Skip to content

Recover the Silent Tune keep-alive after audio interruptions - #745

Open
bjorkert wants to merge 4 commits into
devfrom
fix/silent-audio-keepalive-recovery
Open

Recover the Silent Tune keep-alive after audio interruptions#745
bjorkert wants to merge 4 commits into
devfrom
fix/silent-audio-keepalive-recovery

Conversation

@bjorkert

@bjorkert bjorkert commented Aug 21, 2026

Copy link
Copy Markdown
Member

Silent Tune's looping silent audio is the app's only background claim. When it stops, iOS suspends the app within seconds and it cannot recover on its own, which is what produces the "App inactive for N minutes" notification.

Reactivation now runs inside a UIApplication background task assertion so the retries get runtime even without the audio claim, keeps trying for about 18 seconds, and no longer trusts player.isPlaying, which keeps reporting true after the session is taken. The BGAppRefreshTask stays open until the restart resolves, and scheduleRefresh no longer pushes a pending request further out.

Three other things stop the audio and none were handled. A route disappearing, such as leaving CarPlay, pauses the player with no interruption notification. A media services reset invalidates the session and player. And a process launched into the background by BGAppRefreshTask never runs the backgrounding transition, so it had no observers attached and no background alerts armed. All three are covered now, and a background recovery clears any delivered notification it has made obsolete.

This does not make Silent Tune reliable, since an interruption that outlasts the assertion still suspends the app. Logging is quiet while healthy and detailed once struggling, including one line per lost runtime window measured against a monotonic clock.

The silent audio session is the only background execution claim Silent Tune
has, so when another app interrupts it the process is suspended within
seconds and no app code runs again until iOS resumes it. Reactivation now
happens inside a UIApplication background task assertion, which grants
runtime independently of the audio claim, so the retry attempts actually
run instead of being frozen by suspension. Attempts are spread over about
18 seconds, which fits inside both the assertion and a BGAppRefreshTask
window, and playAudio confirms the player is really playing rather than
assuming play() worked.

The BGAppRefreshTask stays open until the restart resolves, so a failed
first attempt is retried within the same window. It also queues its
successor before doing any work, asks for the earliest possible window when
a restart fails, and cancels itself when the user is no longer on Silent
Tune. scheduleRefresh leaves a pending request alone when it would run at
least as soon, so repeated background transitions no longer push the check
further out.

Logging follows the same shape: quiet when the keep-alive is healthy, and
detailed once it is struggling. A first attempt success stays behind debug
logging, while recovery after a failure reports the attempts and elapsed
time, repeated failures with an unchanged error code are suppressed so a
long ladder cannot bury the log, and session errors are named instead of
being printed as raw four character codes. The task scheduler records one
line per lost runtime window with its length and which background alerts
fired, and the background and foreground transitions record the refresh
mode, Low Power Mode and Background App Refresh status.
Device logs show isPlaying still reporting true at the moment an interruption
is delivered, so the recovery entered on interruption began was skipped
before it took its background task assertion, leaving the app dependent on
iOS delivering interruption ended after all. Recovery no longer consults
isPlaying to decide whether to run. Reattempting against a player that is
genuinely playing is harmless, because the session is activated before the
player is replaced, so a failed attempt leaves a working player untouched.

Recovery on interruption began now waits a second before its first attempt.
A brief interrupter's interruption ended lands inside that window and
supersedes the work, so momentary blips stay as quiet as before, and work
that does run means the claim is really gone whatever isPlaying says. The
immediate background refresh request moved to that point as well, so it is
made once per real interruption instead of once per blip, and the assertion
expiring now arms it unconditionally since reaching expiration means the
claim was never re-established.
Silent audio also stops when the audio route disappears and when media
services reset, and neither posts an interruption notification, so nothing
noticed. A day of device logs shows three silent deaths with zero
interruptions while the phone moved in and out of CarPlay. Route changes and
media services resets are now observed and recover through the same ladder.
Recovery runs for a route appearing or disappearing; a category change never
recovers, because playAudio sets the category itself and an alarm takes over
the session that way. Other reasons are logged so a later log can earn them a
recovery.

A process launched into the background by BGAppRefreshTask never runs the
backgrounding transition, so it had no interruption, route or media services
observers and no background alerts armed. Observers are now attached whenever
audio is restarted, and a background recovery arms the alerts, which also
clears any delivered notification the recovery has just made obsolete. The
task scheduler is kicked on recovery so the alerts are re-armed from the
moment runtime returns. Alerts are only armed while backgrounded, since the
task's work lands on the main queue and the app may have been opened in
between.

The runtime gap is measured against a monotonic clock, so a wall clock
correction cannot hide a stall, and a material difference between the two is
reported. A scheduler park that outlives the moment between a task firing and
its action rescheduling it is now reported with its duration.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant