Repository navigation
v0.6.0
Added
-
CI now builds and smokes a release-configuration Android player. The two
things a shipping build does that a development build does not — R8
minification and Managed Stripping Level High — were the last build
configuration nothing here had ever run, and both bear on this package
specifically: the C# runtime finds the Java class
com.emindeniz99.quickactions.QuickActionsBridgeby name over JNI, which
a minifier is free to rename or strip, and the package ships nolink.xmlfor
the managed linker to read. A fourthandroid-buildmatrix leg,
2022.3-release, builds Testbed2022 through the new
TestbedBuilder.BuildAndroidPhoneRelease— non-development player, stripping
High,minifyReleaseon — against the keep rule README.md tells consumers
to add:-keep class com.emindeniz99.quickactions.** { *; }in
Assets/Plugins/Android/proguard-user.txt, now checked into all three
testbeds and switched on throughuseCustomProguardFile, because the
documented recipe is the supported configuration and so is the thing worth
testing. The minified APK takes the same aapt2/manifest/dex assertions as
every other leg (its dex class names are exactly what R8 could have renamed)
and then the same API 30 emulator smoke, where a shortcut must register and a
warm and a cold tap must each arrive asPerformed. Nothing in the package
itself changed. Its first run (2026-09-02, PR #19 run 64) was green:
UnityLinker ran the High rule set (global-metadata.dat1.70 → 1.18 MB), R8
collapsed four dex files into one and left one-and-two-letter class names
around the three package classes it had been told to keep, the APK came out
at 14.2 MB against the development build's 23.7 MB, and the smoke registered
three shortcuts and got a warm and a cold tap back asPerformed— the
documented recipe holds, and stripping High needed nolink.xml. -
The built-in Android icons now ship an API 26+ adaptive variant. On API 26+
AOSP does not draw a legacy shortcut drawable as authored — it wraps it onto a
white plate at 0.70 of the viewport, so the built-ins rendered as a small
indigo disc inside a white ring, out of place beside their neighbours. Every
Android build now also writes
res/drawable-anydpi-v26/ic_quickaction_builtin_<name>.xml, an
<adaptive-icon>under the same resource name, over two layers of its own
(ic_quickaction_builtin_<name>_background, full-bleed indigo, and
…_foreground, the same glyph scaled into the 66-of-108 safe zone so no
launcher mask clips it). The plain vector stays for API 25 and the resource
qualifier picks between them, so nothing that names an icon changes: the Java
lookup, the staticandroid:icon="@drawable/ic_quickaction_builtin_<name>"
bake, theic_quickaction_*keep rule, the settings opt-out and the define-off
sweep all cover the new files unchanged. About 6 KB of XML for all four icons.
Not yet seen on a launcher — neither variant. -
The
android-smokelegs now photograph the launcher's long-press sheet.
dumpsys shortcutproves each shortcut registered with an icon resource;
nothing has ever seen that art drawn by a launcher. WithCAPTURE_LONGPRESS=1,
tools~/device-smoke/android_device_smoke.shends — after itsPASS:
verdict — by going home, opening the app drawer, locating the app's icon by
its launcher label in auiautomator dump, long-pressing it, and keeping
longpress.pngplus both hierarchy dumps; CI uploads them per leg as
longpress-<leg>. It asserts nothing and can fail nothing: the gesture and
the sheet belong to whatever launcher the system image ships, so the capture
runs witherrexitoff, bounds everyadbcall withtimeout, prints one
grep-ableshortcut sheet visible: yes/no, and leaves the verdict to the
eight steps above it. Its first run (both images ship the Pixel launcher)
found the API 30 swipe opening no drawer and the API 35 swipe-hold opening
no sheet, so the capture now escalates throughKEYCODE_ALL_APPS, the
launcher's own "Apps list" handle and theALL_APPSintent when the icon
is not found, and presses withinput motioneventDOWN / hold / UP, taking
the screenshot and the hierarchy while the finger is still down. That
second attempt (run 56) photographed the Pixel launcher's "App suggestions"
sheet on API 35: the icon it pressed was the hotseat's prediction of the
app, which never opens the shortcut popup, and on API 30 no opener moved
the launcher off the home screen. A predicted-app match now counts as a
miss, so the drawer escalation runs and tries to find, and then press, a
real icon. Its first run recognised the prediction and escalated, and no
opener moved the API 35 launcher off its home screen either. On the run
after that (61) the API 30 swipe opened the drawer for the first time, the
press landed on the real icon, andlongpress-2022.3holds the first
photograph of the popup: the three static entries and the runtime-added
daily, with the built-inaddandfavoriteart drawn on two of the rows.
The verdict line saidnoon that run — it matched the titles while
Launcher3 draws the long label (theSubtitle) when it fits — so
CAPTURE_TITLESentries are nowTitle=Subtitlepairs, either form counts,
and when nothing was pressed the line says so instead of blaming the
hierarchy. The API 35 drawer opened once the swipe stopped starting on the
hotseat's search bar: the dumps of both images put the old 90% start on a
search widget, so the capture now swipes from 65% of the height first — the
API 35 drawer opened on the first run of that (PR #19 run 64), the press
landed on the real icon and the verdict line readyeswith all three long
labels — and from the bottom edge second, the start API 30 had answered
once. Both images' sheets are on file; the capture stays best-effort and is
never a check. -
The settings page says what an
Icondoes on Android, next to the field.
A property drawer forIconTypeadds one line under the popup: "built-in
drawable" forAdd/Compose/Favorite/Play, or the exact
ic_quickaction_<name>the project must ship for the other 25 — which used
to surface first as a build-log warning. The four names come from a second
generated file,Editor/QuickActionsBuiltInIconSet.cs, written by the same
generator from the same list (the Android assembly holding the art is
UNITY_ANDROID-only, so the always-present Editor assembly needed its own
copy), held to the art by--checkand by harness tests that also pin the
drawer's name derivation to the JavaICON_NAMEStable member by member,
that the note stops promising a drawable when Write built-in Android
icons is off, and that in the static list it says what the baker needs:
a static shortcut bakes no icon for a non-built-in choice unless
AndroidDrawablenames one. The note wraps at narrow inspector widths. -
AGENTS.md— the shortest correct integration path for an AI coding
agent (or a hurried human): install, define, asmdef reference, the guarded
snippet, the checks that prove it worked, and the don'ts. The README gains a
60-second quickstart with the same guarded snippet, a contents list, and an
embedded-package (vendoring) install option. -
CI runs the iOS native hooks under a mock coexistence host. Two new
legs,ios-export-coexandios-simulator-coex, re-export Testbed2022 and
Testbed6 with four sources fromExamples~/Coexistence/iOS/copied into
Assets/Plugins/iOS, so one app is at once anIMPL_APP_CONTROLLER_SUBCLASS
host, a category+loadswizzle of the two app-delegate selectors, and a
GoogleUtilities-style isa proxy — the three shapes a project's own
UnityAppControllersubclass, AppsFlyer / OneSignal-style categories and
Firebase take. The Simulator launch must print everyQA-COEX: PASSline by
name plus the package's own install line; the category+loadrecords
whether it ran before or after the package's class+loadand composes
either way, because the first run showed both — category first on the
2022.3.62f3 export, class first on 6000.3.21f1, every file in
UnityFrameworkboth times — where the README had asserted a class
+loadalways runs first. On the scene testbed a second launch shadows
application:configurationForConnectingSceneSession:options:without
calling super and requires the install line to flip from "via
configuration" to "via notification". Until now theios-simulatorleg
only required the process to still exist, which an app with no hooks
installed satisfies too. The lifecycle each leg exercises is asserted, not
described. First run (2026-09-02): on Testbed6 (6000.3.21f1, scene
manifest, iOS 26 Simulator) every check passed on both launches — the
UIScene path's first exercise anywhere: hooks onUnityScenevia the
configuration wrapper, via the notification fallback when the host shadows
the selector, the cold launch item queued once withNOreturned, the
warm tap once with one completion, the host's discardedNOnot
double-delivering. On Testbed2022 the mock host itself was wrong twice: its
category+loaddemanded an order objc4 does not promise, and its
scene-configuration override opted the manifest-less app into the scene
lifecycle — UIKit calls that selector on any delegate that implements it,
manifest or not, and the app then never became active. Both corrected: the
order is recorded, and the override exists only under a manifest. -
One
NSLogper launch names the branch each iOS hook took.
