dsh-live-voice 0.7.0 → 0.7.2

package/lib/server.js

@@ -2052,7 +2052,7 @@ var VAD_ASSET_TYPES = {
 };
 var MAX_VAD_ASSET_BYTES = 32 * 1024 * 1024;
 function registerVadAssetRoutes(ctx, options = {}) {
-  const version = options.version ?? (true ? "0.7.0" : "test");
+  const version = options.version ?? (true ? "0.7.2" : "test");
   if (!/^[a-zA-Z0-9][a-zA-Z0-9.+-]{0,127}$/.test(version))
     throw new Error("Invalid VAD asset version");
   const directory = options.directory ?? fileURLToPath(new URL("./vad/", import.meta.url));
@@ -2146,7 +2146,7 @@ import { lstat as lstat2, readFile as readFile5 } from "node:fs/promises";
 import { fileURLToPath as fileURLToPath2 } from "node:url";
 import { join as join8 } from "node:path";
 function registerBrowserAsrAssets(ctx, options = {}) {
-  const version = options.version ?? (true ? "0.7.0" : "test");
+  const version = options.version ?? (true ? "0.7.2" : "test");
   if (!/^[a-zA-Z0-9][a-zA-Z0-9.+-]{0,127}$/.test(version)) throw new Error("invalid-version");
   const base = "/api/dsh-live-voice/asr/" + encodeURIComponent(version) + "/";
   const directory = options.directory ?? fileURLToPath2(new URL("./asr/", import.meta.url));

package/lib/vad/manifest.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
-  "pluginVersion": "0.7.0",
+  "pluginVersion": "0.7.2",
   "dependencies": {
     "@ricky0123/vad-web": "0.0.31",
     "onnxruntime-web": "1.22.0"
@@ -36,12 +36,12 @@
       "sha256": "2623a2953f6ff3d2c1e61740c6cdb7168133479b267dfef114a4a3cc5bdd788f"
     },
     "vad.capture.js": {
-      "bytes": 1254,
-      "sha256": "da75b304c7a67d2538b96c0310cb9d4d991a46ca495f246b19ab70ce71b9f26b"
+      "bytes": 1694,
+      "sha256": "d24df69dc6326d8c1b37437b0bef459a3a67405237108a3c32868a87c3fa13c4"
     },
     "vad.worker.js": {
       "bytes": 79337,
-      "sha256": "16ae4a3a3d4fa13ad7e65e8a7ce220e816448c4b9899ebaa53a2ad4c287eecbb"
+      "sha256": "6b83547e105d6a61b37f3647687021afcab2162d707071e7bffd9a6ffecade58"
     }
   }
 }

package/lib/vad/vad.capture.js

@@ -1,11 +1,22 @@
 // src/modules/recognition/vad/vad.capture.ts
 var VadCaptureProcessor = class extends AudioWorkletProcessor {
   stopped = false;
+  paused = false;
   pcm = new Float32Array(2048);
   at = 0;
   constructor() {
     super();
     this.port.onmessage = (event) => {
+      if (event.data?.type === "resume") this.paused = false;
+      if (event.data?.type === "flush") {
+        if (event.data.pause) this.paused = true;
+        if (this.at) {
+          const pcm = this.pcm.slice(0, this.at);
+          this.port.postMessage({ pcm, sampleRate }, [pcm.buffer]);
+          this.at = 0;
+        }
+        this.port.postMessage({ type: "flushed" });
+      }
       if (event.data?.type === "stop") {
         this.stopped = true;
         if (event.data.drain && this.at) {
@@ -20,6 +31,7 @@ var VadCaptureProcessor = class extends AudioWorkletProcessor {
   process(inputs, outputs) {
     for (const channels of outputs) for (const channel of channels) channel.fill(0);
     if (this.stopped) return false;
+    if (this.paused) return true;
     const input = inputs[0];
     if (!input?.length) return true;
     for (let i = 0; i < input[0].length; i++) {

package/lib/vad/vad.worker.js

@@ -1632,7 +1632,7 @@ var vo = Ke;
 Object.defineProperty(I.versions, "web", { value: $t, enumerable: true });
 
 // src/modules/recognition/vad/assets.ts
-var sileroVadAssetBase = () => "/api/dsh-live-voice/vad/" + encodeURIComponent(true ? "0.7.0" : "test") + "/";
+var sileroVadAssetBase = () => "/api/dsh-live-voice/vad/" + encodeURIComponent(true ? "0.7.2" : "test") + "/";
 
 // src/modules/recognition/vad/StreamingResampler.ts
 var StreamingResampler = class {

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-live-voice",
-  "version": "0.7.0",
-  "dshTestedVersion": "0.2.2-alpha.2",
+  "version": "0.7.2",
+  "dshTestedVersion": "0.2.2-rc.1",
   "type": "module",
   "scripts": {
     "build": "node --experimental-strip-types ../../scripts/build.ts --package live-voice",

package/src/app/ARCHITECTURE.md

@@ -3,6 +3,8 @@
 `src/app` is the composition boundary between DSH and the voice-domain modules. Client and server entry points remain separate; host-only Node dependencies must never enter the browser bundle.
 
