From 9479279eb99a42085e685a11c3db8f87e0e1b37b Mon Sep 17 00:00:00 2001
From: Ethan <257484394+EthanYoQ@users.noreply.github.com>
Date: Fri, 25 Sep 2026 17:36:17 +0800
Subject: [PATCH] fix(remote-input): ignore stale audio preparation after stop
 or cancel

Invalidate in-flight audio setup on every capture teardown and guard
resume, microphone, worklet and timeout continuations by generation.
Preserve normal microphone reuse and add deterministic regression tests.
---
 openless-all/app/assets/remote-input/app.js   |  19 +-
 .../remote-input-audio-generation.test.mjs    | 246 ++++++++++++++++++
 2 files changed, 258 insertions(+), 7 deletions(-)
 create mode 100644 openless-all/app/scripts/remote-input-audio-generation.test.mjs

diff --git a/openless-all/app/assets/remote-input/app.js b/openless-all/app/assets/remote-input/app.js
index d3ed74b01..a07ac65a8 100644
--- a/openless-all/app/assets/remote-input/app.js
+++ b/openless-all/app/assets/remote-input/app.js
@@ -1537,8 +1537,10 @@
     setStatus(L.preparingMic, 'work');
     clearResult(); // 清掉上一次的识别结果,避免新录音时还显示旧文字
 
+    var gen = audioGen;
     withTimeout(ensureAudio(), MIC_PREP_TIMEOUT_MS, 'TIMEOUT')
       .then(function () {
+        if (gen !== audioGen) return;
         if (!recording) {
           // 期间已被取消/松手
           teardownAudioCapture();
@@ -1549,6 +1551,7 @@
         setStatus(L.preparingBackend, 'work');
       })
       .catch(function (err) {
+        if (gen !== audioGen) return;
         recording = false;
         resetRemoteStreamState();
         // 超时多半是 audioCtx 卡死(resume 永不 settle),彻底重建,否则下次重试会继续卡在
@@ -1639,6 +1642,7 @@
   // 确保 AudioContext / getUserMedia / 采集节点就绪并开始推流。
   // 必须在用户手势调用栈内(startRecording 由手势触发)。
   function ensureAudio() {
+    var gen = audioGen;
     // 不支持 getUserMedia
     if (!navigator.mediaDevices || !navigator.mediaDevices.getUserMedia) {
       return Promise.reject(new Error('UNSUPPORTED:浏览器不支持录音,请升级或换浏览器'));
@@ -1665,11 +1669,10 @@
 
     return resumeP
       .then(function () {
+        if (gen !== audioGen) return null;
         // 2) 麦克风流(已存在则复用)
         if (mediaStream) return mediaStream;
-        // 捕获当前代际:迟到 resolve 时若代际已变(超时重置/断线释放),停掉轨道并放弃,
-        // 避免泄漏麦克风或覆盖重试成功的新流。
-        var gen = audioGen;
+        // 使用准备开始时的代际；停止/取消后迟到的流必须释放，不能覆盖下一次录音。
         return navigator.mediaDevices
           .getUserMedia({
             audio: {
@@ -1701,7 +1704,7 @@
       .then(function (stream) {
         // 3) 建立采集图(若已建好则跳过)。audioCtx 可能在准备超时后被 resetAudioContext
         // 置空(本次 getUserMedia 迟到 resolve),此时直接放弃,避免对 null ctx 建图报错。
-        if (sourceNode || !audioCtx || !stream) return;
+        if (gen !== audioGen || sourceNode || !audioCtx || !stream) return;
         sourceNode = audioCtx.createMediaStreamSource(stream);
         return buildCaptureGraph();
       });
@@ -1709,12 +1712,14 @@
 
   // 建立 AudioWorklet(优先)或 ScriptProcessor(兜底)
   function buildCaptureGraph() {
+    var gen = audioGen;
     var inSr = audioCtx.sampleRate || 48000;
 
