Changelog

Copy Markdown View Source

[0.7.0] - 2026-09-11

Added

  • MobDev.Differential.run/3 drives the iOS/Android view-tree comparator against two live device BEAMs (MOB-157). The orchestrator samples both trees via :rpc.call(node, Mob.Test, :view_tree, [node]), then invokes Mob.Differential.compare/3 on one of the devices, iOS by convention. If iOS returns {:badrpc, {:EXIT, {:undef, _}}} (i.e. the device's mob build predates the comparator) the orchestrator falls over to the Android node. If either device has not rendered yet (a root with children: []) the run short-circuits with {:error, :not_ready} before the comparator is called; a :not_ready answer from the comparator itself is not retried on the other node; any other {:error, _} from the comparator surfaces as {:error, {:comparator_error, node, other}} so the documented error taxonomy stays a closed set; if neither node has Mob.Differential loaded (both apps predate the comparator), the run returns {:error, :differential_unavailable}. An error tuple from Mob.Test.view_tree/1 (e.g. Android's {:error, :not_loaded} when the Kotlin bridge predates MobBridge.uiViewTree()) surfaces as {:error, {:tree_error, node, reason}}, distinct from :not_ready. frame_tolerance_dp is forwarded to the comparator; rpc_timeout bounds every RPC (worst-case wall clock is roughly 3 * rpc_timeout because the three round-trips are sequential). Unknown options raise ArgumentError rather than being silently dropped. The RPC module is injectable so callers can stub it in tests.

Fixed

  • Tier-2 projects from mix mob.new_plugin compile again (MOB-168). The generated module documentation embedded one heredoc inside another, which ended the outer string early and left the example markup as invalid Elixir. The same compile check caught a tier-3 list referring to module data as a screen assign; executable navigation coverage then caught a callback that did not match its row-tap messages. Generated tier-3 lists now render and navigate correctly. Scaffold tests compile a real generated project for every tier.

  • mix mob.connect no longer force-quits every app on an attached iPhone (MOB-70). Clearing other Mob apps off the device before a launch is necessary — each one starts an in-process EPMD on 0.0.0.0:4369, so only one can run at a time — but the code decided what to clear by matching every process under Bundle/Application/, which is where all third-party apps live. Plugging in a personal iPhone and running mix mob.connect terminated every app its owner had open, and the except_bundle argument meant to spare the target app was discarded outright.

    Measured against an attached iPhone SE: the old code would have killed 15 user apps, TestFlight among them. It now kills 0 unless mob_dev installed them itself.

    mob_dev records what it installs (MobDev.IOSInstalls, ~/.mob/ios_installs.json) and kills only that. An absent or unreadable record means kill nothing. Behaviour change: a Mob app installed by another route — Xcode, TestFlight, a colleague's build — is no longer cleared, so it will still hold EPMD 4369 and the launch will fail as it did before mob_dev cleared anything. See decisions/2026-09-06-mob-dev-kills-only-what-it-installed.md.

Changed

  • mix mob.deploy rejects unrecognised options instead of ignoring them. A run that passed an extra or misspelled flag used to carry on regardless; it now fails and names the option. If you have a wrapper script, CI step or shell alias passing a flag this task does not accept, it will start failing. It also takes no positional arguments — mix mob.deploy --native ABC123 deployed to every device and said nothing, and now refuses.

Fixed

  • mix mob.deploy hot-pushed over distribution without writing the BEAMs to disk, so changes reverted on the next restart. The dist and filesystem paths were mutually exclusive: when a device answered over dist, the new modules were loaded into the running VM and the on-disk copies were left stale. The app then reverted to the last filesystem deploy whenever it restarted — and mix mob.connect restarts the app, so connecting to inspect your change was enough to undo it. This is the long-standing "mix mob.deploy didn't do anything" report; --native was unaffected because it skips dist entirely. The dist path now hot-loads and writes, without restarting, and a failed write is reported instead of being hidden behind a successful hot load. Verify with mix mob.attest after a restart. Physical iPhones are excluded and get the hot load plus a warning: they are often LAN-only with no devicectl route, and the write is a no-undo replace that a mid-copy interruption would leave unbootable.

  • mix mob.deploy --help and mix mob.mutate --help refused to help. Strict option parsing turned the flag people try first, on the task they run most, into Unrecognized or invalid option(s): --help — telling the reader to go and read the help. Both now defer to MobDev.TaskHelp, as mix mob.uninstall already did.

  • mix mob.uninstall ignored :ios_bundle_id. The task resolved one bundle id for both platforms and passed it down, so the per-platform clause that was supposed to pick :ios_bundle_id for an iOS device never ran. On a device holding the app under a different iOS id, uninstall removed nothing and reported success. The accompanying test passed an opts shape the task never sends, which is why the suite stayed green.

  • mix mob.deploy's error told you the equals form was required for dash-prefixed values. It stopped being true in the same commit that wrote it, once the spaced form was handled.

  • --beam-flags "-S 4:4 -A 4" aborted the deploy. OptionParser will not consume a dash-prefixed argument as a string value, so the spelling this repo prints in seven places — the moduledoc, five README recipes and both battery-bench workflows — parsed as two unknown options. Under the previous lenient parsing the value was silently dropped and the deploy carried on with whatever mob.exs held; strict parsing turned that into a hard failure whose message named --beam-flags itself as unknown. Both spellings work now.

  • The error said "Unknown option(s)" for options that are known. --schedulers abc reported --schedulers as unknown and discarded the value, sending the reader to a help page that lists it. It now distinguishes an unrecognised flag from a bad value and prints both.

Added

  • mix mob.attest — prove the device is running the code you just pushed. Compares module_info(:md5) on the device against :beam_lib.md5/1 of the local .beam, so a push that never landed, landed in the wrong place, or landed and was never loaded all show up. Exits non-zero on a mismatch, and also when the check could not run — a check that could not run is not a check that passed. Modules the device has not loaded yet are reported and are not a failure. Defaults to the exact set mix mob.deploy pushes, so what is attested cannot drift from what was shipped. Written after a bundle-id divergence let a BEAM push succeed against the wrong app's container, printing a tick while the app kept running old code; it found a second live instance (MOB-161) on its first real use.

  • mix mob.mutate — mutation testing for the lines a branch changed. Breaks the production code one line at a time and reports the changes nothing noticed, which is the only cheap way to tell a test that guards something from one that merely runs. Doing this by hand has found real defects every time it has been tried here, and got the baseline comparison wrong once, reporting a surviving mutation as caught. The baseline is always captured by running, never supplied. --file, --base, --test-command, --max, --json; exits non-zero when anything survives.

  • mix mob.deploy --json — a machine-readable result on stdout listing the deployed, failed and skipped devices with their per-device reasons, and an outcome that mirrors the exit status. Progress output is redirected to stderr for the run, so mix mob.deploy --json | jq receives exactly one document. Emitted on the native-build failure path too, which is when a caller most needs it.

  • :ios_bundle_id in mob.exs, for when Android's applicationId is not a legal Apple bundle id. Apple forbids the underscores Android allows, and com.example.* is often already claimed by another Apple team, so the two frequently cannot be the same string. Every iOS path — deploy, connect, provision, battery bench, uninstall, the simulator and device builds, code-signing, and the release IPA — resolves :ios_bundle_id || :bundle_id. Android keeps using :bundle_id.

Changed

  • mix mob.deploy no longer exits 0 after shipping nothing. Four ways it could, all fixed (MOB-150):

    • --android --native with no sdk.dir in android/local.properties printed a skip warning, built nothing, and succeeded. The old rule was ok_count == length(results), which is 0 == 0 for a run that produced no results at all.
    • --ios / --android where every device of that platform was skipped. A skip stays non-fatal when it is incidental — a phone that happens to be attached — and is fatal when the run named that platform. The rule is per platform, not per device: one simulator deploying while a stale one is skipped is a success, not a failure.
    • --device X that reached X and deployed nothing to it, and --device NOPE matching no device at all.
    • --ios on Linux, which resolves to no platforms and enumerated no devices.

