Repository navigation
Chore: Integrate open PRs (2026-11-07) - #505
Merged
Merged
Conversation
… app metadata dioxus-compose-msix stages a package directory from one executable (or a directory with the renderer beside it): AppxManifest.xml for a full-trust desktop application, the four tile and logo images drawn from the app icon, and a four-part version mapped from Cargo.toml (revision 0 for the Store, a growing build number for App Installer). Identity, publisher, name and description come from Dioxus.toml, with [windows.msix] holding the values Partner Center assigns. On Windows it packs .msix and .msixbundle with MakeAppx and can write an App Installer update feed. Tests are named for the proposed update packaging requirement, which is still a draft.
The workflow builds the packager and a small windowed stand-in, packages it with the notepad sample's metadata, signs it with a throwaway certificate, installs it, and runs the Windows App Certification Kit against the package and the bundle, uploading both reports. A second job publishes two builds to an App Installer feed on a file share and checks that the installed package sees and takes the update. A dispatch can package a real build from another run's artifact instead of the stand-in.
…nifest The certification kit reads DPI awareness from the embedded manifest and warned that the stand-in was not DPI aware. The build script shows the two linker arguments an application uses to embed the same manifest.
The file share the feed named was not reachable on the runner (0x80070003), so the update test now serves the feed and bundles from a local HTTP server.
Resolved by the Windows package workflow, which runs the packager's build; nothing else in the lock file moved.
dioxus_compose::update says how the running application was delivered (AppImage, Flatpak or neither), reads the AppImage's update information from its .upd_info ELF section, and checks for, applies and restarts into an update through the appimageupdatetool the packager places inside the AppImage. A Flatpak is reported as store-managed and never offered an update. Covers the self-update criteria of the draft packaging requirement (provisionally FR-34, Linux part).
packager-linux reads dx's [application], [bundle] and [linux] tables plus a [linux.store] table for what store listings need, and writes the desktop entry, AppStream metainfo, hicolor icons, an AppDir carrying appimageupdatetool, AppImage update information (gh-releases-zsync or zsync) and a flatpak-builder manifest (prebuilt archive or an offline build from source with vendored crates). The demo crate is the small program the packaging workflow builds in two versions to show each channel updating an installed build. Covers the metadata criteria of the draft packaging requirement (provisionally FR-34, Linux part).
appimage.sh fetches pinned appimagetool, runtime and appimageupdatetool, assembles the AppDir and writes the AppImage with its .zsync. flatpak.sh vendors crates with a pinned flatpak-cargo-generator, writes the manifest, and builds and lints it with Flathub's org.flatpak.Builder. The workflow shows an AppImage 1.0.0 updating itself to 1.0.1 from .zsync files served by nginx on the runner, refuses a 1.0.2 signed by another key, updates an installed Flatpak with flatpak update, and starts the released calculator sample from both packages.
A workflow that is only dispatchable cannot be dispatched from a branch until it reaches the default branch.
Windows fetches the bundle a feed names in byte ranges and refused Python's plain server (0x80D05011).
tools/packager/package-macos makes Name.app from the executable cargo built, with Info.plist taken from the project's Cargo.toml and Dioxus.toml, and signs, wraps and publishes it for one of two channels: - sparkle: hardened runtime, no sandbox, Sparkle.framework inside, SUFeedURL and SUPublicEDKey in Info.plist; zip archive, notarytool submission and stapling, and an appcast whose EdDSA signature is checked against the key the application carries before it is written. - app-store: sandboxed with only the declared capabilities, export compliance answered, no Sparkle anywhere (framework, SU keys, or a host binary that names it), and the productbuild package App Store Connect takes. Signing goes inside out without --deep. Ed25519 is implemented after RFC 8032 so keys need no keychain and Linux can sign; signatures are identical to Sparkle's sign_update, which is preferred when given. Covers the draft acceptance criteria 1 to 9 of the proposed FR-34 (update and store packaging, macOS part), which is not yet in the SPEC.
