Files
giza-kmp/docs/dependency-update.md
T

32 KiB
Raw Blame History

Every other dependency: the update the composite decides the order of

quartz-mls-marmot-update.md moved the four libraries the group layer stands on, and moved them yesterday. This plan is for everything else in gradle/libs.versions.toml: the toolchain, Compose, and the eighteen or so libraries around them.

Three findings shape it:

  • Most of it is small, and none of it needs a newer Kotlin. Every library target below asks for a kotlin-stdlib at or under 2.4.10, the version the app compiles with. The libraries can move first, on their own, without touching the toolchain.
  • The toolchain cannot move in Giza alone. A composite build has to settle on one Android Gradle Plugin version across every build that applies it, and Gradle refuses to configure otherwise. That is what 546095b0 ran into in September. AGP 9.4.0 is pinned in Giza, in lightning-kmp-app, and twice in secp256k1-kmp, which is four submodules down. Kotlin 2.4.10 is pinned in all five builds. A toolchain bump is therefore a trip down the submodule chain, the way submodule-refresh.md did it.
  • Material3 has been running against an older Compose than the app's. The pinned material3:1.10.0-alpha05 was compiled against Compose foundation 1.9.3, while the app is on Compose 1.11.1. Every JetBrains material3 since then is published with the Compose version it was built for, and 1.13.0 is the first since 1.9.0 to reach a release candidate.

Researched on 2026-11-20 against 390d4bc3. The composite's pins were read from lightning-kmp-app at 3794eb8, the commit Giza's tree points at.

Where the build stands

Being built on claude/dependency-update-plan-1b853a, one commit per step. Each phase records what it turned out to be in the results table below.

phase result
0 the baseline, at 25f14aa4. check green. jvmTest 1,407 run, 0 failed, 0 skipped, 182 classes; testDebugUnitTest 529 run, 0 failed, 65 classes. m3-audit: every budget 0, adaptive APIs 12 against a floor of 12, navigation components 2 against 2. Resolved graphs: 277 coordinates on jvmRuntimeClasspath, 342 on debugRuntimeClasspath, 336 on releaseRuntimeClasspath. The newest versions were as the tables below say. The build's own warnings are AGP 9's seven deprecated opt-out flags in gradle.properties (each "removed in version 10.0"), the obsolete applicationVariants API, Kotlin's deprecation of KMP beside com.android.application, and three Gradle 10 deprecations (by getting, by registering), all of them in the composite's scripts rather than Giza's
1 built. Eight library entries, the sqldelight plugin alias and six version keys removed, with the six commented-out lines in composeApp/build.gradle.kts that named them or Room 2's aliases. All three resolved graphs are identical to Phase 0's, line for line. check green
2a built. serialization 1.13.0, kermit 2.2.0, okio 3.18.2. The graph diff is those three families and nothing else; serialization moved 1.12.0 → 1.13.0, since quartz had already lifted it past the declared 1.11.0. kermit 2.2.0 publishes one kermit-android where 2.1.0 had kermit-android-debug. jvmTest 1,407 run, 0 failed, once the IdentitySwitchJvmTest 5 s timeout seen under the first full run (the one giza-wallet-removal.md Phase 3 records) passed three times alone and in a second full run
phase what it is rough size
0 the baselines: the resolved graph, the test counts, the warnings, the newest versions checked again half a day
1 the catalogue's dead entries removed, with no version moved an hour
2 the libraries that touch neither the UI nor the composite: network, serialization, persistence, QR a day
3 Compose Multiplatform and the material3 that was built for it one to two days
4 the Gradle wrapper an hour
5 AGP, down the composite and back up a day, most of it in the submodules
6 Kotlin, down the composite and back up, on the same trip as Phase 5 a day
7 the notes and comments that still name the old versions half a day
8 the checks Linux cannot make: a mac, a device, a desktop run on real profiles a mac session and a device session

Where everything is

Newest stable and newest pre-release are as of 2026-11-20, from the repository each one resolves from. Bold is a version that moves.