    A --native run that built the artifact and found no device to push it to still exits 0 — "build the APK now, attach the phone after" is unchanged.

  • mix mob.deploy now exits non-zero when a device fails. A run that printed Failed on 1 device(s) previously still returned status 0, so CI and wrapper scripts read a failed deploy as success. Every device is still attempted and the full summary still printed first — only the exit code changed. A device that is skipped because the app is not installed on it stays non-fatal on both platforms. If you have CI that deploys to several devices and has been passing, it may now fail — check whether it was passing on a partial deploy.

Fixed

  • mix mob.deploy silently ignored unknown options. -d was never aliased to --device here — mob.connect has aliased it all along — so mix mob.deploy -d <udid> deployed to every connected device instead of the one named, without a word. A typo'd --devcie did the same. Parsing is strict now and the run refuses, naming the offending options.

  • iOS deploys installed the app under one bundle id and pushed BEAMs at another. :ios_bundle_id was resolved by the build but discarded by the deployer and connector, so mix mob.deploy --native --device <udid> installed successfully and then failed with "App '…' is not installed on this device." On a machine that also had an older build under the other id it was worse: the push silently succeeded against the wrong app and reported success.

  • The simulator and device builds disagreed about the bundle id. The simulator build took it verbatim from ios/Info.plist while the device build used the configured id, so simctl launch <configured-id> failed. Both paths now stamp and print the id they installed.

  • The device build looked its provisioning profile up by the Android id, so a profile minted by mix mob.provision (which uses the iOS id) was never matched.

  • The release IPA was stamped and signed with the Android applicationId, which App Store Connect rejects when it contains an underscore.

  • mix mob.uninstall targeted the Android id on iOS devices, removing nothing and reporting success.

  • mix mob.doctor warned that bundle_id was unset for a project that correctly set only ios_bundle_id.

  • Stamping CFBundleIdentifier crashed on an Info.plist that lacked the key — plausible for mix mob.adopt projects — instead of adding it.

[0.6.33] - 2026-09-04

Fixed

  • The shared OTP root keeps one version of an app, not every version it has ever seen (MOB-143). The root is keyed by OTP hash rather than by project, so every app on the machine shares one lib/. Installing a dependency without removing its other versions left them side by side, and the code server resolves an application to the highest version it finds there.

    That splits an app in half. Its beams come from its own mix.lock and are pushed to the device; code:lib_dir/1 — and so code:priv_dir/1, which is where load_nif looks — resolves against the shared directory. An app locking exqlite 0.38.0 would load 0.38.0's beams and 0.40.0's sqlite3_nif.so, and on_load would fail with bad_lib, taking down every database connection and the boot with it. The visible error talked about database credentials and said nothing about a cached artifact.

    It was also non-local: an app that built fine yesterday broke because a different app upgraded a dependency. pythonx always swept by wildcard and never had this; exqlite, emlx and the generic installer removed only a malformed empty-version directory and let real versions accumulate.

    All four install paths now sweep other versions, keep the target, and log what they removed. A stale version left by an earlier build is cleaned on the next native build, so the first build after upgrading may report removals.

[0.6.32] - 2026-08-31

Fixed

  • Android plugin component registration is backward-compatible again (regression in 0.6.31 — do not use 0.6.31; upgrade from 0.6.30 directly to 0.6.32 or later). ui_components.android.composable remains the native registry key used by existing plugin bridges instead of being treated as a callable Kotlin symbol. Generated registration is now explicit through android.factory; opted-in factories receive both props and the native event sender. The generated factories run before bridge registration so a bridge-owned factory remains authoritative for the same key.

  • Skew note: android.factory is inert on mob_dev ≤ 0.6.30 (silently ignored — the component renders only if bridge/host-registered).

[0.6.31] - 2026-08-31

Fixed

  • Android hosts now register plugin ui_components composables automatically (mob_scene3d-q03). The manifest's ui_components.android.composable was data nobody consumed: every host had to hand-register the Compose factory in MainActivity.onCreate, and a host that forgot rendered the component as nothing — MobNativeViewRegistry.render returns silently on an unknown key. The generated MobPluginBootstrap.registerAll(this) (already called by every generated/adopted MainActivity) now registers each activated plugin's composable with the app's MobNativeViewRegistry, mirroring iOS's mob_register_plugins() bootstrap (MobDev.Plugin.AndroidBootstrap). The registry key is android.view_module, falling back to ios.view_module; a bare composable is qualified with the bridge_class package. Silent blanks are gone: a typo'd composable fails the Gradle Kotlin compile, a malformed declaration (no key / no composable) fails the mob_dev build with the manifest error, and a declared-but-unresolvable composable (the hand-copied tier-2 workflow) registers a loud red "Missing native component" placeholder + Log.e that the host's own later registration overwrites. The validator also rejects an android.composable that isn't a Kotlin identifier or dotted path, at validate time instead of Gradle time.

[0.6.30] - 2026-08-30

Fixed

  • 0.6.29's published package could not compile. The compile-time Zig-pin reader loaded the repository-root .tool-versions, which Hex omits from packages. The required version now lives in packaged source (lockstep test against the repo's .tool-versions), and a packed-artifact regression compiles the real Hex layout and exercises doctor/native exact, mismatch, and failed probes. Use 0.6.30 instead of 0.6.29.

[0.6.29] - 2026-08-30

Changed

  • Zig preflight now enforces the exact Android build toolchain. mix mob.deploy --native and mix mob.doctor derive the required version from .tool-versions (zig 0.17.0-dev.269+ebff43698) and report missing/mismatched/broken zig with exact mise install commands, replacing the stale zig 0.15.x advice (asdf guidance dropped — asdf cannot fetch historical dev nightlies). Behavior change: a non-exact zig now fails the preflight even for legacy projects with app-owned build.zig that previously built with other versions — install the pinned version (mise install zig@0.17.0-dev.269+ebff43698). Companion to mob_new 0.4.27, which pins the same version in generated projects.

[0.6.28] - 2026-08-28

Added

  • mix mob.doctor warns when a project still carries the pre-MOB-104 sheet dismissal wiring. An app generated before the fix routes Mob.UI.sheet/2's :on_dismiss through MobBridge.nativeSendTap and delivers {:tap, tag}, but the documented contract — and iOS — deliver {:dismiss, tag}. A screen written to the contract never matches it and dies with FunctionClauseError, or the dismissal is silently swallowed and the sheet can never be re-presented. MobBridge.kt is app-owned and never re-rendered, so mob_new's template fix doesn't reach existing projects (see decisions/2026-08-25-detect-dont-autopatch-native-source.md for why this repo detects rather than auto-patches). The check names the two-line port and notes it needs mob >= 0.7.31. Only fires on projects that actually render sheets, so apps predating Mob.UI.sheet/2 stay quiet (MOB-104).

Changed

  • Physical iOS overrides are now replaced, not merged (devicectl --remove-existing-content), and the transfer is verified before the app is restarted. Note the trade-off: because mob_beam.m prefers Documents/otp/<app> on directory existence alone, a transfer interrupted after the copy begins leaves an incomplete override that the next launch will still prefer over the signed bundle. There is no rollback — devicectl has no delete verb, and an empty directory is still preferred — so the deploy now says so explicitly and names the recoveries (re-run the deploy, or reinstall).