dioxus-compose-update loads Sparkle.framework from the bundle's Contents/Frameworks at run time and creates SPUStandardUpdaterController on the main thread through objc2, where the renderer's C entry runs [NSApp run]. The application calls start() before launch and check_for_updates() from a menu item or button. Install::Immediately supplies an updater delegate that installs a silently downloaded update at once and relaunches. The Host rather than the Renderer, because Sparkle is Objective-C and the Rust side reaches it directly, while the native-image Renderer would need hand-written glue; nothing crosses the boundary. Sparkle is not linked, so the crate builds anywhere and is a no-op off macOS. A Mac App Store build leaves the dependency out, and the packager refuses a store bundle that names Sparkle. Covers draft criterion 10 of the proposed FR-34.
experiments/self-update builds a small dioxus-compose application at 1.0.0 and 2.0.0 against the published crate (so the released renderer is fetched, not built), packages both for the sparkle channel signed ad hoc, serves an appcast from 127.0.0.1, and waits for the installed 1.0.0 to replace itself with 2.0.0: once installing immediately and relaunching, once installing on quit. It then builds the same application without the updater for the app-store channel, makes the .pkg, checks nothing of Sparkle is in it, and runs it sandboxed. The macOS packaging workflow runs the packager tests on macOS and Linux, the updater crate's tests and lints, and the rehearsal, uploading its evidence. Covers draft criterion 11 of the proposed FR-34.
A Flathub build runs cargo with --locked and no network, so the lock file has to name every crate before the build does.
…et 26.08 The freedesktop SDK has readelf but not patchelf, so a manifest that has to point the executable at the renderer beside it now builds patchelf in a module whose files are cleaned out of the application. The renderer's lib/ directory is flattened beside the executable, the layout it finds its companions from. The default runtime branch is 26.08, the newest on Flathub. flatpak.sh hands flatpak-builder an absolute manifest path, the demo no longer names a screenshot that does not exist (appimagetool fails on the validator's warning), and the AppImage job also runs on aarch64.
The sample release's Linux rows wrap every staged sample with tools/packager-linux/appimage.sh, embedding gh-releases-zsync update information for this repository and publishing the .zsync beside the AppImage. They are signed when APPIMAGE_SIGNING_KEY is set and the run warns when it is not. A sample that cannot be wrapped keeps its archive. Not run here: this workflow builds the renderer and publishes a release, so it is verified at the next sample tag. The script it calls is the one packager-linux.yml exercises on x86_64 and aarch64.
…d-in The package job now runs twice: once on the stand-in and once on a real build (the notepad archive from a published sample release, or a run's artifact when dispatched with a run id), packaged as a directory with the renderer and its companion files beside the program. Builds use the committed lock file.
make-sample-icons.py renders the same marks at every size in SIZES. The 64 pixel icon.png every sample embeds for its window is byte for byte what it was; icon-256.png is new, because Flathub refuses icons under 128 pixels and an upscaled 64 pixel one looks it. The sample release's AppImages carry both.
The released calculator sample names libdioxus_compose_renderer.so with no RUNPATH, so inside the AppImage it exited with status 127. AppRun now puts the payload directory on LD_LIBRARY_PATH, and a Flatpak built from a prebuilt archive sets the rpath to $ORIGIN with patchelf when the executable has none. Also: prebuilt and renderer archives can be given once per architecture (only-arches), so one manifest serves x86_64 and aarch64 as Flathub builds both; --app-id overrides the identifier for a listing under another ID; and the tests are named after FR-35, the packaging requirement's number (FR-34 belongs to another draft).
org.flatpak.Builder reaches the host's flatpak through the session bus, which a runner does not have, so installing build dependencies failed. Every builder call now runs under dbus-run-session. The workflow generates the Flathub submission for a placeholder ID, lints it and uploads it; FLATHUB.md is the checklist for submitting it under the real ID.
