fix(audio): recover playback when the mpv daemon dies or restarts

src/hooks/useAudio.ts

@@ -20,6 +20,7 @@ import {
 import {
 	createAudioBackend,
 	detectPlayers,
+	PlayerRestartedError,
 	type AudioBackend,
 	type BackendName,
 	type DetectedPlayer,
@@ -502,8 +503,25 @@ async function pause(): Promise<void> {
 	}
 }
 
+/** mpv was killed/crashed: respawn it and restart playback from the saved
+ *  position via the full play path (fresh loadfile, cover art, media
+ *  registry). A bare unpause would target a dead — or freshly-idle —
+ *  daemon and silently do nothing. */
+async function recoverPlayback(): Promise<void> {
+	const ep = currentEpisode();
+	if (ep && ep.audioUrl) {
+		await play(ep);
+	} else {
+		setError("Player is not running");
+	}
+}
+
 async function resume(): Promise<void> {
 	if (!backend) return;
+	if (!backend.isAlive()) {
+		await recoverPlayback();
+		return;
+	}
 	try {
 		await backend.resume();
 		setIsPlaying(true);
@@ -515,6 +533,13 @@ async function resume(): Promise<void> {
 			media.setPlaybackState(true);
 		}
 	} catch (err) {
+		// Race: the daemon died between the liveness check above and the
+		// unpause — backend.resume() respawned it and threw
+		// PlayerRestartedError (the fresh daemon has no file loaded).
+		if (err instanceof PlayerRestartedError) {
+			await recoverPlayback();
+			return;
+		}
 		setError(err instanceof Error ? err.message : "Resume failed");
 	}
 }

src/utils/audio-player.ts

@@ -24,6 +24,10 @@
  *   fills its demuxer cache ahead of time; the first real play just flips
  *   `pause` to false — the ~2s network open is paid at boot, not on the
  *   user's first Play.
+ * - Crash/kill recovery: a dead daemon (process exit or broken IPC socket)
+ *   is detected on the next command; play() respawns a fresh daemon and
+ *   reloads. resume() cannot unpause a freshly-idle daemon — it throws
+ *   PlayerRestartedError so the caller reloads the episode via play().
  */
 
 import { platform } from "os";
@@ -274,12 +278,28 @@ class MpvConnection {
 		}
 		this.handleTeardown();
 	}
+
+	/** True while the Unix socket is open — a live, reachable daemon. */
+	isConnected(): boolean {
+		return this.sock !== null;
+	}
 }
 
 // ── mpv Backend ──────────────────────────────────────────────────────
 // One resident daemon for the app's lifetime, controlled over a single
 // persistent JSON IPC connection with property observation.
 
+/** Thrown by resume() when the daemon restarted (killed/crashed) and the
+ *  previously-loaded file is gone — the fresh daemon is idle, so the
+ *  caller must reload the episode via the full play path instead of
+ *  unpausing (which would silently do nothing). */
+export class PlayerRestartedError extends Error {
+	constructor() {
+		super("mpv restarted; episode must be reloaded");
+		this.name = "PlayerRestartedError";
+	}
+}
+
 /** Property observation ids (correlate property-change events). */
 const OBS_TIME_POS = 1;
 const OBS_PAUSE = 2;
@@ -318,14 +338,35 @@ export class MpvBackend implements AudioBackend {
 	// ── Daemon lifecycle ─────────────────────────────────────────────
 
 	private async ensureDaemon(): Promise<void> {
-		if (this.proc && !this._exited && this.conn) return;
+		// Healthy = process alive AND its IPC socket open. A socket teardown
+		// with a living process (rare) is just as unusable as a dead one —
+		// every command would fail "not-connected" forever.
+		if (this.proc && !this._exited && this.conn?.isConnected()) return;
 		if (this.startPromise) return this.startPromise;
-		this.startPromise = this.spawnDaemon().finally(() => {
+		this.startPromise = this.recoverDaemon().finally(() => {
 			this.startPromise = null;
 		});
 		return this.startPromise;
 	}
 