Fixed

  • mix mob.deploy no longer drops dependencies reachable only through extra_applications. The runtime filter stopped consulting the project's own .app file, so a dep declared runtime: false and opted back in via extra_applications: — the documented idiom — was silently never pushed. The app booted and died with undef on first use. The project's .app is traversed again, with dev-only deps subtracted afterwards rather than by skipping the traversal.
  • BEAM discovery follows Mix.Project.build_path/0 and compile_path/0 instead of a hardcoded _build/dev, so a non-dev MIX_ENV or a custom :build_path is honoured. Projects using build_per_environment: false previously pushed no dependency BEAMs at all.
  • mix mob.deploy from an umbrella root now reports that mob does not support umbrellas instead of leaking a raw Mix.Project.app_path/1 error.
  • A physical-iOS deploy targeting a WiFi-only device returns a per-device error instead of throwing past the run and aborting every other device's summary.
  • iOS staging directories are PID-qualified, so two concurrent deploys cannot delete each other's staging mid-copy.

[0.6.27] - 2026-08-26

Added

  • mix mob.doctor detects the pre-fix Android component-event JNI mismatch (see mob's 0.7.25 CHANGELOG entry) in an already-generated project's MobBridge.kt and points at the fix — this repo doesn't auto-patch hand-editable native source (see decisions/2026-08-25-detect-dont-autopatch-native-source.md), so an existing app stays broken until a human ports the template fix or regenerates; this surfaces it via mix mob.doctor instead of an UnsatisfiedLinkError on first real interaction with a tier-2 native component. (MOB-98)

Fixed

  • mix mob.connect waited the full ~500ms iOS accessibility settle delay even against physical devices, where it does nothing — IOS.enable_accessibility/1 shells out to xcrun simctl, simulator-only tooling. Now gated on simulator type as well as platform, matching the same predicate already used elsewhere in this file for the identical reason. (MOB-99)

[0.6.26] - 2026-08-25

Added

  • MobDev.Plugin.Manifest.validate/1 now warns on unknown top-level keys in priv/mob_plugin.exs. mix mob.plugins previously reported "vetting clean" for a manifest carrying keys that meant nothing — the validator only checked the shape of known keys, never that every key was one of them. A styles: / default_style: key (which belongs in the separate priv/mob_style.exs manifest — see mob's MOB_PLUGINS.md) validated silently and then went nowhere. The warning is spec-versioned: a manifest declaring a plugin_spec_version this mob_dev doesn't support already errors separately (check_spec_version), so an unrecognized key there doesn't pile a second, noisier warning on top — forward-compat stays intact. For a supported spec version, an unrecognized key is almost always author error, and now says so instead of vanishing.

Found via a real report from someone building a style plugin against this system.

[0.6.25] - 2026-08-25

Added

  • Plugin-declared default fonts, part of the mob custom-fonts feature (see mob's MOB_FONTS.md). A capability plugin's priv/mob_plugin.exs manifest can now carry a default_font: %{family:, file:} field — family is the iOS PostScript name (can't be derived from the filename, so the plugin author supplies it), file must be one of the plugin's own assets.fonts entries. MobDev.Plugin.Manifest's validation pipeline checks the shape and the file reference; MobDev.Plugin.Merge.default_font/1 gathers it; MobDev.Plugin.Validator's conflict surface makes two plugins declaring one a build error (only one can win, since Mob.Theme.fonts[:default] is a single slot); MobDev.Plugin.RuntimeManifest.build/1 converts the survivor to the on-device %{ios:, android:} spec read at boot by Mob.Plugins.default_font/0.
  • Mob.Font.android_resource_name/1 is now called from mob core instead of a local copy. The plugin asset planner (MobDev.Plugin.Assets) used to carry its own filename→Android-resource-name normalization, kept in sync with mob's runtime copy (Mob.Theme.font/2) only by convention. Both now call the one implementation in mob, so a build-time bundled font and its runtime theme token can no longer drift apart.

Fixed

  • {:mob, path: "../mob"} only resolved on a machine with a sibling mob checkout — broke mob_dev's own CI, which has no such directory (** (Mix) Cannot compile dependency :mob because it isn't available). Switched to a git dependency during font development, and now to the proper Hex constraint (~> 0.7.25) now that mob has published the version carrying Mob.Font.

[0.6.24] - 2026-08-19

Fixed

  • mix mob.deploy's ERTS preflight silently treated a non-debuggable APK as healthy, so every subsequent run-as-based BEAM push failed and got swallowed too. ensure_erts_on_device/2 only pattern-matched the preflight run-as ... ls ... output for "No such file" / "not found" (missing-ERTS signatures). A run-as: package not debuggable: <pkg> response — the normal signal for an APK built as a release/non-debug variant — matched neither, so the check returned :ok and deploy_android proceeded straight into push_beams_android/push_priv_android/push_exqlite, each of which shells out to run-as #{pkg} tar xf ... 2>/dev/null; true. That trailing ; true exists to swallow Toybox tar's benign chown-to-macOS-UID exit code, but it also swallowed the run-as failure itself — so mix mob.deploy reported ✓ while nothing reached the device. ensure_erts_on_device/2 now detects any run-as:-prefixed response and returns a clear, actionable error (rebuild via mix mob.deploy --native to get a debuggable install) before any push is attempted.

[0.6.23] - 2026-07-12

Fixed

  • cpp_archive / nx_eigen NDK resolution honors ANDROID_HOME. MobDev.Plugin.CppArchive and MobDev.NxEigenNif hardcoded ~/Library/Android/sdk/ndk/<ver> (+ the darwin-x86_64 host), ignoring ANDROID_HOME / ANDROID_SDK_ROOT — so a lang: :cpp_archive plugin build failed with NDK toolchain not found wherever the NDK lives elsewhere (CI, a shared SDK), unlike the main NativeBuild path. Extracted shared MobDev.NdkVersion.{root,host,toolchain_bin,sysroot}/0 (on the existing env-aware sdk_root/0) and routed cpp_archive, nx_eigen_nif, and native_build's ndk_sysroot through them — one source of truth, so the NDK path can't diverge again. (MOB-89)

[0.6.22] - 2026-07-11

Added

  • ios_release_screenshot config — opt mob's public-API screenshot NIF into release builds. Companion to mob 0.7.20 (mob#71), which carves screenshot/3 into #if !MOB_RELEASE || defined(MOB_ENABLE_SCREENSHOT). Setting config :mob_dev, ios_release_screenshot: true exports MOB_ENABLE_SCREENSHOT=1, and the generated release_device.sh compiles mob_nif.m with ${MOB_ENABLE_SCREENSHOT:+-DMOB_ENABLE_SCREENSHOT} — so a release build can ship the screenshot NIF (letting an agent SEE a shipped app's screen to error-correct) while the private synthetic-input NIFs (tap/type) stay stripped. Default is unchanged and byte-identical (screenshot stays stripped); enabling it is a deliberate choice because the NIF captures the app's own window with no OS prompt or indicator. (#39)

[0.6.21] - 2026-07-08

Fixed

  • iOS app icons are flattened opaque so App Store upload validation accepts them. A transparent source icon (a common design — a rounded badge on transparent corners) produced transparent iOS icons and tripped altool error 90717 ("Invalid large app icon … can't be transparent or contain an alpha channel"). IconGenerator.write_ios_icons now composites an alpha-bearing source onto an opaque background (Image.flatten!) before resizing — using the explicit --adaptive-bg colour when given, else the same colour sampled for the Android adaptive background, so both platforms stay consistent; an opaque source is left untouched. Android icons keep their transparency (adaptive foregrounds + legacy launcher icons need it), and the bundled fallback mob_logo iOS assets were pre-flattened. Verified: a transparent-badge icon now passes App Store validation and reaches TestFlight. (#37)

