A mode's own drain-rate bucket starts empty until it accumulates history,
so toggling ANC into an unlearned mode fell straight through to Apple's
optimistic spec rating while the mode just left showed its worse measured
rate. Result: enabling ANC could make the displayed time jump up ~1h.
Fill an empty ANC bucket at display time with the less-optimistic of the
mode-agnostic UNKNOWN reading and a sibling mode's learned rate, scaled by
the ratio of the two modes' rated drain. Scoped to spec'd models and
device-supported modes; picks the best-evidenced sibling, tie-broken by
closest rated drain then recency. No persistence or UI change.
The existing spec ceiling and display clamp still backstop the borrowed rate.
- Show the time until the case is full inside its charging chip, learned from
the case's own rising level with the existing charge-band model; no rating
exists for case charging, so the first charge learns before it shows
- Derive a case battery health from observed transfer efficiency: pod percent
gained per case percent spent while docked and unplugged, corrected by each
pod's own health, compared against Apple's "with charging case" totals
- Case data is only genuine while a pod is docked — both transports silently
freeze the last value otherwise. Battery updates now flag whether the case
entry is live, and BLE gains strict same-frame case accessors, so estimates
never learn from frozen echoes
- The case deliberately gets no runtime estimate: idle-then-burst drain has no
meaningful hourly rate
- Mark the case metrics as experimental in the Battery settings card
- Show a "Still determining — check back after a few more listening sessions"
placeholder under Battery Health in the device info sheet while the estimate
is still accumulating data, so the feature is discoverable from day one; no
placeholder for profiles without a paired device, where it could never resolve
- Reorder the Battery card: time remaining & health settings first, then a
divider, then the charging-side settings (charge limit, charged notification)
- Mention the battery-health component in the estimate toggle and reset texts
- Ground the charge-band boundaries in Apple's documented fast-to-80%/trickle
charging behavior and standard lithium CC/CV charging references
- Replace the single linear charge rate with a three-band model (bulk / taper /
trickle) matching lithium CC/CV charging: each band learns its own rate, the
ETA walks the remaining bands, and the spec seed gets a taper haircut for the
slow bands — no more over-promising above 80%
- Base the battery-health figure exclusively on drain observed while the pod is
worn, audio is playing, AND this device is the system's audio sink; idle wear
previously diluted health upward against Apple's listening ratings
- Listening segments are flushed for persistence the moment their gate breaks
(playback stop, docking, transport flip) instead of being discarded with the
cleared window
- The time-remaining estimate keeps learning from all usage — actual current
drain, idle included, is the right basis for "how long will they last"
- The gauge line under the percentage now always shows the runtime estimate
("if used now") even while charging; the time-until-charged moved into the
charging chip itself ("Charging · 25m"), so the two can't be confused
- Battery health is now computed and shown per pod (Left/Right paired row in
the info sheet, mirroring the serial rows) — single-pod listening habits or
a replaced earbud make the sides genuinely diverge, and a combined figure
would mask a failing pod
- While charging, the gauge showed the bare runtime projection when no charge
rate existed yet — "1% · 4m" next to a charging chip reads as a four-minute
charge. The line now shows the "until charged" ETA or nothing
- Seed the charge rate from Apple's published quick-charge claims ("5 minutes
in the case provides around 1 hour of listening"), normalized against the
rated listening hours, so the ETA is present from the very first charge; a
live fit still takes over within minutes
- While a pod charges, fit its rising level and show the time until full in the
gauge instead of the runtime estimate; learned charge rates are persisted per
slot so the ETA appears immediately on later charges
- Suppress the charge ETA during Optimized Battery Charging holds, the final
trickle phase, and whenever the level stalls longer than one visible step
should take (granularity-aware: 1% AAP steps vs 10% BLE steps)
- Clear a slot's fit window when its readings switch between AAP and BLE — the
granularity jump would otherwise read as a fake level step
- Derive a battery-health percentage (median of accumulated drain rates vs the
model's rated life) and show it in the device info sheet; the info button now
also appears for BLE-only devices once health data exists
- Tag learned rates with the model they came from so re-pointing a profile at
different hardware starts learning fresh instead of inheriting foreign rates
- Track how many sessions blended into each learned rate and require three
before a health figure is shown
- Replace the global estimate toggle with a per-device toggle stored on the profile
- Seed the estimate from each model's rated battery life and show it immediately, using
the rating as a hard upper bound on displayed life while the measured rate converges
- When the ANC mode is unknown, seed from the shorter of a model's ANC-on/off ratings
- Show a projection while charging ("if used now") without ever learning from a rising battery
- Consolidate charge limit, "notify when charged", the estimate toggle and reset into one
Battery card; the charge notification now works for any live device, not only classic
audio connections
- Smooth the displayed time asymmetrically (drop fast, rise slow) so a faster-than-rated
drain stops over-promising within a couple of updates
Without tint, white PNG product icons (AirPods photos) are invisible on
light notification backgrounds. Restore src_in tint with notification_icon_tint
(a reliable day/night color resource) in both the PodInfoItemIcon.Notification
style (covers small layouts) and the big layout product ImageViews that do
not use that style.