+	/** Bring up a usable daemon. If the old process still lives with a dead
+	 *  IPC connection, kill it so the fresh spawn owns the socket path and
+	 *  no orphan lingers — and await its exit so its exit handler can't run
+	 *  after spawnDaemon() and clobber the new daemon's `_exited` flag. */
+	private async recoverDaemon(): Promise<void> {
+		const stale = this.proc;
+		if (stale && !this._exited) {
+			try {
+				stale.kill();
+			} catch {
+				/* already gone */
+			}
+		}
+		if (stale) await stale.exited.catch(() => {});
+		this.conn = null;
+		await this.spawnDaemon();
+	}
+
 	private async spawnDaemon(): Promise<void> {
 		// Clean up stale socket
 		try {
@@ -360,9 +401,15 @@ export class MpvBackend implements AudioBackend {
 		this._exited = false;
 		this.proc.exited
 			.then(() => {
+				// Daemon died (crash or external kill): every per-file state
+				// is gone with it. _loadedUrl null forces the next play()
+				// down the full reload path; _position/_volume/_speed are
+				// kept so a recovery reload can carry them over.
 				this._exited = true;
 				this._intentPlaying = false;
 				this._loadedUrl = null;
+				this._loadedPaused = false;
+				this._ended = false;
 				this._paused = null;
 			})
 			.catch(() => {});
@@ -590,6 +637,13 @@ export class MpvBackend implements AudioBackend {
 	}
 
 	async resume(): Promise<void> {
+		// The daemon may have died while we were paused (crash/kill): bring
+		// a fresh one up. It starts idle — no file to unpause — so throw
+		// PlayerRestartedError and let the caller reload the episode.
+		await this.ensureDaemon();
+		if (!this._loadedUrl) {
+			throw new PlayerRestartedError();
+		}
 		if (this._ended && this._loadedUrl) {
 			// Play pressed on a finished episode: replay from the top.
 			this._ended = false;
@@ -609,7 +663,12 @@ export class MpvBackend implements AudioBackend {
 			this._loadedPaused = false;
 		}
 		this._ended = false;
-		await this.send(["set_property", "pause", false]);
+		const resp = await this.send(["set_property", "pause", false]);
+		// Never claim success when the unpause didn't land: a dead/restarted
+		// daemon would otherwise leave the UI "playing" with no audio.
+		if (resp.error && resp.error !== "success") {
+			throw new Error(`mpv resume failed: ${resp.error}`);
+		}
 		this._intentPlaying = true;
 	}
 

tests/audio-backend.test.ts

@@ -12,6 +12,10 @@
  *    it by unpausing — the boot-restore fast path with no second load.
  * 6. EOF: the episode ends → isPlaying() goes false on its own; pressing
  *    resume() afterwards replays from the top.
+ * 7. Daemon death: a killed/crashed mpv is detected (isAlive drops);
+ *    resume() refuses to unpause the fresh idle daemon (throws
+ *    PlayerRestartedError) and play() recovers by respawning a fresh
+ *    daemon and loading the file.
  *
  * All playback runs silent (volume 0). Requires a real mpv on PATH;
  * tests skip where it is missing.
@@ -19,7 +23,10 @@
 import { test, expect } from "bun:test";
 import { tmpdir } from "os";
 import { join } from "path";
-import { MpvBackend } from "../src/utils/audio-player";
+import {
+	MpvBackend,
+	PlayerRestartedError,
+} from "../src/utils/audio-player";
 
 const SAMPLE_RATE = 22050;
 const FREQ = 440;
@@ -165,6 +172,46 @@ test.skipIf(!hasMpv)(
 	{ timeout: 20000 },
 );
 
+test.skipIf(!hasMpv)(
+	"daemon killed mid-play: resume() rejects on the fresh idle daemon; play() recovers a new one",
+	async () => {
+		fixtureWavs();
+		const backend = new MpvBackend();
+		try {
+			await backend.play(wavA, { volume: 0, speed: 1, startPosition: 0 });
+			await waitFor("playing", () => backend.isPlaying());
+			await waitFor(
+				"position advances",
+				async () => (await backend.getPosition()) > 0.5,
+			);
+
+			// Simulate a crash: SIGKILL the daemon out from under us.
+			const proc = (backend as unknown as { proc: { pid: number } }).proc;
+			expect(proc).toBeTruthy();
+			process.kill(proc.pid, "SIGKILL");
+			await waitFor("death observed", () => !backend.isAlive());
+
+			// resume() must NOT silently no-op on the dead daemon: it
+			// respawns, finds the fresh daemon idle (no file loaded), and
+			// throws — the hook falls back to the full play path.
+			await expect(backend.resume()).rejects.toThrow(PlayerRestartedError);
+
+			// play() (the hook's recovery) reuses the respawned daemon and
+			// plays the file — audio must actually advance again.
+			await backend.play(wavA, { volume: 0, speed: 1, startPosition: 0 });
+			expect(backend.isAlive()).toBe(true);
+			await waitFor(
+				"recovered playback advances",
+				async () =>
+					(await backend.getPosition()) > 0.5 && backend.isPlaying(),
+			);
+		} finally {
+			await cleanup(backend);
+		}
+	},
+	{ timeout: 20000 },
+);
+
 test.skipIf(!hasMpv)(
 	"EOF marks playback ended; resume() then replays from the top",
 	async () => {

