Hot-swapping the delegate under a forwarding player

Some designs need a new player object per track — one media item each, two alive at once during a transition. Everything above (the session, the notification, the UI) wants one object that never changes. A ForwardingPlayer whose delegate can be replaced bridges the two:

internal class DelegatingForwardingPlayer(initial: Player) : ForwardingPlayer(initial) {
    fun swapDelegate(newDelegate: Player) { … }
}

The swap is: snapshot the listeners → detach the video output from the old delegate → remove the listeners → replace the delegate → re-add the listeners → re-attach the video output → verify.

Traps

Listeners are bound to the delegate at the moment they are added. The wrapper registers an internal adapter per listener against the current delegate; replacing the delegate leaves those adapters pointing at the old object. From outside this looks like "everything worked for the first track and then the notification and the UI froze". Track every listener yourself and cycle them:

private val tracked = mutableListOf<Player.Listener>()
override fun addListener(l: Player.Listener)    { tracked.add(l);    super.addListener(l) }
override fun removeListener(l: Player.Listener) { tracked.remove(l); super.removeListener(l) }
// in swapDelegate: tracked.toList() → super.removeListener each → set field → super.addListener each
// toList() snapshots defensively: a listener that re-enters add/removeListener mid-cycle would
// otherwise mutate tracked while this loop iterates it.

The delegate field is not writable through the public API. Reach it once, reflectively, in a companion, and cache the handle; a per-swap lookup is both slow and gives you no single place to fail. Fail loudly when the lookup returned nothing, rather than skipping the swap and leaving the old delegate in place — a silent no-op swap is far harder to diagnose than a thrown error:

private val PLAYER_FIELD: Field? = try {
    ForwardingPlayer::class.java.getDeclaredField("player").apply { isAccessible = true }
} catch (e: Exception) { Logger.e(TAG, "…", e); null }

Treat this as version-coupled: pin the library version, and re-run one swap after every upgrade.

Detach the video output from the old delegate before swapping, then re-attach after. A native video surface can be connected to one decoder at a time; leaving it attached makes the incoming player's attach fail and the app stop. And nothing re-sends the surface to a new delegate on your behalf, so without the re-attach video simply stops rendering while audio continues. Remember which kind of output was set — surface, surface view, texture view, surface holder — because you have to call the matching setter again.

Answer the playlist questions yourself. Each delegate holds one item, so its own hasNextMediaItem()/hasPreviousMediaItem() are always false, and the session concludes there is nothing to skip to: next/previous vanish from the notification and from external controllers. Inject a tiny provider backed by your real queue and override both the per-command query and command set:

override fun getAvailableCommands(): Player.Commands {
    val nav = playlistNavigationProvider ?: return super.getAvailableCommands()
    return super.getAvailableCommands().buildUpon().apply {
        add(Player.COMMAND_SEEK_TO_PREVIOUS)                       // always: restart current track
        if (nav.hasNextMediaItem())     { add(Player.COMMAND_SEEK_TO_NEXT); add(Player.COMMAND_SEEK_TO_NEXT_MEDIA_ITEM) }
        if (nav.hasPreviousMediaItem()) { add(Player.COMMAND_SEEK_TO_PREVIOUS_MEDIA_ITEM) }
    }.build()
}

Overriding only isCommandAvailable is the common half-fix: some surfaces read the Commands set instead and still show nothing.

Do NOT override the item count or the current index. It is tempting to report the real queue position, and it is the one override that will stop the app: the session layer validates currentMediaItemIndex < timeline.windowCount, and your delegate's timeline genuinely has one window. Index and count must stay consistent with the delegate (count 1, index 0); only the navigation answers come from your queue.

"Previous" and "previous item" are two different commands. seekToPrevious carries the restart-the-current-track-if-past-three-seconds rule, which users expect from a hardware button. seekToPreviousMediaItem must always move. Keep both in the provider, with the unconditional one defaulting to the other so implementers can ignore the distinction when it does not matter:

interface PlaylistNavigationProvider {
    fun hasNextMediaItem(): Boolean
    fun hasPreviousMediaItem(): Boolean
    fun seekToNext()
    fun seekToPrevious()
    fun seekToPreviousMediaItem() = seekToPrevious()
}

play()/pause()/setPlayWhenReady() need the same routing, or the intent falls through to the raw delegate. ForwardingPlayer's own implementations act on whichever delegate is currently attached, not on whatever tracks your play-intent across a swap. A system surface (notification, Bluetooth, headset, Android Auto) calling through the wrapper resumes that delegate's audio while your own "should the next item auto-play" bookkeeping stays exactly where it was — so the next track-change swap reads stale intent and loads the incoming track paused. The current track keeps playing; the transition silently does not. Override all three and route them through the same provider that already answers the navigation questions above:

override fun play() {
    val nav = playlistNavigationProvider
    if (nav != null) nav.play() else super.play()
}

Same shape for pause() and setPlayWhenReady(); super.X() is correct only when there is no provider at all — a swap without this trio plays the current track fine and goes silent on the next.

Nobody replays the events the new delegate already fired. The incoming player was prepared, given its item and possibly started before the swap, so the listeners you just re-attached missed its item-transition and metadata events, and the notification keeps showing the previous track. Dispatch them by hand right after the swap — item transition, metadata, then available-commands so the buttons re-evaluate — each in its own try/catch, so one bad listener can't abort the rest.

Guard the degenerate swap and verify the real one. Return early when the new delegate is the current one (the cycle would remove and re-add every listener for nothing), and after setting the field, read it back and log loudly if it did not take. A swap that silently failed presents as "the app plays the old track's audio with the new track's artwork", which nobody diagnoses quickly.

Verifying it

Run from the core submodule root. List the provider interface's members, then the wrapper's own overrides, and confirm every provider name reappears as an override fun — a gap there is a transport method routed straight to super, silently:

FILE=media/media3/src/main/java/com/maxrave/media3/exoplayer/DelegatingForwardingPlayer.kt
sed -n '/interface PlaylistNavigationProvider/,/^    }/p' "$FILE" | grep -oE 'fun [a-zA-Z]+'
grep -oE 'override fun [a-zA-Z]+' "$FILE" | sort -u

Behaviourally: start playback, pause and resume from the notification and a headset button, let a track end naturally, and confirm the next item auto-plays.