Remove Theme.Material3.DynamicColors.DayNight from the notification
container style; it does not resolve correctly in the notification host
process, causing light-mode colors to appear in dark mode. Replace all
?attr/colorAccent and ?android:attr/textColorPrimary refs in RemoteViews
layouts with explicit day/night color resources (notification_icon_tint)
and TextAppearance.Compat.Notification.Title defaults. Change tintMode
from multiply to src_in for status glyphs; remove tint from product PNG
icons to prevent silhouette rendering.
Fixes#613
The conversation volume slider committed a profile write on every
drag frame — laggy dragging, and racing async writes persisted stale
mid-drag values (observed: 78 stored while the UI showed 70). Use
local state while dragging and commit on release, matching the
adaptive-noise and tone-volume sliders.
With one AirPod in the case the pod deterministically drops the
terminal 0x4B frame: the wind-down flurry ends on a transitional
status (1,2,3,0xB,4 then nothing), so the reaction never disengaged
and the volume stayed low until the 5-minute backstop (#608).
Reproduced on Pro 3 (fw 6589, 6503) and Pro 2 USB-C (fw 6814) —
all three share one firmware train; with both pods worn the terminal
always arrives.
The pod sends no frames during active speech, so any non-START frame
means the wind-down has begun and a terminal is imminent. A HOLD now
arms a short 6s fuse instead of the 5-minute backstop: if no terminal
(or fresh START) follows, the reaction disengages anyway. The fuse
must stay above ~5s — gaps up to 2.8s were observed between
consecutive wind-down frames. STARTs still arm the long backstop,
since the pod stays engaged and silent against ambient noise for
20-30s+ and a short timeout there resumes media mid-conversation.
Validated on hardware on both models: 7/7 single-pod conversations
restored exactly 6s after the last wind-down frame.
AirPods Gen4 emit a one-off type 0x07 frame from their identity address
on connect whose payload prefix is 0x07 instead of 0x01. Its bytes are
not the plaintext status format, but the model bytes happen to match,
so it minted a fresh device tracker and showed as a duplicate device
card until the stale timeout.
Every known plaintext status broadcast (including pairing mode) uses
prefix 0x01; pairing state lives in the suffix byte. Reject anything
else at the decoder, logging the dropped frame's hex.
Fixes#603
@mvanhorn independently fixed#598 in #605 by no longer resetting the
case cooldown on close, throttling the re-pop. We cherry-picked that
commit above to keep his authorship/credit, but this PR instead removes
the underlying lid-state flapping at its source (out-of-case pod's stale
lid byte -> UNKNOWN), so the cooldown can keep resetting on close and a
genuine close->reopen still shows the popup. Reverting his change here so
the two approaches don't stack; thanks @mvanhorn for the parallel work.
When one pod is out of the case, the out-of-case pod broadcasts bit4-only
frames whose lid byte is stale and decodes to a phantom OPEN even while the
case is shut, interleaved ~1:1 with the correct in-case-pod (bit6) frames.
The derived lid state flapped OPEN<->CLOSED, so the case-open popup re-popped
~0.5s after closing and sometimes lingered. Verified on AirPods Pro 1 (issue
log) and Pro 3 (live BLE capture).
- Trust the lid bit only from in-case-pod (bit6) or both-in-case (bit2) frames;
bit4-only frames now decode to UNKNOWN, and the real state is recovered from a
recent in-case broadcast (matches LibrePods).
- getLatestCaseLidState recovers from history within a 2s age window instead of a
fixed frame count, so a missed CLOSED can't keep a stale OPEN.
- Don't refresh the show-cooldown on a non-CLOSED hide, so a transient UNKNOWN
can't suppress a genuine re-open.
- Add a freshness backstop: dismiss the popup if a fresh OPEN broadcast stops
arriving (device left BLE range while open).
- Key popup/auto-connect de-duplication on the derived lid state, not just raw
advertisement bytes, since the effective lid can change while bytes don't.
Closes#598
Adds a dashboard hint card pointing at the existing Troubleshooter when a profile is connected to the system (audio) but CAPod is receiving no live data — the symptom of a phone dropping AirPods BLE broadcasts (e.g. some HyperOS devices, #603). Debounced ~15s so it doesn't flash during the gap between an audio connection and the first broadcast, and suppressed whenever any pod is live so it never claims 'no data' while data is visibly arriving.
Probe compatibility options through a transient in-memory override on BlePodMonitor instead of writing the user's persisted settings on every attempt. Only the winning combo is persisted, and only on success; failure or cancellation clears the override, restoring the user's original settings. Combos are tried fewest-disables-first so a phone that only needs batching disabled isn't left with filtering disabled too. Per-attempt cache clearing plus a freshness cutoff stop a previous combo's cached devices from satisfying the next one, and the 'found' checks now require a fresh live BLE observation rather than cached/AAP state.
The pod only signals CA start and end, not continuous keep-alives. On fw 6861 it held CA engaged for 21s with zero 0x4B frames while the wearer kept talking, so the 12s stale-timeout fired mid-speech and resumed media; the pod never re-sent a start frame, so it didn't re-pause.
Disengage now waits for the explicit not-speaking frame (status 5 added as a terminal STOP, since that firmware winds down 3->5 and never reaches 6/8/9). Transitional/unknown statuses stay engaged. The stale timer is demoted to a long 5min backstop for a fully-dropped terminal frame; constants moved to Kotlin Durations.
Reacts to the AirPods speaking-detection event (AAP 0x4B): lowers media volume by a configurable amount or pauses playback when you start talking, and reverts when you stop. Per-profile, Pro-gated, opt-in (default off).
Decodes the speaking status from the last payload byte ({1,2}=start, {6,8,9}=stop, else keep-alive); engage/disengage with a frame-idle stale timeout to recover a dropped stop, plus disconnect and service-stop cleanup.
Consolidate the nudge availability DataStoreValue with the rest of the persisted settings instead of carrying it in a separate Hilt module + DataStore file. Matches the existing convention where compat-style flags (offloaded filtering, indirect callbacks, etc.) live alongside theme/notification settings.
NudgeCapabilityStore keeps the verdict-mapping behavior; only the source of the persisted value changes.
- Seed connectedDevices flow with onStart so ALWAYS autolaunch fires before HEADSET profile binds
- Gate UnavailableMissingPermission classification on Android 12+ (BLUETOOTH_CONNECT only exists from API 31)
- Treat blank/empty profile.address as unpaired (matches AutoConnect)
- Subscribe DeviceSettingsViewModel to nudgeCapabilityStore.availability so UI updates immediately on verdict change
- Seed availability StateFlow with the persisted value via valueBlocking to close the cold-start race
Matches Apple's iOS/macOS behavior: stem-press pauses and sleep-detection pauses are explicit user intent, not eligible for auto-resume on next pod-in. sendPause gains a rememberForResume parameter (default false); only PlayPause's auto-pause branch passes true. New sendStop wrapper clears the auto-resume flag for stem-mapped MEDIA_STOP.
Replaces the origin-tracking machinery with a simpler model that matches Apple's iOS/macOS behavior: auto-play strictly resumes a CAP-dispatched pause, gated by a sticky boolean cleared on inactive→active transitions. The original fire-on-cold-wear behavior is preserved as a per-device opt-in 'Start music on wear' setting.
Two SIGSEGV native crashes recurred on Android 10 in 5.1.4-rc0 inside JIT-cached code at toBatteryFloat+4 (popup) and mergeBatterySlot+40 (cache merge). Both functions had a boxed Float? unbox at function entry that R8 horizontally merged into stdlib host classes, where Android 10 ART JIT miscompiled the unbox.
Convert PodDevice battery getters to non-null Float with BATTERY_UNKNOWN sentinel and propagate primitive Float through every display/persistence consumer. mergeBatterySlot now takes primitive Float; toBatteryFloat and toBatteryOrNull are deleted. Add isKnownBattery and batteryProgress helpers used everywhere instead of scattered nullable checks. Raw live extraction in toCachedState avoids touching the unified getter so cached values aren't refreshed as live.
Move consumeUpgradeExtra() out of MainActivity.onNewIntent and into a Compose-level LaunchedEffect that collects a new MainActivity.warmIntents SingleEventFlow.
onNewIntent could fire before the setContent{} lambda registered the back stack with NavigationController, leaving navCtrl.goTo(Nav.Main.Upgrade) to throw IllegalStateException("NavigationController not initialized"). Same race shipped a fix for in octi (d4rken-org/octi#284).