The toolchain

what where it is pinned today newest stable newest pre-release target
Gradle gradle/wrapper/gradle-wrapper.properties 9.7.1 9.9.2 — 9.9.2
Android Gradle Plugin agp, and the composite (C2) 9.4.0 9.5.1 9.6.0-alpha04 9.5.1
Kotlin (multiplatform, compose compiler, serialization) kotlin, and the composite 2.4.10 2.4.21 2.5.0-RC 2.4.21
KSP ksp 2.3.6 2.3.14 — 2.3.14
Compose Hot Reload composeHotReload 1.2.0 1.3.0 (published today) — 1.3.0
foojay resolver settings.gradle.kts 1.0.0 1.0.0 — 1.0.0
Magpie settings.gradle.kts 0.2.3 0.2.3 — 0.2.3
Daemon JVM gradle/gradle-daemon-jvm.properties 21 — — 21

Compose

library today newest stable newest pre-release target
Compose Multiplatform (plugin, runtime, foundation, ui, resources, tooling) 1.11.1 1.12.2 1.13.0-rc01 (2026-11-18) 1.13.0, or 1.12.2 (D3)
org.jetbrains.compose.material3:material3 and -adaptive-navigation-suite 1.10.0-alpha05 1.9.0 1.13.0-rc01 1.13.0, or 1.12.0-alpha03 (D3)
org.jetbrains.compose.material3.adaptive:adaptive 1.2.0 1.3.0 1.4.0-alpha01 1.2.0 (see below)
org.jetbrains.compose.material:material-icons-* 1.7.3 1.7.3 — 1.7.3; nothing newer exists
org.jetbrains.androidx.navigation:navigation-compose 2.9.2 2.10.0 — 2.10.0
org.jetbrains.androidx.lifecycle:* 2.11.0 2.11.0 — 2.11.0
androidx.activity:activity-compose 1.13.0 1.13.0 1.14.0-alpha03 1.13.0

Which material3 goes with which Compose, from each material3 pom:

material3 built against Compose its navigation suite asks for adaptive
1.10.0-alpha05 (today) 1.9.3 1.2.0, plus Google's androidx.compose.material3:material3-adaptive-navigation-suite:1.5.0-alpha08
1.12.0-alpha03 1.12.0-beta01 1.2.0
1.13.0-rc01 1.13.0-rc01 1.2.0

The catalogue's comment on material3Adaptive still holds at every target. The navigation suite asks for adaptive 1.2.0 and window-core 1.4.0, and adaptive 1.3.0 would bring in window-core 1.5.0 beside them. Adaptive stays at 1.2.0. Only the version the comment names changes (Phase 7).

From 1.11 onward material3 also depends on material3-window-size-class from the same group. That is a new transitive dependency, and Phase 3 checks that nothing in the app collides with it.

Libraries

library today newest stable target asks for kotlin-stdlib
kotlinx-serialization (json, cbor) 1.11.0 (Gradle already resolves 1.12.0, via quartz) 1.13.0 1.13.0 2.4.10
kotlinx-coroutines 1.11.0 1.11.0 1.11.0 —
kotlinx-datetime 0.8.0 0.8.0 0.8.0 —
io.ktor:* 3.5.2 3.6.1 3.6.1 2.3.21
io.coil-kt.coil3:* 3.5.0 3.6.3 3.6.3 2.4.10
com.squareup.okio:okio 3.18.1 3.18.2 3.18.2 2.1.21
co.touchlab:kermit 2.1.0 2.2.0 2.2.0 2.3.21
io.github.alexzhirkevich:qrose 1.1.2 1.3.1 1.3.1 2.4.0
androidx.room3:* (runtime, compiler, plugin) 3.0.1 3.0.3 3.0.3 2.1.20
androidx.sqlite:sqlite-bundled 2.7.0 2.7.1 2.7.1 2.1.20
androidx.datastore:* 1.2.1 1.3.0 1.3.0 2.2.20
app.cash.sqldelight:* 2.3.2 2.4.1 2.3.2 (D5) 2.3.10

Nothing in the last column is above 2.4.10, so Phase 2 does not wait on Phase 6.