[0.6.20] - 2026-07-07

Fixed

  • iOS release builds now compile + link activated-plugin NIFs. mix mob.release --ios produced a binary that failed to link for any app with NIF plugins — Undefined symbols: _<module>_nif_init, referenced from _erts_static_nif_tab in driver_tab_ios.o. iOS statically links every NIF into the single app binary (no dlopen under the App Store sandbox), so the generated driver_tab_ios.c references each activated plugin's <module>_nif_init, but the release path (release.ex → release_device.sh) hand-compiled a fixed object list and never touched plugins (the dev path already compiled them via build.zig -Dplugin_c_nifs). release_env/2 now emits MOB_PLUGIN_IOS_NIF_SOURCES + MOB_PLUGIN_IOS_FRAMEWORKS from MobDev.Plugin.activated() (via the new pure, tested MobDev.Release.plugin_ios_build_env/1); release_device.sh compiles each plugin NIF source with -DSTATIC_ERLANG_NIF_LIBNAME=<basename> + -fmodules (framework autolink) and links the objects plus the declared frameworks. Verified end-to-end: a 10-plugin app links and the IPA validates against its App Store distribution profile. Scope: covers nif_sources (lang: :c | :objc) + ios_frameworks; plugin swift_files and :cpp_archive static archives on the release path remain a follow-up. (#36)
  • Test: MobDev.Release.HelpersTest's git fixture no longer inherits the ambient git environment. Run inside a git hook (.githooks/pre-push), git exports GIT_DIR/GIT_WORK_TREE/GIT_INDEX_FILE; the fixture's git commands inherited them and operated on the outer repo instead of their tmpdir, crashing setup — so the suite passed standalone but failed only on push. The fixture now clears those vars on every git invocation. (#36)

[0.6.19] - 2026-07-07

Added

  • Plugins can contribute AndroidManifest <application> components and res/ files. Two new optional android: manifest keys — manifest_application_snippets (XML fragments spliced into the app's <application> block, idempotent per android:name) and res_files (plugin-relative paths copied into the app res/ tree at their derived res/<type>/<file> destination, path-contained + host-clobber-guarded + signed). This closes the gap that forced plugins needing a <service>/<receiver>/<provider> + resource (e.g. mob_nfc's HCE HostApduService + apduservice.xml) to make it a manual host_requirement. New MobDev.Plugin.Merge.android_manifest_snippets/1 + android_res_files/1 gatherers (classified in the cross-plugin conflict surface), manifest-schema validation, and NativeBuild splice + copy (ledger-pruned like bridge_kt). (MOB-39)

Changed

  • Plugin contributions merged into host-owned build files are now reversible (going forward). Plugin <uses-permission>s, <application> components, and Gradle deps are fenced in a regenerated-each-build managed region (MobDev.Plugin.ManagedBlock) in AndroidManifest.xml / build.gradle, so removing a plugin drops its fenced contributions (no more dangling <service> / orphan permission / orphan dep) while hand-authored content outside the fence is untouched. Idempotent (strip(place(x)) == x); de-dupes against existing entries; strip is orphan-BEGIN-safe (never deletes host lines between a stray marker and a real region). Forward-only: entries an older mob_dev already appended unfenced are indistinguishable from hand-authored ones, so they're left in place (not double-added, not auto-removed) — regenerate the app, or remove them by hand, to fence them. (MOB-40)

Fixed

  • mix mob.new_plugin no longer scaffolds plugins pinned to the abandoned mob ~> 0.6. MobDev.Plugin.Scaffold hard-coded {:mob, "~> 0.6"} in the generated mix.exs and mob_version: "~> 0.6" in every tier's manifest, so a freshly scaffolded plugin could not activate against the published mob 0.7.x (installed :mob 0.7.x does not satisfy mob_version "~> 0.6"). The requirement is now derived at scaffold time from the mob actually resolved in the project (Scaffold.detect_mob_requirement/0), falling back to a single @fallback_mob_requirement constant ("~> 0.7") when mob isn't loadable. mix.exs and the manifest always agree. A Scaffold test pins the default and asserts a generated manifest validates against a matching mob version, so the pin can't silently lag a future mob release. (#21)

[0.6.18] - 2026-07-05

Added

  • mix mob.provision can authenticate via an App Store Connect API key, so an unattended user — a CI runner or a headless agent account with no GUI login — can provision without a signed-in Xcode Apple ID account. Set APP_STORE_CONNECT_KEY_ID, APP_STORE_CONNECT_ISSUER_ID, and APP_STORE_CONNECT_API_KEY_PATH (the downloaded AuthKey_<id>.p8) and the task passes them to xcodebuild as -authenticationKeyID / -authenticationKeyIssuerID / -authenticationKeyPath. All three or none (a partial set raises, naming what's missing); with none set the signed-in Xcode account is used exactly as before. Pure, tested Mix.Tasks.Mob.Provision.asc_auth_args/1. Signing still needs the certificate + private key in an unlocked keychain — the API key only authorizes the profile/device calls. (#31)

[0.6.17] - 2026-06-29

Fixed

  • mix mob.connect crashed on a Mac set up only for iOS. The dist-port collision check shelled out to adb forward --list, and System.cmd("adb", …) raises :enoent for a missing binary rather than returning a non-zero exit — inside a linked Task, so the crash propagated and killed the whole connect. MobDev.Tunnel.run_adb/1 now resolves adb via System.find_executable/1 first (mirroring Discovery.Android.list_devices/0) and returns {:error, …} when it's absent, so the port scan degrades to "no forwards" and iOS-only Macs work with no Android platform-tools installed. (#29)

Added

  • mix mob.connect --ios-only / --android-only restrict discovery to one platform — handy when a phone for another project is plugged in. Set it project-wide in mob.exs with config :mob_dev, platforms: [:ios]. New pure helpers MobDev.Config.parse_platforms/1 and Mix.Tasks.Mob.Connect.resolve_platforms/2. (#29)

[0.6.16] - 2026-06-24

Added

  • Plugins can contribute array-valued iOS plist keys (e.g. UIBackgroundModes). apply_plugin_plist_keys! previously skipped any non-scalar ios.plist_keys value as unsupported; a list value now merges into the host Info.plist array (creating it if absent, appending only the missing string entries, deduped) instead of clobbering it. So one plugin can add bluetooth-central while another (e.g. mob_background) keeps audio. Merge decision extracted to the pure, tested NativeBuild.plist_array_additions/2.

[0.6.15] - 2026-06-23

Security

  • Bumped req 0.5.18 → 0.6.2 (pulls finch 0.22.0 → 0.23.0), clearing EEF-CVE-2026-49755 (HIGH) and EEF-CVE-2026-49756 (LOW) flagged by mix mob.security_scan. req is a transitive dep (via igniter); the bump stays within igniter's ~> 0.5 requirement.

[0.6.14] - 2026-06-23

Added

  • :extra_static_libs hook on :static_nifs entries — a Mob app can now link external per-ABI static archives alongside its project NIF archives. Some project NIFs intentionally declare extern symbols and don't host-link their backing archive (avoiding a host/device archive mismatch during mix compile); this lets the native app link resolve those symbols against the correct per-ABI .a. Entries require concrete per-ABI keys (:ios_sim, :ios_device, :android_arm64, :android_arm32, :android_x86_64), and the matching archive paths are appended to the existing -Dproject_rust_libs= link argument. -D<module>_static=true is emitted only on the ABIs where the entry applies, and iOS project-NIF filtering is now per-ABI so a device-only guarded entry doesn't leak into simulator args. Also passes Zigler 0.16's required generated-build flags when re-driving staged Zig NIF builds and resolves Zigler dotfile sources with match_dot: true. Verified on a physical SM-T577U (arm64-v8a) tablet via a Ghostty VT NIF. (#24)

[0.6.13] - 2026-06-20

Changed

  • mix mob.deploy --native now preserves on-device app data when the signing key matches. The Android install path previously ran an unconditional adb uninstall before adb install, which wiped MOB_DATA_DIR (on-device identity, screen stores) on every native deploy — even an in-place update signed with the same (e.g. committed) debug keystore. It now attempts adb install -r first and only falls back to uninstall + install when the in-place update is genuinely rejected (INSTALL_FAILED_UPDATE_INCOMPATIBLE from a signature mismatch, INSTALL_FAILED_VERSION_DOWNGRADE, etc.). Apps that pin a committed debug keystore now keep their identity across --native redeploys. Decision logic extracted to NativeBuild.needs_clean_reinstall?/2 and unit-tested.