Use AAP-connected mocks for the dark dashboard to showcase full connectivity badges (BLE + IRK + encrypted + AAP) and the ANC mode selector. Populate the add-profile preview with a demo name, model, and paired device. Add user-facing labels to the dashboard mock devices.
Refactor BluetoothDevice2 so name/address are primary constructor fields, letting previews construct one without a real Android BluetoothDevice. Bump the screenshot test JVM heap to 4g — the smoke batch (42 renders) was hitting the test executor's default ceiling.
Mirrors permission-pilot's policy: only the 6 smoke locales (en-US, de-DE, ja-JP, ar, zh-CN, pt-BR) check phoneScreenshots PNGs into the repo. Non-smoke locales are excluded via .gitignore. Drops fastlane/metadata/android/ from ~67 MB to ~7 MB and prevents future bloat from full regens. Play Store's supply retains previously-uploaded screenshots for locales not pushed, so full localization is maintained by occasional manual regen + screenshots_only upload.
BluetoothSocket.connect() is a blocking JNI call that ignores coroutine cancellation, so the previous withTimeout in AapAutoConnect only cancelled the suspending wrapper while the native thread stayed pinned. Hung threads accumulated and could trigger ANRs.
Move the timeout inside AapConnection and run the blocking connect on a daemon thread; on timeout, close the socket from the caller thread to unblock the native call (the documented Android pattern for cancelling in-flight L2CAP connects).
Also cancel appScope before delegating uncaught exceptions so coroutines have a best-effort window to release resources before the system handler terminates the process.
Battery slot percent comparisons in mergeBatterySlot/hasStateChanged compiled to Intrinsics.areEqual on boxed Float; R8 optimization on Android 10/11 dropped a null check during inlining and the resulting NPE escaped onEach { persistLiveDevices }, cancelling the upstream combine and freezing every observer of DeviceMonitor.devices.
Comparisons now operate on primitive float (cmpg-float in dex) so no Intrinsics.areEqual call remains in the merge path. The persist loop also catches and reports per-profile, and AAP-only profiles with active DeviceInfo are now persisted even when no BLE pod is in range.
Wrap logging, reporting, and Looper resume calls so the foreground service timing exception suppression cannot itself trigger another crash. Extract handler into a dedicated class with seams for unit tests.
Bit 5 of pubStatus is always set on A3454 and no longer carries the wear flag (unlike Max gen 1). Read bits 1 and 3 instead — the per-earcup sensors. OR rather than AND so phones that only see one bit reliably still report worn correctly.
Closes#548
Sync flags with what the BLE classes actually report and what iOS exposes:
- AirPods Gen 1/2/3: enable hasEarDetection (already parsed via DualApplePods) and hasEarDetectionToggle
- AirPods Gen 3, Pro 1: enable hasEndCallMuteMic (force-sensor stems)
- Powerbeats Pro, Beats Fit Pro: enable hasEarDetectionToggle (iOS exposes it)
- Beats Solo Pro, Studio 3: drop hasEarDetection (over-ear, BLE class is bare SingleApplePods)
- FAKE_AIRPODS_GEN1/2/3: enable hasEarDetection to match HasEarDetectionDual
- Generalize microphone mode description from 'AirPod' to 'earbud'
Tests rewritten as exhaustive set assertions plus implication invariants.
Glance only calls provideGlance() once per widget session; subsequent update() calls recompose the existing composition without re-running provideGlance. The previous capture-once approach left widgets frozen at initial state because the composition had no reactive State to read.
Subscribe to a widgetDeviceFlow(profileId) inside provideContent that pre-filters by WidgetDeviceKey, so the composition recomposes on visible state changes without firing on every BLE advertisement. Replace updateAll() with explicit per-GlanceId update() calls in WidgetManager, with a platform-id fallback when Glance returns no IDs.
Replaces the user-facing scanner mode setting with an automatic policy that picks LOW_LATENCY when a profile-paired device is connected, BALANCED in the foreground, and LOW_POWER in the background. The TroubleShooter scopes a temporary LOW_LATENCY override via a refcounted withTemporaryOverride block so overlapping callers stay correct.
Fixes a regression where the controller could block BLE scanning entirely if BLUETOOTH_CONNECT was missing or the HEADSET profile proxy stalled, and adds a reactive bondedDeviceAddresses flow so bond changes propagate without waiting for an unrelated input. Cleans up the now-dead scanner mode strings across all locales and unused ScannerMode fields.
Matthias Urhahn