     // 优先 AudioWorklet
     if (audioCtx.audioWorklet && typeof AudioWorkletNode !== 'undefined') {
       return loadWorklet()
         .then(function () {
+          if (gen !== audioGen) return;
           workletNode = new AudioWorkletNode(audioCtx, 'ol-pcm-worklet', {
             numberOfInputs: 1,
             numberOfOutputs: 0,
@@ -1723,12 +1728,13 @@
           });
           workletNode.port.onmessage = function (e) {
             // e.data 是已转换好的 Int16 LE ArrayBuffer
-            sendAudio(e.data);
+            if (gen === audioGen) sendAudio(e.data);
           };
           sourceNode.connect(workletNode);
           usingWorklet = true;
         })
         .catch(function () {
+          if (gen !== audioGen) return;
           // worklet 加载失败 → 回退 ScriptProcessor
           usingWorklet = false;
           buildScriptProcessor(inSr);
@@ -2035,6 +2041,7 @@
   // ============================================================
   // 仅停止"采集/推流"(断开节点),保留 audioCtx & mediaStream 以便快速重启。
   function teardownAudioCapture() {
+    audioGen++; // 停止/取消也作废在途的 resume、麦克风、worklet 和准备超时回调。
     releaseWakeLock();
     if (wakeLockHint) wakeLockHint.textContent = L.wakeLockHint;
     try {
@@ -2072,7 +2079,6 @@
 