Fixed

  • mix mob.deploy --native --android now fails fast with the real cause when zig is missing (landed in code before 0.6.13; previously undocumented). The Android JNI build is driven by build.zig; with zig off PATH it used to print a yellow "skipping build.zig step" warning and fall through to a CMake fallback that references C sources mob 0.7+ no longer ships, dying ~150 lines later with a misleading Cannot find source file: .../mob_nif.c. It now aborts before Gradle with an actionable message (install zig 0.15.x, verify with mix mob.doctor) when build.zig is present, zig is absent, and the legacy C sources are gone. Decision extracted to NativeBuild.zig_build_plan/3. (#20)

[0.6.12] - 2026-06-19

Fixed

  • 16 KB page-size alignment enforced at build time for every app. Android 15+ devices use 16 KB memory pages and Google Play requires every bundled .so to have 16 KB-aligned LOAD segments. The -Wl,-z,max-page-size=16384 link flag lives in the app's build.zig, which is copied once at mix mob.new and never regenerated — so apps generated before the template carried the flag kept linking 4 KB-aligned .so and failed Play. mix mob.deploy --native now reads the app's build.zig and, if its -shared link lacks the flag, injects it before linking (idempotent — a no-op when already present, e.g. the current mob_new template or a hand-fixed app). Pure core in inject_page_size_flag/1.

[0.6.11] - 2026-06-19

Added

  • mix mob.adopt — installs Mob into an existing Phoenix project, the Igniter-based install-into-existing counterpart to mix mob.new (which generates from scratch). Composable: the orchestrator runs the sub-installers mob.adopt.{deps,bridge,screen,mob_app,mob_exs,native,finalize}, each invokable independently. Default LV-bridge mode wires window.mob through a LiveView phx-hook and generates a mob_app.ex that boots the host Phoenix endpoint on-device (SQLite Repo assumed); --no-live-view generates a thin-client shell whose WebView opens a deployed server. Pre-1.0: refuses loudly (via Igniter issues) on umbrella / non-Phoenix / heavily-customised app.js or root layout / non-SQLite LV hosts rather than risk breaking the app. The native Android/iOS trees (--android / --ios) render from mob_new's templates and require the mob_new archive installed (mix archive.install hex mob_new) — mob_new stays the single source of native templates; the Elixir-side adoption needs no archive. Contributed as mob_new#8 by @ken-kost and relocated here — adopt is an Igniter task that mutates an existing project (like mob.add_nif / mob.enable), so it belongs in mob_dev (a Hex dep), not in mob_new (a self-contained Mix archive that can't carry Igniter). See decisions/2026-06-19-mob-adopt-lives-in-mob_dev.md. The shared patcher/generator helpers are duplicated from mob_new into MobDev.Adopt.{Patcher,Generator} pending the Phase-5 Igniter reunification.

[0.6.10] - 2026-06-19

Added

  • Plugin :cpp_archive NIFs — ship a C++ static library (e.g. an Nx backend) from a plugin. Manifest-driven cross-compile + --whole-archive static link (rule #11), with <module>_nif_init symbol verification and duplicate-init cross-validation. (#18)

Fixed

  • cpp_archive builds now fail fast with a named error on an unsupported Android ABI (e.g. the x86_64 emulator) instead of silently skipping — which previously deferred an unresolved <module>_nif_init to an on-device link failure.
  • The plugin manifest now requires a lowercase :module atom for :cpp_archive entries (fail-loud instead of fail-open / libnil.a).

[0.6.9] - 2026-06-18

Fixed

  • mix mob.publish --android now commits when Google requires changesNotSentForReview=true. Uploading a release while the app is under policy review (or otherwise can't auto-send for review) made the Play Edits :commit fail with HTTP 400 ("Please set the query parameter changesNotSentForReview to true"), discarding the whole edit so nothing reached the track. commit_edit/3 now detects that 400 and retries the commit with ?changesNotSentForReview=true; the changes land on the track and are sent for review from the Play Console UI. Verified uploading Io v13 to the internal track.

[0.6.8] - 2026-06-18

Added

  • mix mob.connect --only <serial> (alias --device/-d, repeatable). Restricts the run to devices whose serial/udid contains the given substring. Without it, connect attaches to every running device, so one slow or locked device (typically a plugged-in physical iPhone whose app restart blocks) could stall the whole session before any node connected. Verified end-to-end against a single Android phone: mix mob.connect --only ZY22CRLMWK tunnels, restarts, and connects livebook_mob_android_zy22crlmwk@127.0.0.1 on its serial-derived port 9633, then RPC into the device BEAM works (read live state, eval code).

[0.6.7] - 2026-06-18

Breaking

  • MobDev.Tunnel.setup/2 → setup/1. The dist port is now derived from the device serial internally, so callers no longer pass an index. setup/1 is @doc'd public API, so out-of-tree callers (a custom deploy Mix task, for example) break at compile time with MobDev.Tunnel.setup/2 is undefined or private. Did you mean: setup/1. Fix: drop the second argument — the returned %Device{} carries the assigned dist_port.

    (Noted retroactively. This shipped in 0.6.7 mentioned only as a parenthetical inside a "Fixed" bullet about dist-port keying, where anyone scanning for breaking changes would miss it — reported by @dl-alexandre in GenericJam/mob_dev#25.)

Fixed

  • mix mob.connect reliability — dist ports keyed by device serial, not run index. The Mac runs one shared EPMD; assigning ports as 9100 + index meant every project's first device claimed 9100, so two phones (or two projects' device-0) registered the same port and adb forward tcp:9100 could only reach one — the other silently timed out. Ports are now derived from the device serial (Tunnel.serial_base_port/1, a crc32 hash into 9100..9899) and bumped past any port another live node/forward already holds (assign_dist_port/2). A given phone always gets the same unique port across runs and projects, and deploy and connect agree on it.
  • Stale-tunnel cleanup. Tunnel.setup removes the device's own old forwards first (scoped to that serial), so prior runs no longer leave duplicate/wrong forwards that poison the next connect.
  • Real diagnostics on connect failure. A timed-out node now reports why — app not running / Standby-killed, dist never registered in EPMD, registered at a different port, no forward, or cookie mismatch — instead of a black-box "timed out".

[0.6.6] - 2026-06-18

Fixed

  • Android native build skips ABIs the app's build.zig doesn't handle instead of hard-failing. mob_dev builds arm64-v8a/armeabi-v7a/x86_64 by default, but an app's app-owned build.zig (copied at mix mob.new time) may predate x86_64 support (mob_new < 0.4.5) and reject -Dabi=x86_64. That used to fail the whole native build — aborting before the io.mob.plugin.MobPluginBootstrap regen, so the next gradle build then failed on an unresolved bootstrap. Now each ABI is pre-flighted against the build.zig (build_zig_supports_abi?/2) and unsupported ones are skipped with a warning (gradle abiFilters wouldn't ship them anyway). Real failures of a SUPPORTED ABI still halt the build.