The released notepad sample fails the kit's optional Blocked executables test (the renderer DLL references CreateProcessW, and ICU data matches blocked names), which leaves the overall result at WARNING. The script annotated it as an error while passing the step. It now reports optional failures as warnings, lists them in the summary, and fails the step on any required failure.
From the command line the kit inspects a full-trust package without launching it, which the script's description claimed it did.
The first published dioxus-compose links the renderer as @rpath/... and adds no rpath to the application, so the packager cannot find the renderer from the executable. --renderer-dir names it; without it the packager says so instead of building a bundle with no renderer. The rehearsal application turns on the published crate's native-renderer feature (off by default in 0.0.0) and the script fetches and checks the released renderer it needs.
The update packaging requirement is FR-35, not FR-34.
…-35) The certification kit reads DPI awareness from the executable's embedded manifest and reported every application as not DPI aware, because the renderer only declares it at run time. The crate's build script now writes the manifest as a compiled resource file and names it as a native library, which, unlike a linker argument, reaches the final link of every program that depends on the crate; the MSVC linker recognises it as a resource by its contents. The manifest declares PerMonitorV2, the awareness the renderer asks for, and asInvoker. On by default through the windows-manifest feature. An application that embeds its own manifest leaves the feature out, since two manifest resources fail the link.
…3D made the maximum
…to WeakReference in the macOS sheets
…se has another schema
…re the builder looks
…he SDK probe in the user installation
… installed runtimes
This was referenced Nov 7, 2026
This was referenced Nov 14, 2026
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.
Integration branch for the open PRs of 2026-11-07 (one branch per repo, owner instruction). The original PRs are still open; the leader closes them after review.
Issues
Closes #98
Closes #100
Closes #101
Closes #105
Closes #401
Closes #404
Closes #409
Closes #476
Refs #483 (stage 1 is #486 plus #503; the other platforms still answer "unavailable"), #278 (FR-39 stack #286, #285, #289, #290 is merged but not built), #4, #66 and #3 (packagers, drafts).
Classification
feat/file-dialog-483developWINDOW_REVISION-> e3088cc. Part of #483.feat/esc-dismiss-476developfeat/release-0-0-7developfeat/file-dialog-stage1developfilteris now mapped back to the position in the list the Host sent (FileDialogSpec.sentPositions), testfr72_result_filter_is_the_position_in_the_list_the_host_sent. Refs #483.feat/native-skip-macosdevelopfeat/samples-releaseddevelopreleasemainfeat/all-keys-100developfeat/fr45-focus-keyup-98developfeat/fr53-linux-drop-c-crdevelopfeat/fr53-linux-zoom-c-crdevelopfeat/fr53-linux-color-c-crdevelopfeat/fr53-linux-live-size-c-crdevelopfeat/resize-sync-knfeat/resize-sync-spikespike/resize-sync-measuredevelopfeat/resize-sync-specfeat/resize-probe-discriminationfeat/resize-sync-spikefeat/resize-probe-discriminationfeat/resize-probe-discriminationdevelopfeat/fr39-call-macro-spansdevelopfeat/fr39-call-macrosdevelopfeat/composable-attribute-namedevelop#[composable]; renamed to#[Composable]in a separate Fix commit. Stray tracked filecompose-rust/tests/pointer_events.rs-E(a sed backup on develop) still says#[composable]; it is not compiled.feat/fr39-copy-statedevelopWindowAmbient(added on develop after the PR) now releases its arena state on drop and keeps it out of any openremembercapture. Known risk from the PR body:mutable_state_ofneedsT: Send; develop code added since may still.clone()states.feat/sample-release-packagingdevelopfeat/fr41-3-3d-transformdevelopfeat/span-style-helperdevelopspanStylenow takes the resolved family