   // 彻底释放(断线时):停止麦克风轨道并关闭 ctx。
   function teardownAudio() {
-    audioGen++; // 代际推进:作废所有在途的 getUserMedia 迟到回调
     teardownAudioCapture();
     if (mediaStream) {
       try {
@@ -2093,7 +2099,6 @@
   // 与 teardownAudio 的区别:这里 close 并置空 audioCtx —— 超时根因往往是 ctx 自身坏掉
   // (resume 永不 settle),保留它只会让下次继续卡。
   function resetAudioContext() {
-    audioGen++; // 代际推进:作废所有在途的 getUserMedia 迟到回调
     teardownAudioCapture();
     if (mediaStream) {
       try {
diff --git a/openless-all/app/scripts/remote-input-audio-generation.test.mjs b/openless-all/app/scripts/remote-input-audio-generation.test.mjs
new file mode 100644
index 000000000..bd991d2f7
--- /dev/null
+++ b/openless-all/app/scripts/remote-input-audio-generation.test.mjs
@@ -0,0 +1,246 @@
+import assert from 'node:assert/strict';
+import { readFileSync } from 'node:fs';
+import test from 'node:test';
+import vm from 'node:vm';
+
+// Exercise the production lifecycle functions, with only browser/transport effects
+// replaced. Deferred promises and explicit timers make late callbacks deterministic.
+const source = readFileSync(new URL('../assets/remote-input/app.js', import.meta.url), 'utf8');
+const names = [
+  'withTimeout',
+  'startRecording',
+  'stopRecording',
+  'cancelRecording',
+  'ensureAudio',
+  'buildCaptureGraph',
+  'clearPendingPcm',
+  'resetRemoteStreamState',
+  'teardownAudioCapture',
+  'teardownAudio',
+  'resetAudioContext',
+];
+const lifecycle = names
+  .map((name) => {
+    const start = source.indexOf(`  function ${name}(`);
+    const end = source.indexOf('\n  }', start);
+    assert.ok(start >= 0 && end > start, `production function ${name} must exist`);
+    return source.slice(start, end + '\n  }'.length);
+  })
+  .join('\n');
+const settle = () => new Promise((resolve) => setImmediate(resolve));
+const deferred = () => {
+  let resolve;
+  let reject;
+  const promise = new Promise((yes, no) => {
+    resolve = yes;
+    reject = no;
+  });
+  return { promise, resolve, reject };
+};
+const makeStream = () => {
+  const track = { stopped: false, stop() { this.stopped = true; } };
+  return { track, getTracks: () => [track] };
+};
+const makeNode = () => ({
+  disconnected: false,
+  port: {},
+  connect() {},
+  disconnect() { this.disconnected = true; },
+});
+
+function harness({ worklet = false, suspended = false } = {}) {
+  const calls = { mic: [], resume: [], worklet: [], nodes: [], sent: [], statuses: [], pcm: [] };
+  const timers = new Map();
+  let timerId = 0;
+  let scriptBuilds = 0;
+  class AudioContext {
+    constructor() {
+      this.state = suspended ? 'suspended' : 'running';
+      this.sampleRate = 48000;
+      this.audioWorklet = worklet ? {} : undefined;
+    }
+    resume() { const d = deferred(); calls.resume.push(d); return d.promise; }
+    suspend() { this.state = 'suspended'; return Promise.resolve(); }
+    close() { this.state = 'closed'; return Promise.resolve(); }
+    createMediaStreamSource() { return makeNode(); }
+  }
+  const state = {
+    recording: false, startSent: false, awaitingResult: false,
+    ws: { readyState: 1 }, audioGen: 0, audioCtx: null, mediaStream: null,
+    sourceNode: null, workletNode: null, scriptNode: null, usingWorklet: false,
+    remoteSessionId: '', remoteSequence: 0, finishAfterStarted: '',
+    pendingPcm: [], pendingPcmBytes: 0, resampleState: { phase: 0, last: 0, hasLast: false },
+    wakeLockHint: null, TARGET_SR: 16000, MIC_PREP_TIMEOUT_MS: 10000,
+    L: { preparingMic: 'mic', preparingBackend: 'backend', micTimeout: 'timeout', ready: 'ready' },
+    window: { AudioContext },
+    navigator: { mediaDevices: { getUserMedia() {
+      const d = deferred(); calls.mic.push(d); return d.promise;
+    } } },
+    AudioWorkletNode: class { constructor() { const n = makeNode(); calls.nodes.push(n); return n; } },
+    setTimeout(fn) { const id = ++timerId; timers.set(id, fn); return id; },
+    clearTimeout(id) { timers.delete(id); },
+    loadWorklet() { const d = deferred(); calls.worklet.push(d); return d.promise; },
+    buildScriptProcessor() { scriptBuilds++; state.scriptNode = makeNode(); },
+    wsSendJSON(message) { calls.sent.push(message.type); },
+    sendAudio(pcm) { calls.pcm.push(pcm); },
+    setStatus(message) { calls.statuses.push(message); },
+    micErrorText(error) { return error.name; },
+    interruptRecording() { throw new Error('unexpected interruption'); },
+  };
+  for (const name of [
+    'clearRecoveryTimer', 'acquireWakeLock', 'releaseWakeLock', 'clearReadyTimer',
+    'clearWorkTimeout', 'updateRecordBtnUI', 'clearResult', 'detachHoldEnd',
+    'setLevel', 'enterTranscribing', 'armWorkTimeout', 'saveRecoverySession',
+  ]) state[name] = () => {};
+  vm.runInNewContext(lifecycle, state);
+  return { state, calls, timers, get scriptBuilds() { return scriptBuilds; } };
+}
+
+async function startPending(h) {
+  h.state.startRecording();
+  await settle();
+}
+async function resolveMic(h, index) {
+  const stream = makeStream();
+  h.calls.mic[index].resolve(stream);
+  await settle();
+  return stream;
+}
+function assertActive(h, stream, node) {