Left alone, and why

what today why
quartz, marmot-kmp, mls-kmp, crypto-kmp 2.20.1, 2.3.0, 1.3.1, 2.1.1 quartz-mls-marmot-update.md owns them, and they are the newest released as of yesterday. mls-kmp 1.4.0 is its Phase 5b
fr.acinq.secp256k1:secp256k1-kmp-jni-android:0.24.0 — substituted by the composite and never resolved; the version is a placeholder
android-compileSdk, -targetSdk, -minSdk 37, 37, 26 not revisited here; a target SDK move is a behaviour change with its own checks
docs/tools/engine-interop/old, docs/tools/old-state-blob old engines pinned to old versions on purpose; they exist to reproduce them

Dead catalogue entries

Nothing in any build script reads these:

entry note
androidx-core-ktx, androidx-appcompat
androidx-testExt-junit, androidx-espresso-core, junit there are no instrumented tests; jvmTest uses kotlin-test-junit
androidx-sqlite-bundled-jvm, androidx-room3-sqlite-wrapper
lightning-kmp-core commented out in composeApp; giza-wallet-removal.md Phase 7 step 3 also drops it
plugins.sqldelight not applied, as the comment in composeApp's plugins {} explains

plugins.androidLibrary is read once, in the root apply false. It stays. Taking it out changes which class loader AGP's library plugin is first loaded in, and there is nothing to gain.

Constraints

C1 — what is on a device today still opens tomorrow. No change in this plan touches the group state. The Magpie four do not move. But three of the moves sit under a profile's files:

  • androidx.datastore 1.2.1 → 1.3.0 reads every *.preferences_pb. That includes the ones lightning-kmp-app's DataStoreManager writes, because Gradle resolves the highest version in the app's graph, and the library's own pin of 1.2.1 loses to the app's.
  • Room 3.0.1 → 3.0.3 is two patches. There is no schema change and no migration. The exported schemas/…/36.json should come out byte-identical, and Phase 2 checks that.
  • androidx.sqlite 2.7.0 → 2.7.1 is the bundled SQLite that opens the database file.

The proof is the one giza-wallet-removal.md Phase 7 names: an Android install over the current build, and a desktop run on an existing data directory, with the same profiles and groups opening. That is Phase 8.

C2 — one AGP across the composite. The error, from 546095b0:

Using multiple versions of the Android Gradle Plugin [9.1.1, 9.4.0] across Gradle builds is
not allowed.

These builds pin it:

build where today
Giza gradle/libs.versions.toml, agp 9.4.0
lightning-kmp-app gradle/libs.versions.toml, agp 9.4.0
secp256k1-kmp build.gradle.kts, a buildscript classpath entry 9.4.0
secp256k1-kmp settings.gradle.kts, a resolutionStrategy that rewrites every com.android plugin request 9.4.0

lightning-kmp and bitcoin-kmp apply no Android plugin. They take part only to carry the submodule pointer.

C3 — one Kotlin across the composite. This is a choice, not something Gradle enforces (D2). Every build pins 2.4.10 today:

build where
Giza kotlin in the catalogue (multiplatform, compose compiler, serialization)
lightning-kmp-app kotlin in its catalogue
lightning-kmp kotlin in its catalogue
bitcoin-kmp kotlin in its catalogue
secp256k1-kmp kotlin("multiplatform") version "2.4.10" in build.gradle.kts
docs/tools/engine-interop/new kotlin("jvm") version "2.4.10", whose comment says to keep it at the app's versions

C4 — green at every commit. Each commit passes:

  • ./gradlew check, which runs m3Audit, brandAudit, walletAudit, marmotAudit and magpieLockAudit;
  • ./gradlew :composeApp:compileDebugKotlinAndroid :composeApp:jvmTest, with the test count equal to Phase 0's;
  • ./gradlew -Pmagpie.frozenLock=true :composeApp:jvmTest once the lock is frozen. None of the moves here may add a lock line. If one does, a library has started asking for a newer quartz or crypto, and the strictly pins will already have failed resolution. That is the intended tripwire.

C5 — iOS is built on a mac or not at all. The isMacOsX gate in composeApp stays. So does the -Xdisable-phases=RemoveRedundantCallsToStaticInitializersPhase workaround for Kotlin 2.4.10's release linker, until a mac links a release binary without it (Phase 6). Every phase that touches iOS-visible code leaves an iOS line for Phase 8.

Decisions

The recommended option is first, and the plan below assumes it.