instead of the font list.feat/win32-window-from-forkdevelopfeat/key-handler-before-focusabledevelopfeat/win32-resize-syncfeat/win-linux-window-defaultfeat/shared-window-logicdevelopfeat/remove-awt-shipped-pathdevelopfeat/win-linux-window-defaultdevelopfeature/window-chrome-to-forkdevelopfeature/samples-completedevelopfeature/packager-windowsdevelopfeature/packager-macosdevelopfeature/packager-linuxdevelopfeat/web-direct-calldevelopcargo run --bin codegen, and codegen.rs conflicts heavily with develop. Needs a builder; the integration PR's "Generated files are current" artifact can supply the regenerated files once the code merges.Conflicts and how they were resolved
schema.rsand the three generated Kotlin copies merged by hand in tag order.schema-hash.txt,vectors.jsonhash andSCHEMA_HASHinProtocol.gen.ktwere left at develop's value: they are stale until the generated files are refreshed (see below).HostTextField.ktcombines the Escape-during-composition rule from develop with the new focus-gained and key-up dispatch.events.binand the vector description incodegen.rs/vectors.jsonwere rebuilt by hand: Feat: Send every key to the Host (FR-12.1) #461's 969 byte file plus develop's 73 byteFileDialogResultrecord at offset 969 (file is now 1042 bytes, the Kotlin vector test says so).KeyUptakes theKeyDownlayout (32 byte record, repeat, code, text) so both stay one payload type in the generator; Rust decode/encode,host.rs, the KotlinKeyUpdata class anddispatchKeyUpfollow. Key handlers see key-ups first through the key-down dispatch so repeat tracking stays right.test-samples.ymlmerge (restore-only Maven cache from develop plus the released-renderer condition; the save step got the same condition).LinuxWindow.kt,LinuxWindowBackend.kt,RenderNode.kt, test imports; zoom, colour scheme and drop targets all kept. Feat: Linux live size hints (needs fork PR #179 pin first) (FR-53, #407) #449 was left out (blocked on fork PR Docs: Right-click menu defaults to the OS menu, an app override replaces it (FR-33.1) #179).#[Composable], kept),state.rs(WindowAmbientowns and frees its state). Tests added on develop after Refactor: Rename #[composable] to #[Composable] (#278) #286 was opened were renamed from#[composable]to#[Composable].RenderNode.kt(key handlers moved ahead of the node modifiers; key-up handler and focus-target logic combined) andKeyEventTest.kt.TextResolution.ktspan style helper takes the resolved font family.Cargo.lockmerged package by package (a script that keeps ours and adds the packages and workspace members theirs introduces; no cargo was run);Cargo.tomlmembers union;test-samples.ymlhand merged with Feat: Samples build against the released renderer #466 (see its row).compose-rust/android-kotlin/mirror was refreshed with./scripts/stage-android-kotlin.shafter the merges that touch shared Kotlin.Generated files that may be stale
No cargo or gradle was run. These are stale or hand merged and are checked by the "Generated files are current" job, which uploads the regenerated copies as an artifact:
compose-rust/schema-hash.txt,compose-rust/tests/vectors/vectors.json(hash line),compose-rust/tests/vectors/events.binandmutations.bin,renderer/desktop/src/protocol/Protocol.gen.ktand its Android mirror (SCHEMA_HASH, KeyUp/FileDialogResult/Transform3D parts merged by hand),compose-rust/src/boundary_wasm.gen.rs,renderer/web/src/bridge/HostBridge.gen.ktand the webcompose-rust-host.gen.mjsfiles (untouched, may need the new events),Cargo.lock(merged by script).v0.0.7 items
All present after this PR: #455, #473 and #479 (via #489), #458, GutterBar (#477), Windows K/N static (#465, #333), #467 and #468, #476 and #502, #480.
Not integrated
#449 (fork PR #179 open), #84 (needs codegen run), #268, #170, #155, #135, #134, #118, #117 (stale, branches kept), #490 and #463 (release PRs), #354, #365, #366, #367, #368 (parked). The classification of each is in the table below and is not final: the leader decides.