[0.6.5] - 2026-06-17

Changed

  • Default OTP runtime → Elixir 1.20.1 (@otp_hash 7d46fdd4 → 5c9c69fc). Same OTP-29 / erts-17.0 / OpenSSL 3.4.0 base; the bundled Elixir stdlib (elixir/logger/eex) is swapped rc.5 → 1.20.1 across all five tarballs (android, android-arm32, android-x86_64, ios-sim, ios-device), published as the otp-5c9c69fc release on GenericJam/mob. bundled_versions.exs adds the 5c9c69fc bundle and flips active_hash. Backward compatible — apps pick up 1.20.1 on their next mix mob.deploy (recompile against the new runtime). NOTE: the major.minor skew check treats rc.5 and 1.20.1 as both "1.20", so it does NOT warn on this transition; the stdlib swap is what makes beams load.

[0.6.4] - 2026-06-16

Added

  • Android x86_64 emulator support (resolves GenericJam/mob#20). The x86_64 ABI is now wired throughout: OtpDownloader.ensure_android("x86_64") (#11), the x86_64 native build path (zig_build_android_objects, ensure_jni_libs, otp_dir_for_abi), and an android_x86_64 target in mix mob.release.otp plus the scripts/release/*x86_64* build scripts. The otp-android-x86_64-<hash>.tar.gz runtime is published on the otp-<hash> release. This is the slice x86_64 Linux / CI hosts need, where ARM emulation isn't available.

[0.6.3] - 2026-06-16

Fixed

  • exqlite NIF symlink now picks the device's actual ABI. The Android deployer hardcoded lib/arm64 for the sqlite3_nif.so symlink target, so on a 32-bit (armeabi-v7a) device the link dangled, exqlite was :nif_not_loaded, and any generated app using ecto_sqlite3 crashed on boot. It now probes lib/<abi>/libsqlite3_nif.so (Android extracts only the active ABI). Found + fixed verifying the showcase on a 32-bit Moto E — SQLite migrations now run and the app boots.

[0.6.2] - 2026-06-15

Fixed

  • Create an app-level driver_tab when a NIF-bearing plugin is active. A generated app ships no drivertab and links against mob's core static-NIF table — which has no plugin entries. So a plugin's <module>_nif_init linked but never registered, and the NIF was :nif_not_loaded on device (the home rendered, the Kotlin/permission bridge worked, but the actual capability call crashed). regen_driver_tab! now creates `priv/generated/driver_tab*.zig` (core + plugin entries) when plugins contribute NIFs and the app has none. Found + fixed verifying the showcase app on a physical Android phone and the iOS simulator (location demo now returns a real fix on both).

[0.6.1] - 2026-06-15

Fixed

  • Prune orphaned plugin artifacts when a plugin is removed. Plugin tier-3 merges copy files into the host tree (bridge Kotlin into the Kotlin sourceSet, migrations, images); these lingered after a plugin was dropped, and an orphaned bridge .kt could break the Gradle compile. NativeBuild's kotlin / migration / image merges now ledger what they write (per concern, under priv/generated/.mob_plugin_artifacts/) and delete what a prior build produced but the current one no longer does — so add/remove of a plugin is clean in both directions.

[0.6.0] - 2026-06-12

Added

  • Style packages, tokens-only tier: MobDev.Style (priv/mob_style.exs loader + validator), config :mob, :styles/:default_style activation, runtime-manifest emission (misconfiguration fails the build), and mix mob.styles.
  • ui_components expand: form honored (pure-Elixir composites): validated native-XOR-expand; expand-only plugins classify tier 2 but hot-push; the runtime manifest carries composites for boot registration.
  • mix mob.doctor: pre-plugin build.zig detection (missing -Dplugin_* options) and the host_requirements lane.
  • host_requirements manifest key printed by every native build; mix mob.new_plugin scaffolds starter test suites for every tier.

Changed

  • Native builds auto-regenerate the static-NIF driver_tab (was a checked-in artifact whose staleness produced runtime :nif_not_loaded).
  • mix mob.regen_plugin_manifest loads the host app first — spec-v2 generators may call host modules (mob_ash), not just read config.

Fixed

  • Dep detection uses Mix.Project.deps_paths, not stale _build dirs — ends the spurious MLX-404 downloads for apps that never dep emlx.
  • ExSlop is registered as a credo PLUGIN (it had silently never run).

[0.5.17]

Added

  • mix mob.new_plugin scaffolds a starter test suite for every tier (test/test_helper.exs + test/<name>_test.exs): tiers 1–4 get stdlib-only structural manifest checks (required keys, NIF stub loadable, native_dir exists, screen modules compile) with a pointer at mix mob.validate_plugin for the full validator; tier 0 gets a compile smoke test. New plugins start covered instead of starting at zero tests.
  • Plugin host_requirements manifest key. A plugin can declare human-readable host-app obligations the build can't automate (e.g. the AndroidManifest <service android:foregroundServiceType="mediaProjection"> fragment mob_screencast needs, or a capture FileProvider). The manifest validator enforces the shape, MobDev.Plugin.Merge.host_requirements/1 gathers them, and every native build prints them as a warning block — previously forgetting the manual step built + booted clean and only failed at first feature use.
  • mix mob.doctor detects pre-plugin build files. When plugins are activated but a build.zig/build_device.zig declares no b.option for the -Dplugin_* flags the native build emits, doctor warns with the exact missing options — previously Zig rejected the unknown flag half a build in.

Changed

  • mix mob.deploy --native regenerates the static-NIF driver table on every build (whatever formats the project uses, zig and/or c), exactly like the runtime plugin manifest: it's derived state. A stale checked-in driver_tab_* used to link a newly activated plugin's <module>_nif_init without registering it, so every NIF call raised :nif_not_loaded at runtime with nothing pointing at the cause.

Added

  • mix mob.new_plugin scaffolds tiers 3 (multi-screen) and 4 (sub-app), not just 0–2. Tier 3 emits two Mob.Screen modules + a :screens/:migrations manifest + a namespaced Ecto migration; tier 4 emits a lifecycle module + supervised worker + notification handler + settings editor screen + a :lifecycle/:settings/:notifications manifest. Generated manifests validate and modules compile against real mob.
  • Cross-plugin conflict detection. MobDev.Plugin.Validator.conflict_surface/0 classifies every merge gatherer; cross_validate/1 fails the build when two activated plugins clash on any shared resource — screen route, component atom, iOS/Android native view key, migration repo_namespace, NIF module, Swift/JNI source basename, Android bridge class, iOS plist key, supervised worker name, or notification match. A completeness meta-test forces every new shared-resource field to be classified; a property-based fuzzer checks detection is sound + complete across random N-plugin sets. A single plugin declaring a cross-platform NIF (one iOS + one Android entry sharing a :module) is correctly not flagged.

Changed

  • mix mob.deploy --native regenerates the runtime plugin manifest (priv/generated/mob_plugins.exs) on every build, not only when config :mob, :plugins changes. Adding/changing a plugin's tier-3/4 sections previously shipped a stale manifest (the new sections silently didn't activate on device); it is now derived state, always rebuilt before bundling — like the driver table.