# question recommended alternative
D1 Libraries or toolchain first? Libraries first (Phases 2–3), then the toolchain (4–6). The libraries need no newer Kotlin and no submodule commit, so each is a single bisectable Giza commit. Toolchain first, the usual order. Here it only puts the most expensive phase, the one that leaves this repository, in front of the cheap ones.
D2 Kotlin in Giza alone, or down the chain as well? Down the chain, on the same trip as AGP. The composite compiles all five builds in one daemon, and keeping them on one compiler is what has been done every time so far. The trip is already being made for AGP, so the extra cost is one line per submodule. Giza alone. A newer compiler reads klibs from an older one, so it would probably work, but nothing here has been run that way.
D3 Compose 1.13.0 or 1.12.2? 1.13.0 with material3 1.13.0, if both are final when Phase 3 starts. Otherwise 1.12.2 with material3 1.12.0-alpha03 now, and a follow-up that changes two lines once 1.13.0 ships. Either way the material3 alpha is retired onto the Compose it was built for. Hold at 1.11.1 until 1.13.0. That keeps a material3 built for Compose 1.9.3 for however long that takes.
D4 Pre-releases? Stable only, except material3, which has had no stable release in JetBrains' line since 1.9.0 and is already an alpha here. Not taken: Kotlin 2.5.0-RC, AGP 9.6.0-alpha, room3 3.1.0-alpha, sqlite 2.8.0-alpha, activity 1.14.0-alpha, core 1.20.0-alpha, adaptive 1.4.0-alpha. Take release candidates where they exist.
D5 SQLDelight? Hold at 2.3.2. Nothing under mobi.giza uses it. The runtime and drivers are there for lightning-kmp-app's Phoenix databases, whose code its own 2.3.2 plugin generates. A 2.4.1 runtime under 2.3.2-generated code is a mismatch for no gain, and the whole dependency goes with wallet-removal Phase 7. Bump the runtime to 2.4.1 and the library's plugin with it. That is a submodule commit for a dependency already on its way out.
D6 One commit per library, or per phase? One per group that moves together: a plugin with its runtime, Compose with material3, AGP with its composite. Each commit is green on its own (C4). One commit per phase. That is fewer commits, but a regression found in Phase 8 then bisects to a commit that moved eight things.
D7 What if wallet-removal Phase 7 lands first? Then Phases 5–6 shrink. The composite becomes bitcoin-kmp → secp256k1-kmp: one AGP pin pair, and two Kotlin pins. Nothing else in this plan changes, so neither plan waits for the other. —

Phase 0 — the baselines

Half a day. No commit.

  1. Check the newest versions again. This plan's tables are from 2026-11-20, and Compose 1.13.0 and Kotlin 2.5.0 are both days from final. For each coordinate, read the maven-metadata.xml from the repository it resolves from:

    • Google Maven: androidx.*, com.android.tools.build:gradle;
    • Maven Central: everything else, including the KSP plugin as com.google.devtools.ksp:symbol-processing-gradle-plugin;
    • the Plugin Portal: Hot Reload, foojay, Magpie;
    • services.gradle.org/versions/current.

    Update the tables before going on. A pairing that has changed, such as a new material3, gets its pom read the way the table above was.

  2. Dump the resolved graph for the configurations that ship:

    ./gradlew :composeApp:dependencies --configuration jvmRuntimeClasspath > build/deps-jvm-before.txt
    
    ./gradlew :composeApp:dependencies --configuration debugRuntimeClasspath > build/deps-android-before.txt
    

    Every later phase diffs against these. A transitive move nobody asked for shows up there first. For example, serialization is already at 1.12.0 through quartz, and stdlib is already at 2.4.21.

  3. Record the counts. jvmTest's run, failed and skipped counts. m3-audit's numbers, including the adaptive floor, which sits exactly at 12. The debug and release APK sizes.

  4. Record the warnings: ./gradlew :composeApp:compileKotlinJvm --warning-mode all, saved beside the dependency dumps. Compose and Gradle deprecations show up here a release before they become errors.

Phase 1 — the dead entries

An hour. One commit.

Delete the entries in Dead catalogue entries, along with the [versions] keys only they referenced: androidx-appcompat, androidx-core, androidx-espresso, androidx-testExt, junit and lightningKmpCore. Also delete the commented-out lines in composeApp/build.gradle.kts that name an alias which no longer exists (the Room 2 ones: androidx.room, room.runtime, room.sqlite.wrapper, room.compiler) or which this phase deletes (lightning.kmp.core).

Check: both dependency dumps are identical to Phase 0's. If they are not, something did read an entry, and it goes back.

Phase 2 — the libraries around the app

A day. Four commits. None of them touches the composite or the toolchain.

2a — serialization and logging. kotlinxSerialization 1.13.0, kermit 2.2.0 and okio 3.18.2.

  • Serialization changes only which version is declared, since Gradle already resolves 1.12.0.
  • Check: jvmTest. The event and CBOR round trips are the ones that would notice.