+  assert.equal(h.state.recording, true, 'old callback must not stop the new recording');
+  assert.equal(h.state.startSent, true);
+  assert.equal(h.state.mediaStream, stream, 'old stream must not replace the new microphone');
+  assert.equal(h.state.sourceNode, node);
+  assert.equal(node.disconnected, false);
+  assert.equal(stream.track.stopped, false);
+  assert.deepEqual(h.calls.sent, ['start'], 'only the current attempt may send start');
+}
+
+for (const end of ['stopRecording', 'cancelRecording']) {
+  test(`late microphone success after ${end} cannot replace a retry`, async () => {
+    const h = harness();
+    await startPending(h);
+    h.state[end]();
+    await startPending(h);
+    const current = await resolveMic(h, 1);
+    const node = h.state.sourceNode;
+    const stale = await resolveMic(h, 0);
+    assert.equal(stale.track.stopped, true, 'cancelled microphone request must release its tracks');
+    assertActive(h, current, node);
+  });
+}
+
+test('late microphone rejection cannot stop a retry', async () => {
+  const h = harness();
+  await startPending(h);
+  h.state.stopRecording();
+  await startPending(h);
+  const current = await resolveMic(h, 1);
+  const node = h.state.sourceNode;
+  h.calls.mic[0].reject(Object.assign(new Error('old request'), { name: 'NotAllowedError' }));
+  await settle();
+  assertActive(h, current, node);
+  assert.equal(h.calls.statuses.at(-1), 'backend');
+});
+
+test('old preparation timeout cannot close the current AudioContext', async () => {
+  const h = harness();
+  await startPending(h);
+  const oldTimeout = [...h.timers.values()][0];
+  h.state.stopRecording();
+  await startPending(h);
+  const current = await resolveMic(h, 1);
+  const node = h.state.sourceNode;
+  const context = h.state.audioCtx;
+  oldTimeout();
+  await settle();
+  assertActive(h, current, node);
+  assert.equal(h.state.audioCtx, context);
+  assert.equal(context.state, 'running');
+});
+
+test('late resume cannot start another microphone request after cancellation', async () => {
+  const h = harness({ suspended: true });
+  await startPending(h);
+  h.state.stopRecording();
+  h.calls.resume[0].resolve();
+  await settle();
+  assert.equal(h.calls.mic.length, 0, 'cancelled resume must not request microphone permission');
+  assert.deepEqual(h.calls.sent, []);
+});
+
+test('late resume cannot borrow a new recording and send a second start', async () => {
+  const h = harness({ suspended: true });
+  await startPending(h);
+  h.state.stopRecording();
+  await startPending(h);
+  h.state.audioCtx.state = 'running';
+  h.calls.resume[1].resolve();
+  await settle();
+  const current = await resolveMic(h, 0);
+  const node = h.state.sourceNode;
+  h.calls.resume[0].resolve();
+  await settle();
+  assertActive(h, current, node);
+});
+
+for (const outcome of ['resolve', 'reject']) {
+  test(`late worklet ${outcome} cannot rebuild the current capture graph`, async () => {
+    const h = harness({ worklet: true });
+    await startPending(h);
+    const current = await resolveMic(h, 0);
+    h.state.stopRecording();
+    await startPending(h);
+    h.calls.worklet[1].resolve();
+    await settle();
+    const node = h.state.sourceNode;
+    const currentWorklet = h.state.workletNode;
+    h.calls.worklet[0][outcome](new Error('old module'));
+    await settle();
+    assertActive(h, current, node);
+    assert.equal(h.state.workletNode, currentWorklet);
+    assert.equal(h.calls.nodes.length, 1);
+    assert.equal(h.scriptBuilds, 0, 'stale worklet failure must not start fallback capture');
+    assert.equal(h.state.usingWorklet, true);
+  });
+}
+
+test('current worklet failure still falls back to ScriptProcessor', async () => {
+  const h = harness({ worklet: true });
+  await startPending(h);
+  await resolveMic(h, 0);
+  h.calls.worklet[0].reject(new Error('worklet unavailable'));
+  await settle();
+  assert.equal(h.scriptBuilds, 1);
+  assert.equal(h.state.recording, true);
+  assert.deepEqual(h.calls.sent, ['start']);
+});
+
+test('current timeout still resets the context and releases a late microphone', async () => {
+  const h = harness();
+  await startPending(h);
+  [...h.timers.values()][0]();
+  await settle();
+  assert.equal(h.state.recording, false);
+  assert.equal(h.state.audioCtx, null);
+  assert.equal(h.calls.statuses.at(-1), 'timeout');
+  const stale = await resolveMic(h, 0);
+  assert.equal(stale.track.stopped, true);
+  assert.deepEqual(h.calls.sent, []);
+});
+
+test('normal stop still pairs start/stop and reuses the microphone on the next attempt', async () => {
+  const h = harness();
+  await startPending(h);
+  const stream = await resolveMic(h, 0);
+  h.state.remoteSessionId = 'active-session';
+  h.state.stopRecording();
+  assert.deepEqual(h.calls.sent, ['start', 'stop']);
+  assert.equal(h.state.awaitingResult, true);
+  assert.equal(stream.track.stopped, false);
+  h.state.awaitingResult = false; // Backend completed the preceding recording.
+  await startPending(h);
+  assert.equal(h.calls.mic.length, 1);
+  assert.equal(h.state.mediaStream, stream);
+  assert.deepEqual(h.calls.sent, ['start', 'stop', 'start']);
+});