Fixed

  • iOS simulator deploy now boots from a clean mix mob.deploy --native (was device-only). Three gaps made the sim deploy incomplete vs the device path, so the sim crashed on boot even though the device worked:
    • The Elixir-distribution apps elixir/logger were staged only under lib/<app>/ebin, which the sim's mob_beam.m doesn't add to the code path — boot failed at ensure_all_started(:elixir) with "elixir.app not found". They're now flattened into the flat BEAMS_DIR alongside eex (which already needed this), where the path resolves.
    • priv/ was only partially staged (repo/migrations), so Application.app_dir(:<app>, "priv/cacerts.pem") was :enoent and Mob.Certs.load_cacerts! crashed the boot. The whole priv/ is now rsynced into the flat dir (cacerts, mix/hex ebins, vendored static, …), matching the device release.
    • Paths.sim_runtime_dir/0 fell back to /tmp/otp-ios-sim for zig-based projects (no ios/build.sh), but the runtime is synced to ~/.mob/runtime/ios-sim — so the launcher and staging disagreed. It now recognizes ios/build.zig and returns the default runtime dir. Verified: a clean mob.deploy --native to an iPhone 11 Pro Max sim boots Io, Phoenix endpoint up, embedded Livebook home renders — no manual runtime fixups.

[0.5.16]

Fixed

  • Gate the plugin flags on the iOS device build path too. 0.5.15 gated the plugin-flag emission for Android and the iOS simulator build, but zig_build_binary_ios_device still emitted -Dplugin_swift_files/-Dplugin_frameworks (and generated the bootstrap, making plugin_swift_files always non-empty) unconditionally — so mix mob.deploy --native to a physical iPhone broke on an app scaffolded before the plugin system (invalid option: -Dplugin_swift_files). The device path now mirrors the sim path: bootstrap + flags only when plugins are activated. Verified mix mob.deploy --native to a physical iPhone — full OTP, Phoenix endpoint up, LiveView connected, embedded Livebook home rendered.

[0.5.15]

Fixed

  • mix mob.release --android --no-slim now actually ships the full OTP tree. The Android release stripped OTP libs unconditionally (OtpAssetBundle.build/2 was called with no opts), so --no-slim was silently ignored on Android. slim is now threaded build_aab → OtpAssetBundle.build(slim:); with slim: false the OTP tree ships untouched. Required for apps that run arbitrary user code at runtime (e.g. an embedded Livebook host doing Mix.install) — stripping any OTP lib (inets, ssl, xmerl, runtime_tools, …) is a latent crash when a user's deps need it. Default stays slim: true.
  • iOS --no-slim release passes App Store validation. The always-on Apple-policy strip cleared erts-*/bin and priv/bin but missed standalone executables inside OTP libs (e.g. erl_interface/bin/erl_call), which App Store validation rejects (90171). Now lib/*/bin/* executables are stripped too (always on), keeping every lib's .beam/.app — so a full-OTP --no-slim bundle is still Apple-compliant.
  • Native builds no longer break on pre-plugin app scaffolding. native_build.ex emitted -Dplugin_c_nifs/-Dplugin_zig_nifs/-Dplugin_jni_sources (Android) and -Dplugin_swift_files/-Dplugin_frameworks (iOS) unconditionally, but an app scaffolded before the plugin system has no such options in its build.zig and Zig rejects the unknown -D flag. These flags (and the iOS plugin bootstrap) are now emitted only when plugins are activated; a plugin-aware build.zig defaults them to "" so behaviour is unchanged there.

[0.5.14]

Fixed

  • iOS release: erl_errno_id_unknown shim written with a literal \n. The weak-stub line in release_device.sh used printf '%s\\n' inside the ~S (raw) heredoc, so bash received both backslashes and printf wrote a literal backslash-n into erl_errno_id_compat.c — clang then rejected the trailing }\n and mix mob.release --ios failed. Now printf '%s\n' (one backslash) emits a real newline. Regression guard added to release_script_test.
  • iOS release: clear preflight when priv/generated/driver_tab_ios.c is missing. The release links a per-app static-NIF driver table, but the dev build uses the built-in Zig table and mix mob.regen_driver_tab defaults to Zig, so a project that never ran it with --format c died deep in release_device.sh with a cryptic cc: no such file. build_ipa now fails early with the exact command to run (mix mob.regen_driver_tab --format c).

[0.5.13]

Fixed

  • iOS deploy now ships the whole priv/, not just priv/repo/migrations + priv/static. MobDev.Release's iOS bundler copied only migrations and priv/static, so apps that bundle extra runtime assets under priv/ — :mix/:hex ebins for on-device Mix.install, or a vendored library's own priv/ (e.g. Livebook's priv/static + priv/livebook) — silently never reached the device. Now rsyncs all of priv/, matching the Android deployer. Unblocks on-device Mix.install and embedded Livebook on iOS. Verified on a physical iPhone: priv/mix/ebin (103 beams) and priv/livebook/static present on device, embedded Livebook serves, and Mix.install([{:short_uuid, "~> 0.1"}]) returns :ok.

[0.5.12]

Changed

  • OTP runtime bumped to 7d46fdd4 (Elixir 1.20.0-rc.5). @otp_hash now points at the otp-7d46fdd4 release: all four platform tarballs (ios-sim, ios-device, android, android-arm32) bundle Elixir 1.20.0-rc.5 matched to the OTP-29 erts. Completes the 1.19.5 to 1.20 runtime migration the bundled-versions manifest was already staged for. Verified end-to-end on a physical iPhone and a physical Android (Moto G): System.version 1.20.0-rc.5, OTP 29, Mix.install([{:short_uuid, "~> 0.1"}]) returns :ok with the dep compiled on-device.

Fixed

  • iOS {spawn, <linked-in driver>} now works (erts erts_open_driver). The iOS-build #ifdef __IOS__ guard returned BADARG for any open_port whose spawntype included the EXECUTABLE bit (i.e. plain {spawn, Name}), firing before the linked-in-driver name lookup. This broke ram_file ({spawn, "ram_file_drv"}), and therefore `file:open(, [:ram]),erl_tarin-memory extract,hex_tarball.unpack, andMix.installon iOS. The guard now fires only when no linked-in driver matched the name. Bundled in the7d46fdd4` OTP tarballs.

[0.5.11]

Added

  • mob.exs :project_swift_sources config key — optional list of extra Swift sources to compile into the iOS app module alongside Mob's bridge sources. Threaded into both zig_build_binary_ios_sim and zig_build_binary_ios_device as -Dproject_swift_sources=<absolute,paths>. Comma-containing entries are rejected at the boundary; nil/[] is a no-op. Pairs with mob_new's project_swift_sources build hook (mob_new#5). Originally proposed by @dl-alexandre.

[0.5.10]

Added

  • mix mob.deploy --dist-port N and --node-suffix S flags — manual override path for the BEAM-distribution surface. When set, all targeted devices use the same value (use with --device <id> to be explicit). Nil falls back to per-device auto-allocation (Tunnel.dist_port(idx) + Discovery.Android.device_node_suffix / SIMULATOR_UDID-derived suffix). Resolves the register/listen error: no_reg_reply_from_epmd symptom seen when running multiple sims/emulators of the same app concurrently for cross-platform visual comparison.
  • MobDev.Device struct gains a :node_suffix field for plumbing the override per-device alongside :dist_port. Nil keeps auto-derive.
  • MobDev.Discovery.IOS.launch_app/3 accepts :node_suffix opt and forwards as SIMCTL_CHILD_MOB_NODE_SUFFIX to the launched sim. Companion to mob 0.6.10's MOB_NODE_SUFFIX support in mob_beam.m.
  • MobDev.Discovery.IOS.build_simctl_env/2 — pure helper extracted from launch_app/3 so override behaviour is unit-testable without spawning simctl. 7 new tests cover dist_port + node_suffix override paths.

Changed

  • MobDev.Connector.restart_app/1 pattern-matches :node_suffix from Device in both Android + iOS-sim variants, threading the value to the launchers.
  • MobDev.Deployer.deploy_all/1 accepts top-level :dist_port + :node_suffix opts; threaded through deploy_android and deploy_ios_simulator.