2b — network and images. ktor 3.6.1 and coilCompose 3.6.3, together, because coil-network-ktor3 is built against a ktor 3 and the two should land at once.

  • lightning-kmp asks for ktor 3.4.1. Gradle gives it 3.6.1, as it already gives it 3.5.2. ktor keeps binary compatibility inside 3.x, and nothing of the node runs any more, but compileDebugKotlinAndroid compiles the library against it, and that is the check.
  • Check: jvmTest's relay and websocket tests, then a desktop run that connects to relays, loads a room, and shows avatars. That run goes in Phase 8 if it cannot be made here.

2c — persistence. room3 3.0.3 (runtime, compiler, plugin), sqlite 2.7.1, datastorePreferences 1.3.0 and ksp 2.3.14, in one commit. Room's compiler runs under KSP, and the four are the stack under every profile (C1).

  • Check: ./gradlew :composeApp:kspKotlinJvm, then confirm that git diff composeApp/schemas is empty. The migration tests in jvmTest must pass, RemoveWalletsMigrationJvmTest among them.
  • Confirm in KSP's release notes that 2.3.14 supports Kotlin 2.4.10 and 2.4.21, so Phase 6 does not have to move it again.
  • An Android install over the current build is Phase 8's first item.

2d — QR codes. qrose 1.1.2 → 1.3.1. This is two minors of a library with a rendering surface.

  • Read its changelog for API changes to the painter the app builds.
  • Check: the share-profile and key-package QR codes still scan from another device. The m3-color-exempt site keeps its pure black on pure white.

Each commit diffs the dependency dumps against the one before it. The diff should show only what that commit named, plus its own transitive dependencies.

Phase 3 — Compose and its material3

One to two days. One commit. Compose and material3 are released as a pair (D3), and a commit that moves only one of them is the skew this phase exists to remove.

Move together:

  • composeMultiplatform (plugin, runtime, foundation, ui, resources, tooling);
  • material3, which also versions material3-adaptive-navigation-suite;
  • navigationCompose 2.10.0;
  • composeHotReload 1.3.0.

material3Adaptive stays at 1.2.0. lifecycle and activity are already current.

What to look at, in this order:

  1. The compile, and the deprecations. Diff --warning-mode all against Phase 0.

    • MaterialTheme.motionScheme has to still be public. CLAUDE.md rests the motion rule on it.
    • compose.desktop.currentOs and compose.desktop.uiTestJUnit4 are the two plugin accessors the build still uses. If either is deprecated, replace it with its coordinate in this commit.
  2. Compose Resources. CLAUDE.md records that it does not unescape \' and does not collapse %%, but does process \n. That is behaviour, not a contract, and it can change in any Compose release. StringCatalogueJvmTest asserts each escape the app depends on, and it must pass unchanged. If an assertion has to change, the strings using that escape change in the same commit.

  3. The audits. m3Audit, which needs at least 12 adaptive API uses (currently exactly 12) and at least 2 navigation components. A renamed adaptive API drops the count and fails the floor. Then ColorSchemeContrastTest, which runs in all six schemes, and the readableContent() measurement test.

  4. The Android graph. Today's navigation suite pulls Google's androidx.compose.material3:material3-adaptive-navigation-suite:1.5.0-alpha08 onto the Android classpath, and the new one should not. Confirm that in the dump diff. Confirm too that material3-window-size-class arrives once, at the new material3 version.

  5. Looking at it. Run the desktop app at a compact, a medium and an expanded width:

    • the home screen's list and detail panes;
    • a pushed screen's back button;
    • the navigation bar against the rail;
    • a dialog, a snackbar, and a screen-state transition.

    material3 jumps two or three alpha lines here, and component defaults move between alphas. Screenshots before and after go in the commit message's notes or the PR.

Phase 4 — Gradle

An hour. One commit.

./gradlew wrapper --gradle-version 9.9.2

Run it twice, so the second run updates the wrapper jar and scripts with the new version's own code.

  • The submodules' wrappers do not matter to a Giza build, because the composite runs on the root build's Gradle. They are left alone.
  • gradle-daemon-jvm.properties should not change, since toolchainVersion stays 21.
  • The configuration cache stays off (gradle.properties) for the reason written there.

This comes before AGP so that AGP 9.5.1's minimum Gradle version, whatever its release notes say it is, is already met.