[QuickActions] iOS hooks: didFinishLaunching=… performAction=… sceneConfig=… manifest=…at the end of+load, and
[QuickActions] iOS scene hooks installed on <class> via configuration(or
via notification) when the scene hooks land — the only way an adopter, or
CI, can tell a working install from an inert one. -
Samples~/AndroidIcons— a working.androidlibicon example. The
custom-Android-icon recipe as an importable plug-in rather than five written
steps: Package Manager ▸ Samples ▸ Import copies a bare
QuickActionIcons.androidlib—AndroidManifest.xmlandres/at its root,
the shape CI measures on the 2022.3 testbed — withic_quickaction_search
andic_quickaction_homedrawn like the built-ins, underAssets/, where
Unity picks it up on the next Android build. Its README says how a runtime
Addand a static item each reach the drawable, how to confirm it in the
APK withaapt2, the Unity 6 (AGP 8)namespacechange, and that the
.unitypackageinstall cannot carry an.androidlib. The folder ships with
its Android-onlyPluginImporter.meta;tools~/gen_meta.pynow treats a
*.androidlibas one plug-in (a.metafor the folder, none for its
contents) in both its generate and--checkpasses.
Changed
-
The Android emulator smoke relaunches an app whose player never came up.
Twice on 2026-09-02 (runs 52 and 65, both on the API 30 image, both on APKs
that passed unchanged the next time) the smoke's launch reached the
foreground and then nothing happened: noVkInstance, not one line under
logcat'sUnitytag, and after the full budgetdumpsys shortcutstill
showed zero calls. That is the emulator's ARM translation stalling the
player, not the package. Step 5 now spends half its budget, and if nothing
has been published and the player has not logged at all, force-stops and
launches once more before spending the other half — loudly, so a run that
needed the second launch says so. A player that did come up and published
nothing is not relaunched: that would be the package's failure to report. -
The Unity 6000.6 canary retries an editor that died before the suite
ran.tests (unity6-latest)has twice (main run 41, PR #16 run 53) ended
with exit 137 at "Begin MonoManager ReloadAssembly" — the editor killed
during its own start-up, no test executed — and started fine on the runs
between. The job now starts the editor once more when, and only when, the
first attempt left noeditmode-results.xml; an attempt that produced
results is final, so a real test failure is never retried, and a verdict
step turns the job red unless the last attempt passed. -
The iOS status claims are scoped to the path that was actually run. The
Simulator run every doc cited was logged as "Unity 6.3" with no patch number
and its exportedInfo.plistwas never inspected, so which of the two iOS
delivery paths carried that tap is unknown — 6000.3.8f1 is where Unity
starts emittingUIApplicationSceneManifestand Apple stops calling the
app-delegate selector. README Status,PRODUCTION_READINESS.md(the one
iOS row becomes two, and the older Simulator run is credited to neither —
what each row claims now comes from the coexistence leg),CLAUDE.md
andROADMAP.mdnow say so, and the ROADMAP's scene items start with
"confirm the app still starts": the package adds
application:configurationForConnectingSceneSession:options:where Unity
has none, so a wrong return value there would mean no engine at all. The
README gains Coexisting with other native iOS plugins: the five
selectors the package touches, why it hooks the root class from a class
+load, and what each surveyed SDK family does — Firebase / GoogleUtilities
isa-swizzling inherits the package's IMPs because it adds no ivars, Firebase
C++ / OneSignal / AppsFlyer category swizzles chain after it, the listener
SDKs (Facebook, Helpshift, Unity IAP) cannot collide, Singular's scene
swizzle composes, and Unity as a Library and the Swift project type are
unsupported. It is a source audit, and says so. -
The custom-asmdef integration shape is compiled by CI in both
configurations. README tells a project whose scripts live in their own
assembly definition to referenceEminDeniz99.QuickActions; with the define
off that assembly is not compiled, so whether the referencing assembly still
builds was an assertion about Unity, not a measurement. Testbed2022 now
carriesAssets/Integration/Testbed.Integration.asmdef, which references the
package and holds the guarded quickstart component; the define-on legs and
the define-offgate-offbuild compile it on every push, and the README
also describes the gated-glue-asmdef alternative. -
The
.androidlibicon recipe is measured on a real APK — and the
measurement corrected it. README's recipe for a consumer's own Android
shortcut icons rested on Unity's documentation and on reading Unity's
com.unity.mobile.notifications, never on a build. Testbed2022 now carries
an.androidlibunderAssets/and one inside an embedded UPM package
(Packages/com.quickactions.testlib/), and the 2022.3 leg of
android-buildrequires their drawables in the APK resource table on every
push. The first run proved the documented layout wrong: a bare.androidlib
(nobuild.gradleof its own) takesAndroidManifest.xmlandres/at its
