward_scrolls_toward_the_top joins them: a window with a box at the corner and a ScrollColumn and a LazyColumn started in their middle, the real pointer moved over a container and a WM_MOUSEMOVE posted, then a WM_MOUSEWHEEL of +120 (forward, away from the reader) and one of -120. The container must move toward the top by the system's wheel lines of a twentieth of its height and come back; the other must stay. Against 1.11.1-ext.4 the forward notch moves the container toward its end, three times as far, which is what fork #240 fixes. check-windows-consumer.sh and fr53-windows-backend.test.sh read every `fr<N>_windows_*` name from the Kotlin instead of only `fr53_windows_*`, so the new check is required to print its ok line. It runs in the consumer job of the Windows static renderer workflow (windows_static_consumer in ci.yml). Covers FR-13.6.
…t it on a CI runner (FR-81) On the shared macOS runner the test stopped at its own precondition, "frames drawn: 19": three passes of an animated scroll give thirty frames on a developer's display and not on a runner that draws a few frames a second (batch run 43755840225, Test the macos renderer). The other eleven NativeViewWindowTest tests pass there. The test now scrolls back and forth until it has thirty frames, up to twenty-four passes, and measures each page three times in turn, taking the lowest of the three medians: other processes only add time, and on a busy developer machine a single median missed the budget two runs in five. On a CI runner (the CI environment variable is "true") it prints both figures with the words "budget not judged on CI" and does not judge them, because a shared runner's frame times swing by more than the tenth the budget allows; too few frames still fails there. The budget itself is unchanged.
Test: Give the native view frame budget frames enough, and only report it on a CI runner (FR-81)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
Criterion 15 says the measured numbers go beside it. Five runs on macOS arm64 at the window pin a187dd1e: 1.34 to 1.43 ms a frame without views, 1.47 to 1.51 ms with four, so 0.08 to 0.14 ms for the four, and 0.04 ms to spare against the allowance in the closest run. The margin is thin, so a later failure is to be read as cost, not noise. Also written down, as the leader accepted them on 2026-11-16: the lowest of three interleaved medians as the method, and that a CI runner prints the two figures without judging them while too few frames still fails there.
Docs: Record the measured cost of four native views beside FR-81
Feat: One right-click text menu on the Windows windows, the system's (FR-33.1)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
…ences, and prove the method catches a cost (FR-81) The budget test took the lowest of three medians for each page and compared the two. On a busy machine it failed at random: one run had medians without views [1.37, 2.03, 2.06] ms and with four [1.85, 2.04, 2.13] ms, and the lowest-of-three paired a lucky fast baseline with no equally fast partner. What the data says (macOS arm64, this machine's usual load): - No order effect. Measuring both halves on the same page (A/A, 20 pairs) put the second half slower in 10 and faster in 10. - No regression since 4a3b31c: 30 pairs over both commits give a cost of 0.166 ms a frame at a 1.92 ms baseline, against an allowance of 0.242 ms. A single pair's difference swings by about 0.3 ms. - A bootstrap of those pairs: the best single pair (with a median baseline guard) fails a fine run 24% of the time over five pairs and misses a +0.20 ms regression 54% of the time; the mean of the paired differences misses it 8% of the time over twelve pairs. The rule now: twelve interleaved pairs; the cost is the mean of the paired differences, the allowance is a tenth of the mean baseline plus 0.05 ms, and the test fails when the cost minus two standard errors is above the allowance. Its resolution is stated: two standard errors are about 0.17 ms at the measured spread, so it reliably catches a cost of about 0.3 ms a frame over the allowance; smaller overruns would need about 65 pairs. Each pair, the frame count and the cost, standard error, allowance and margin are printed; on CI the budget is printed and not judged, as before. fr81_the_frame_budget_method_catches_an_added_cost proves the method fails on a real cost, everywhere including CI: the Host busy-waits on each scroll report, which the renderer sends from the frame clock inside the measured frame time, for as long as the page without views takes (at least 0.5 ms). Five pairs are enough for it. Four runs of NativeViewWindowTest (two under three extra busy loops): cost 0.120/0.092/0.188/0.097 ms, standard error 0.013/0.013/0.067/0.010, allowance 0.217/0.211/0.260/0.186 ms, all inside; the added-cost check over in all four (cost 1.73 to 2.21 ms). Wall time: 102 s for the budget test, 43 s for the method check.
…erences (FR-81) FR-81 criterion 15: the method line now states twelve interleaved pairs, the mean of the paired differences judged with two standard errors, its resolution of about 0.3 ms a frame and the test that proves it fails on a cost. The measured values gain the figures under this machine's usual load (0.166 ms at a 1.92 ms baseline over 30 pairs, allowance 0.242 ms); the quiet-machine figures stay. Leader decision, 2026-11-16.
Test: Judge the native view frame budget by the mean of paired differences (FR-81)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
Both sides added checks to the Windows real-window harness: develop the two text edit menu checks, this branch the wheel check. Both are kept, and the consumer script and its script test now read every fr<N>_windows and fr<N>_<M>_windows name.
Chore: Move the Compose fork to 1.11.1-ext.5, with wheel tests through the real window
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
compose-rust and compose-rust-macros 0.0.11 in the manifests and Cargo.lock, the install lines in the README, its Korean translation and both getting-started pages, and the changelog's Unreleased section becomes v0.0.11. The section opens with the known issues: the gap at the moving edge during a drag resize on Windows 10 and Server 2022 (the fork's window, compose-multiplatform-core-extended#245; found by the resize probe in run 43812830794, the same window code as v0.0.10), and the window size after a live change of display scale, fixed in the fork after this release's pin. develop bb9cddb was checked by run 43812830794: 76 jobs green, the two resize probe jobs red for the gate's own bug and for #245.
Chore: Bump the version to 0.0.11
Generated by .github/scripts/release/sync-release.sh. The internal planning documents and experiments/ listed below are kept on develop only: PROJECT.md AGENTS.md CLAUDE.md experiments docs/INTENT.md docs/SPEC.md docs/platforms.md
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Rebuilt from
developby.github/scripts/release/sync-release.sh.The internal planning documents (
PROJECT.md,AGENTS.md,CLAUDE.md,docs/INTENT.md,docs/SPEC.md) and theexperiments/folder are excluded; the source,README.mdanddocs/guide/are not.Merging this leaves
mainwith exactly the tree the script produced.