Check: C4, and diff --warning-mode all against Phase 0. Anything new marked "will fail in Gradle 10" gets a line in this phase's record, not a fix.

Phases 5 and 6 — AGP and Kotlin, down the composite

Two days together, most of it in the submodules. Both versions have to move in every build that pins them (C2, C3), so they share one trip down the chain, bottom up, as submodule-refresh.md did:

level repository branch it tracks AGP edit Kotlin edit
4 kngako/secp256k1-kmp threshold build.gradle.kts classpath and settings.gradle.kts resolutionStrategy → 9.5.1 build.gradle.kts → 2.4.21
3 kngako/bitcoin-kmp threshold pointer only catalogue → 2.4.21, and the pointer
2 kngako/lightning-kmp threshold pointer only catalogue → 2.4.21, and the pointer
1 kngako/lightning-kmp-app master catalogue → 9.5.1 catalogue → 2.4.21, and the pointer
0 Giza giza agp → 9.5.1 kotlin → 2.4.21; engine-interop/new too

In each repository that changes a version, commit AGP and Kotlin separately, so either can be reverted alone. Fast-forward each tracked branch and push before the level above moves its pointer. 546095b0 left two pointers at unpushed commits, and a fresh clone could not resolve them.

Giza gets two commits as well: Phase 5 (AGP, with lightning-kmp-app's pointer at its AGP commit) and Phase 6 (Kotlin, with the pointer at its tip).

If wallet-removal Phase 7 has landed by now, the table is two rows, secp256k1-kmp and bitcoin-kmp (D7).

Phase 5 checks — AGP 9.5.1:

  • The submodules: secp256k1-kmp's :tests:jvmTest, bitcoin-kmp's jvmTest, and lightning-kmp-app's :library:compileKotlinJvm and :library:compileAndroidMain. These are the same checks submodule-refresh.md made at each level.
  • Giza: C4, then assembleDebug and assembleRelease. mergeDexRelease is where duplicate classes have failed before (the SQLDelight comment in composeApp).
  • Read the merged manifest, and diff it against Phase 0's.
  • Every flag under #Android in gradle.properties must still be accepted. Several of them (builtInKotlin=false, newDsl=false, the r8 settings) are AGP 9's opt-outs back to AGP 8 behaviour. Record any that 9.5 now warns about, because the next AGP major removes them (see Out of scope).
  • Android Studio syncs the project. That sync is what failed in September.

Phase 6 checks — Kotlin 2.4.21:

  • The same submodule checks, then C4.
  • The Phase 0 dump should no longer show stdlib resolving above the declared Kotlin. The skew quartz-mls-marmot-update.md recorded (declared 2.4.10, resolved 2.4.21) is gone.
  • docs/tools/engine-interop/new builds, and its test passes.
  • Leave on a mac: link a release iOS framework without the -Xdisable-phases=RemoveRedundantCallsToStaticInitializersPhase workaround. If it links, remove the workaround and its comment in a commit of their own. If it does not, keep the workaround, and change the comment's "Kotlin 2.4.10" to the versions it has been seen on.

Phase 7 — the notes that still name the old versions

Half a day. One commit.

git grep -nE '2\.4\.10|9\.4\.0|1\.11\.1|1\.10\.0-alpha05|9\.7\.1|room3 3\.0\.1' -- ':!docs/retired' ':!docs/tools/engine-interop/old' ':!docs/tools/old-state-blob'

Each hit is one of three things:

  • A comment that states the current version. Update it. Examples are the material3Adaptive comment in the catalogue, the Kotlin/Native comment in composeApp, and the "Kotlin 2.4.10" line in quartz-mls-marmot-update.md's transitive table.
  • A record of what was true then. Leave it. A phase's result in another plan is history, and rewriting it would make that plan wrong about its own past.
  • A table that claims to be current. Update it, or add a Since line the way the wallet removal did, for example jvm-target.md's dependency table.

Add a row for this plan to README.md, and fill this plan's status table with each phase's result.

No new audit. The one guard this plan could want, AGP agreeing across the composite, Gradle already enforces by refusing to configure (C2).

Phase 8 — what Linux cannot check

These are the checks every earlier phase deferred. They can share one session with quartz-mls-marmot-update.md's Phase 6 mac half and Phase 8, since all three need a mac and devices.