root — thesrc/main/layout the README prescribed, Unity's own package's
layout only because that package ships abuild.gradle, was silently
ignored. The recipe, its first "trap" and the keep-file example now say so,
and a decoy planted undersrc/main/res/must stay absent from the APK on
every push, so the trap is asserted rather than assumed. -
The Unity 6 Android leg now builds twice and asserts the second APK. CI
had only ever built a clean Gradle project on a fresh runner, so nothing knew
whether the icons,res/raw/quickactions_keep.xml, the baked shortcut
resources and the trampoline<activity>survive an INCREMENTAL build, where
Unity reuses the project it staged underLibrary/. The leg runs
unity-buildera second time over the same project directory
(TestbedBuilder.BuildAndroidPhoneSecond, same configuration to
Builds/QuickActionsDemo-phone-2.apk), uploads that APK as
quickactions-demo-apk-unity6-incremental, and runs the same aapt2
assertions against it — one function over the leg's APKs rather than a copy
per APK. The job summary names which of the two was the incremental one. -
The
.unitypackagecarries the docs the README links to. The classic
install now includesGETTING_STARTED.md,SECURITY.md,
PRODUCTION_READINESS.mdandAGENTS.mdnext to the four root docs it
already shipped, so a reader of the packed README no longer hits dangling
links. The README's usage example is#if QUICKACTIONS_ENABLED-guarded,
says where the settings asset lives, and the API table states that anull
item throws — the one way any call throws. -
CI proves the gate and measures the footprint on every push. A new
gate-offjob builds the 2022.3 testbed withQUICKACTIONS_ENABLEDoff —
an IL2CPP APK inandroid-build's exact configuration and an iOS Simulator
export — and asserts the package left nothing behind: no trampoline
<activity>or shortcuts meta-data in the manifest, noquickactions_*/
qa_*/ic_quickaction_builtin_*resources, nothing of the package's
assembly in the IL2CPP metadata, noQUICKACTIONS_ENABLEDmacro in the.pbxproj, no
markedUIApplicationShortcutItems. Every negative is paired with a positive
control — the define-on artifacts from the same run must trip the same
probes, or the check declares itself blind instead of green. The two APKs
are then diffed entry by entry, plus the whole-file size: that difference is the package's footprint,
printed in the job summary and held under 1 MiB, so an asset or code path
that starts shipping turns the job red.PRODUCTION_READINESS.md's two gate
rows move from "build-proven once, 2026-07-17" to CI.TestbedBuildergains
BuildAndroidPhoneNoDefine/BuildiOSSimulatorNoDefine, and its
DisableDefine/EnableDefinenow flip both mobile targets.The first measurement (run 51) read 1,077,821 bytes and failed the ceiling —
and the entry-by-entry diff said why:libunity.soalone accounted for
1.4 MiB uncompressed, because the define-off build had dropped the IMGUI,
TextRendering, InputLegacy and both TextCore engine modules. Nothing in the
package uses them (its runtime reaches AndroidJNI and JSONSerialize, present
in both APKs); the Demo sample does, and the Demo wrapped its whole file
in#if QUICKACTIONS_ENABLED, so with the define off itsMonoBehaviour
vanished fromAssembly-CSharp, engine-code stripping removed the modules
only it used, and their weight was billed to the package. The Demo now keeps
its component and IMGUI shell compiled in both configurations and guards
only the package calls — the README's own guidance, and what the
Testbed.Integrationcomponent already did — so the same engine modules
stay on both sides of the diff. A new harness config,SampleOff, compiles
that#elsebranch the way a define-off device build does (no define, no
Runtime), which no config did before. The Demo's own namespace,
EminDeniz99.QuickActions.DemoSample, therefore stays in a define-off
build's IL2CPP metadata — as any consumer namespace would — and run 52
flagged exactly that as a leak, because the metadata probe grepped for the
namespace prefix. It now names what only the assembly leaves behind: the
image nameEminDeniz99.QuickActions.dll, theInternalnamespace and the
root namespace'snamespace|Typestrings — 12 lines on the define-on APK,
none on the define-off one — and a leak report prints every metadata line
mentioning the namespace, probe-matching or not. With the Demo on both
sides, run 52's APK pair diffed through the job's own script measures
144,061 bytes (140.7 KiB) compressed:libil2cpp.so+149.7 KiB (arm64)
and +122.0 KiB (armv7) uncompressed,global-metadata.dat+31.4 KiB,
libunity.so+27.0 / +19.9 KiB, and about 10 KB of resources — the
eighteen files the package writes. -
The Demo sample survives a define-off build as a real component. With
QUICKACTIONS_ENABLEDoff,Samples~/Demo/QuickActionsDemo.csused to
compile to nothing, so a define-off build of its scene carried a "missing
script". The component and its on-screen shell now compile either way; with
the define off the buttons report that the package is not compiled instead
of calling it. Define-on behaviour is unchanged — the same catalog, log lines
and autotest hook the device smoke drives.