tests/external-pause-reconcile.test.ts

@@ -16,6 +16,11 @@
  * property the same way the OS does and asserts useAudio reconciles in
  * both directions. Skipped when mpv isn't installed.
  *
+ * Also covers daemon crash recovery: killing mpv out from under the app
+ * must drop the UI out of "playing" (finalizeTrackEnd), and the next Play
+ * press must respawn a fresh daemon and resume audio from the saved
+ * position — the play button may never silently no-op on a dead player.
+ *
  * Real-timer note: the reconcile path runs on useAudio's real 150ms poll
  * interval against a real mpv process, with no injectable clock — the
  * deliberate-exception case from the no-real-timers rule (same as
@@ -190,6 +195,27 @@ async function waitFor(
 	}
 }
 
+/** SIGKILL the backend's mpv daemon — a crash/kill out from under the app.
+ *  The mpv command line carries the IPC socket path, so pgrep finds it by
+ *  that (the socket name is unique to this test process). */
+async function killMpvDaemon(): Promise<void> {
+	const socket = mpvSocket();
+	if (!socket) throw new Error("backend mpv socket not found");
+	const pids = (await Bun.$`pgrep -f ${socket}`.quiet().text())
+		.split("\n")
+		.map((s) => s.trim())
+		.filter((s) => s.length > 0)
+		.map(Number);
+	expect(pids.length).toBeGreaterThan(0);
+	for (const pid of pids) {
+		try {
+			process.kill(pid, "SIGKILL");
+		} catch {
+			/* already gone */
+		}
+	}
+}
+
 const episode = {
 	id: "ep1",
 	podcastId: "pod1",
@@ -235,6 +261,42 @@ test.skipIf(!hasMpv)(
 	{ timeout: 30000 },
 );
 
+test.skipIf(!hasMpv)(
+	"mpv killed mid-play: UI drops out of playing; pressing play recovers a fresh daemon",
+	async () => {
+		const audio = useAudio();
+		await audio.play(episode);
+		// Instant assertion: play() sets isPlaying synchronously when it
+		// succeeded. (In a shared worker that leaked a store mock from
+		// another test file, play() fails and this catches it at 0ms
+		// instead of burning the waitFor timeout below.)
+		expect(audio.isPlaying()).toBe(true);
+		// Let the clock advance past the 5s progress-save floor so recovery
+		// has a saved position to resume from (positions <5s are not stored).
+		await waitFor(() => audio.position() > 6);
+		const crashPos = audio.position();
+
+		// Crash the player out from under the app.
+		await killMpvDaemon();
+		await waitFor(() => !audio.isPlaying());
+		expect(audio.isPlaying()).toBe(false);
+		// The episode stays current — recovery can restart it.
+		expect(audio.currentEpisode()?.id).toBe("ep1");
+
+		// Press play: must respawn mpv and resume from the saved position —
+		// not silently flip the UI to "playing" with no process behind it.
+		await audio.togglePlayback();
+		expect(audio.isPlaying()).toBe(true);
+		expect(audio.position()).toBeGreaterThanOrEqual(crashPos - 0.5);
+		// Audio actually advances again — proof a fresh daemon is playing.
+		await waitFor(() => audio.position() > crashPos + 0.5);
+
+		await audio.stop();
+		expect(audio.isPlaying()).toBe(false);
+	},
+	{ timeout: 30000 },
+);
+
 // ── Teardown ──────────────────────────────────────────────────────────────
 
 afterAll(async () => {