  1. Android, over the current build. Install a Phase 7 debug build over a build from 390d4bc3, on a device that has a recovery-phrase profile, an nsec profile and an npub profile, and is in at least one group. Every profile opens. Every group opens, and sends and receives. This is C1: DataStore 1.3.0, Room 3.0.3 and SQLite 2.7.1 on real files.
  2. Desktop, on an existing data directory. The same profiles and groups open. The relays connect (ktor). Avatars load (coil). The window resizes across the three breakpoints.
  3. iOS on a mac. A debug build and a release framework link (C5, Phase 6). The app runs on a device. The keychain-held profiles open.
  4. QR codes from both platforms scan on the other (2d).
  5. A release APK installs, and runs the same as debug.

Until these pass, the status line says "built but for a mac and devices", the way the other plans in this folder do.

Risks, and what each one costs

risk where it shows cost if missed
DataStore 1.3.0 reads a 1.2.1 preferences file differently Phase 8.1 profiles fail to open on update. High, which is why it is C1 and not a footnote
A material3 default changed between alphas: padding, shape, colour role Phase 3.5 screenshots visual regressions. Low cost, and caught by looking
A renamed adaptive API drops the m3Audit floor Phase 3.3 check fails. That is the guard working
Compose Resources changes its escape handling StringCatalogueJvmTest strings render with a stray backslash or a doubled %, a visible bug on every screen using that string
A submodule push is forgotten a fresh clone fails to resolve the pointer broken clones for everyone else. Push before the pointer moves, every level
The Kotlin/Native workaround is still needed at 2.4.21 Phase 6, on a mac none, if the rule is followed: keep it unless a link without it passes
ktor 3.6 changes websocket behaviour under the relay pool jvmTest, then Phase 8.2 relays that connect and then go quiet. Medium, and it looks like a network problem

Out of scope

  • AGP 10 and the KMP application module. The AGP 9 opt-out flags in gradle.properties (android.builtInKotlin=false, android.newDsl=false, and the others) exist because composeApp applies com.android.application beside the Kotlin Multiplatform plugin. AGP's next major is expected to drop those opt-outs. That would mean splitting the Android entry point into its own module, with composeApp as a com.android.kotlin.multiplatform.library. It is a restructure, not an update. It needs its own plan, and no AGP 10 exists yet: the newest is 9.6.0-alpha04.
  • Kotlin 2.5.0. It is RC as of 2026-11-11. Once it is stable it takes the same route as Phase 6, as its own change.
  • The Magpie four. These are quartz, marmot-kmp, mls-kmp and crypto-kmp, under the rules of quartz-mls-marmot-update.md: marmot's pins first, each step past them its own commit, and lock lines for every target.
  • The composite's own libraries. lightning-kmp asks for ktor 3.4.1, coroutines 1.10.2 and serialization 1.9.0, and lightning-kmp-app asks for okio 3.17.0. Gradle resolves the highest version in Giza's graph, so Giza's catalogue already decides them. Only the plugins have to agree (C2, C3).
  • SQLDelight and lightning-kmp-core. These go with wallet-removal Phase 7 (D5).
  • Material icons. 1.7.3 is the last version published, and no newer artifact exists in either JetBrains' line or Google's. Moving to vendored icons or Material Symbols is a design change, not a version bump.
  • Re-enabling the configuration cache. It is tied to secp256k1-kmp's two tasks, not to any version here.

Where the numbers come from

  • Newest versions: each coordinate's maven-metadata.xml, on 2026-11-20, read from:

    • Google Maven (dl.google.com/android/maven2);
    • Maven Central (repo1.maven.org/maven2);
    • the Gradle Plugin Portal (plugins.gradle.org/m2);
    • and, for Gradle, services.gradle.org/versions/current.

    "Newest stable" excludes alpha, beta, rc, RC and dev versions.

  • Pairings and stdlib floors: each target's .pom. For kotlinx-serialization and okio, which declare stdlib only in Gradle module metadata, the .module file's requires.

  • The composite's pins: lightning-kmp-app at 3794eb8 and the submodules it pins, read in the main checkout.

  • The AGP rule: 546095b0, "build: unify the composite on AGP 9.4.0 so Android Studio can sync". The error text is quoted in its message.

  • The last time these moved: 086b2f4e (2026-08-15, Kotlin, Compose, Room, ktor and others), 546095b0 (2026-09-04, AGP), and 4e939c45 (2026-09-04, Gradle 9.7.1, carried up from the chain).