Fixed
- Both scene-discovery routes bind only to Unity's own scene delegate. The
UISceneWillConnectNotificationfallback learned theUnityScene-ancestry
rule in this release; the configuration wrapper still installed the scene
hooks on whatever class the first connecting session declared. A manifest
may declare more than one role — CarPlay, an external display — and when
such a session connects before Unity's window, that install burned the
process-wide one-shot on the host's class (Unity's scene never hooked) and
addedwindowScene:performActionForShortcutItem:completionHandler:to a
class that lacked it, with the owner still recorded as confirmed. One
shared rule now sits in front of both routes: bind only toUnitySceneor
a subclass when this process has that class (a non-Unity class is logged and
left alone, one-shot intact), and bind to the first declared class with the
owner recorded as unconfirmed when it does not. Found by review; no CI leg
drives a multi-role manifest yet. - A hostile local app can no longer crash the game through the trampoline.
QuickActionsTrampolineActivityis exported (the launcher has to start it),
so any app can start it with arbitrary extras — and reading the action id
unparcels the whole bundle, so aParcelablethis app cannot load threw
BadParcelableExceptionout ofonCreate: a process crash attributed to the
game. The read is now contained — a bundle that cannot be read is logged and
treated as no tap — and the launch-intent lookup sits inside the existing
catch too. A new Java smoke scenario drives the hostile bundle (111 checks). - A failed native read is never mistaken for an empty shortcut set.
QuickActionList.Parsereturnsnullfor a payload the realJsonUtility
rejects; it used to return an empty list, which the facade treats as
authoritative and prunes against, so a serializer failure would have removed
the user's real shortcuts on the next write. The Android bridge's
SetShortcutspropagates that null as a failed write, and the iOS
QABuildShortcutsJsonreturnsNULLinstead of{"items":[]}when
NSJSONSerializationfails. A Unity-only test pins theParsecontract
(the Test Runner suite is 77). - The Android SDK-level read can neither throw out of the facade nor pose
as "API < 25". TheBuild.VERSIONJNI read is guarded like every other
JNI path in the bridge, read once and cached; a read that fails is kept
distinct from an unsupported device — the bridge's read and write members
return their failed-read signal (null) for it, so the facade retries instead
of adopting an empty set and pruning the user's real shortcuts against it. - The
UISceneWillConnectNotificationfallback binds only to Unity's own
scene delegate. It used to bind to the first scene that connected,
whatever it was, and burn its one-shot doing so; on a host that owns the
scene manifest — Unity as a Library, a SwiftUI@mainapp, a second iPad
window, an external-display or CarPlay scene — the real Unity scene was
never hooked, and because the package adds
windowScene:performActionForShortcutItem:completionHandler:to a class
that lacks it, it then looked terminal for the host's own quick actions:
adopted into a queue nothing on their side drains, completedYES. The
fallback now reads the declared class from the connecting session's
UISceneConfiguration(never aGUL_/NSKVONotifying_proxy; set before
the delegate exists) and binds only toUnitySceneor a subclass when this
process has one. With noUnitySceneclass — a host-authored manifest on an
older trampoline — it keeps the first-connecting behaviour but marks the
owner unconfirmed: there it never adopts an unmarked item and completes
NO, which is what UIKit would have seen had the selector never been added.
The configuration-wrapper path is unchanged. - The defensive
class_addMethodtype encoding is built from
@encode(BOOL)rather than the hardcoded armv7-erac(BOOLisbool,
encodingB, on arm64).