[0.5.9]

Changed

  • mix mob.enable tflite now injects {:nx_tflite_mob, "~> 0.0.3"} (Hex) instead of the GitHub-branch form. nx_tflite_mob v0.0.3 went live on Hex with 16 integration tests + a reproducible Mac host build path (see its CHANGELOG). Downstream Mob apps now get version-pinned deps + clean mix deps.tree output, instead of a transient github: checkout.

Notes

  • The Mac host-build path in nx_tflite_mob is for that package's own test suite, not for downstream consumers — production phone builds via mix mob.deploy --native continue to use the prebuilt Android AAR + iOS xcframework that mob_dev's MobDev.TfliteDownloader fetches.

[0.5.8]

Added

  • End-to-end mix mob.enable tflite — what 0.5.7 promised as "lands in 0.5.8". MobDev.NativeBuild now auto-detects the :nx_tflite_mob dep and threads the full TFLite path through Android + iOS sim + iOS device build pipelines:
    • maybe_build_tflite/1 → MobDev.TfliteDownloader.ensure/1 + MobDev.TfliteNif.build/2 for each target arch
    • tflite_zig_args_android/1 emits -Dtflite_static=true -Dtflite_lib=… for the per-ABI Android link
    • tflite_zig_args_ios/1 emits -Dtflite_static=true -Dtflite_dir=… -Dtflite_framework_dir=… for the iOS link
    • copy_tflite_runtime_lib_android/2 drops libtensorflowlite_jni.so into android/app/src/main/jniLibs/<abi>/ during the assemble step
  • copy_tflite_frameworks_ios/3 (kept as future-compat hook) — see the iOS-deploy-fix gotcha below
  • 13 new tests covering the public NativeBuild plumbing (native_build_tflite_test.exs), bringing the TFLite suite to 76 passing total

Fixed

  • iOS deploy: TFLite framework binaries are MH_OBJECT, not MH_DYLIB. TFLite's iOS xcframework slices ship their binaries as filetype=1 relocatable objects, which the linker statically pulls into the app's main Mach-O at build time. Trying to embed them as runtime .framework bundles tripped iOS install twice during this cut: first on missing per-framework Info.plist (which CocoaPods generates), then on "code signature version no longer supported" (iOS 26+ rejects v1 signatures, and codesign only makes v3 sigs for MH_EXECUTE/MH_DYLIB). The fix is to do nothing — the framework search-path arg already covers everything at build time.
  • Resolve :nx_tflite_mob via Mix.Project.deps_paths() rather than Path.join(deps_path, "nx_tflite_mob"). The latter assumes the dep landed in deps/ (hex / git deps do), but path: deps consume in-place from the user's source tree.

Verified on real hardware

  • Moto G Power 5G (BXM-8-256, Android 15): 75-117 ms YOLOv8n via NNAPI / mtk-gpu_shim
  • iPhone SE 3rd gen (A15, iOS 26.4): 24 ms YOLOv8n via Core ML → ANE (FP16 model; 214/385 nodes delegated)

[0.5.7]

Added

  • mix mob.enable tflite — wires TensorFlow Lite into a Mob project on iOS AND Android. Adds {:nx_tflite_mob, ...} to deps and generates <App>.TfliteInit (returns per-platform default delegate opts — NNAPI/mtk-gpu_shim on Android, Core ML delegate on iOS). The static-NIF table entry %{module: :tflite_nif, guard: "MOB_STATIC_TFLITE_NIF"} is registered in MobDev.StaticNifs.default_nifs/0, so the zig build picks it up automatically once tflite_static=true is set.
  • MobDev.TfliteDownloader — fetches tensorflow-lite-2.16.1.aar (Maven Central, Android) and TensorFlowLiteC-2.17.0.tar.gz (dl.google.com, iOS) into ~/.mob/cache/. Honours MOB_CACHE_DIR for test redirection and MOB_TFLITE_LOCAL_TARBALL_DIR for offline iteration.
  • MobDev.TfliteNif — cross-compiles tflite_nif.c (from the :nx_tflite_mob dep) per-arch and archives as libtflite_nif.a for static linking. Mirrors MobDev.NxEigenNif shape. Validates the produced symbol (tflite_nif_nif_init) before declaring success.

Bundle size impact: ~3-4 MB extracted (Android libtensorflowlite_jni.so), ~20-30 MB on iOS (TensorFlowLiteC + CoreML + Metal frameworks). Apps that don't enable TFLite pay zero size cost — the guard keeps the static-NIF table entry inactive.

End-to-end deploy (mob.deploy --native auto-build + runtime-lib embedding) lands in 0.5.8; this release ships the building blocks.

[0.5.6]

Added

  • CLAUDE.md "Release flow" section pointing at the canonical process in mob/RELEASE.md (URL form so it resolves without a local mob checkout). mob_dev specifics: the pre-push hook additionally runs mix mob.security_scan here (this is the only repo that ships the scanner), and the OTP tarball workflow stays separate from mix.exs version bumps.
  • .githooks/pre-push — same script shipped in mob (cheap preflight always, release preflight when mix.exs changed). The mob.security_scan step is gated via mix help availability so the same hook script works in all three repos.

[0.5.5]

Fixed

  • Android 15 segfault on launch (Pixel 7+, after the OS rolled out via OTA). Bumps @otp_hash from 550d7b78 → d9045670 to pick up OTP tarballs cross-compiled with -Wl,-z,max-page-size=16384. Without the flag, every .so in the bundled OTP runtime (crypto.so, asn1rt_nif.so, dyntrace.so, etc.) shipped with 4KB-aligned ELF PT_LOAD segments. Android 15 enforces 16KB alignment on devices with 16KB-page kernels and refuses to load misaligned libs, crashing the app at startup. New tarballs are 16KB-aligned (Align=0x4000).

[0.5.4]

Fixed

  • HexDocs source links pointed at the non-existent main branch — corrected to master so each </> glyph in generated docs opens the actual source file.

Added

  • .github/workflows/test.yml — runs mix test, mix format --check-formatted, mix credo --strict, and mix mob.security_scan on push to master and on every PR.
  • .github/workflows/release.yml — on tag push, creates a GitHub Release whose body is the matching ## [X.Y.Z] section from this changelog.

[0.5.3]

Changed

  • guides/nifs.md — rewrote the "Nx backends on mobile" section to match the current state (mix mob.enable nxeigen now real, mix mob.enable mlx includes the Metal GPU path on iOS device, EXLA "why not" preserved).
  • guides/nifs.md — restructured the multi-Rust-NIF section to lead with filmor's preferred shape (one Rustler crate per app, multiple #[rustler::nif] functions inside it). Multi-crate static linking remains supported and documented as an escape hatch, with the specific tradeoffs called out.

[0.5.2]

Added

  • mix mob.enable nxeigen — wires NxEigen (Eigen C++ CPU backend) into a Mob app. Builds as a C++ :static_nifs entry, cross-compiled per arch (arm64-ios, arm64-iossim, arm64-android, armv7a-android). FFT support uses Eigen's bundled kissfft.
  • EMLX Metal GPU enabled on iOS device. lib/mob_dev/mlx_downloader.ex now fetches the Metal-enabled libmlx.a + mlx.metallib bundle; lib/mob_dev/native_build.ex#maybe_bundle_mlx_metallib/2 copies the precompiled kernel library into the .app at build time, so EMLX.Backend with device: :gpu works on device without runtime kernel compilation.
  • scripts/release/mlx/ios_device_metal.sh + supporting build scripts for producing the Metal-enabled tarball; scripts/release/mlx/patches/0001-ios-metal-build.patch patches MLX 0.25.1's CMakeLists to switch SDK from macosx to iphoneos based on CMAKE_SYSTEM_NAME.

[0.5.1] and earlier

Earlier releases predate this changelog; consult the tag list and the per-tag commit messages for history.