 - `client/apply.tsx` injects DSH client services, creates session controllers and voice ownership, wires engines and settings, and disposes resources. `client/slotDefinitions.ts` and `client/registerSlots.tsx` own the established slot IDs, order, and registration. Preserve those contracts when changing placement. Microphone mode stop invalidates pending input ownership and application continuity without ending current playback; explicit takeover and full End retain their separate handoff/lifecycle semantics.
+- Retained session entries re-resolve the current DSH chat target on committed slot attachment. A replaced target releases its old subscription and republishes through stable readers/listeners without resetting session speech memory; detached-session admission still prevents background playback.
+- Composer detachment parks eligible recording-backed microphone controllers in the existing session registry rather than retiring them. Their capture clock pauses, recognition requests continue, and final results stay on the originating controller. Returning uses the existing ownership barrier before resuming capture and delivering queued results against the latest draft. Explicit stop clears parked intent; plugin disposal retires both mounted and parked entries. This memory is page-lifetime only, not reload recovery.
 - `server/apply.ts` composes host engines and providers. `server/registerRoutes.ts` registers authenticated same-origin routes; do not bypass their validation or expose host access directly to browser code.
 - `client/transcriptionLeaveGuard.ts` owns one activity-scoped native `beforeunload` listener across microphone, meeting sources and Settings test capture. It covers acquisition, listening and unsettled transcription, not playback or voice-mode preference alone. Cancelling the browser dialog does not stop capture; existing `pagehide` cleanup runs only after leaving. Idle and disposed apps remove the listener. Browser wording, user-activation requirements and mobile/forced-close limitations apply; this is not refresh recovery.
 - `client/i18n` owns the only UI translation tree: typed, alphabetized catalogs for every supported locale, runtime selection synchronized with `ctx.locale`, DSH registration, and the boundary wrapping slot components. UI language must not change STT/TTS language or user-defined command phrases.

package/src/app/client/apply.tsx

@@ -457,7 +457,12 @@ export function apply(ctx) {
       entry.speechState.checkpoint = entry.controller.checkpointSpeech();
     entry.speechAttached = false;
     ++entry.speechRequest;
-    const stopped = entry.controller.suspendSpeech();
+    const preserveInput =
+      entry.controller.snapshot.conversation &&
+      entry.controller.snapshot.listening &&
+      typeof entry.controller.recognition.pauseForNavigation === 'function';
+    entry.parkedInput = preserveInput;
+    const stopped = entry.controller.suspendSpeech(preserveInput);
     entry.speechRelease = Promise.resolve(stopped).then(() => {
       if (entry.controller.speechStopError) throw entry.controller.speechStopError;
       if (entry.controller.inputReleaseError) throw entry.controller.inputReleaseError;
@@ -478,6 +483,13 @@ export function apply(ctx) {
       await entry.speechRelease;
       if (disposed || entry.closed || !entry.speechAttached || request !== entry.speechRequest)
         return;
+      if (entry.parkedInput) {
+        const resumed = await entry.controller.resumeInputFromNavigation();
+        if (disposed || entry.closed || !entry.speechAttached || request !== entry.speechRequest)
+          return;
+        if (!resumed) throw new Error('The paused microphone recording could not resume.');
+        entry.parkedInput = false;
+      }
       entry.controller.restoreSpeech(entry.speechState.checkpoint);
       entry.speechState.checkpoint = null;
       entry.refreshSpeech(!entry.speechState.attached);
@@ -488,6 +500,7 @@ export function apply(ctx) {
   function retire(entry) {
     if (entry.closed) return;
     suspendSessionSpeech(entry);
+    if (entry.parkedInput && !disposed) return;
     entry.closed = true;
     if (meetingEntry === entry) {
       void meeting.attendee.stop();
@@ -677,7 +690,8 @@ export function apply(ctx) {
       entry.chatListeners.add(listener);
       return () => entry.chatListeners.delete(listener);
     };
-    entry.readChat = entry.chat.getSnapshot.bind(entry.chat);
+    // Stable React store readers must follow a replacement DSH session binding.
+    entry.readChat = () => entry.chat.getSnapshot();
     const refresh = (reset = false) => {
       if (disposed || entry.closed) return;
       const snapshot = entry.readChat();
@@ -716,10 +730,22 @@ export function apply(ctx) {
     };
     entry.refreshSpeech = refresh;
     refresh();
-    entry.unsubscribe = entry.chat.subscribe(() => {
+    const publishChat = () => {
       refresh();
       for (const listener of entry.chatListeners) listener();
-    });
+    };
+    entry.unsubscribe = entry.chat.subscribe(publishChat);
+    entry.reconnectChat = () => {
+      if (disposed || entry.closed) return;
+      const chat = ctx.uiConversation.binding(sessionId).target('chat');
+      if (chat === entry.chat) return;
+      entry.unsubscribe?.();
+      entry.chat = chat;
+      entry.unsubscribe = chat.subscribe(publishChat);
+      // Subscribe first: activation may materialize the target synchronously.
+      // Existing speech memory and detached-session admission rules still apply.
+      publishChat();
+    };
     // DSH 0.1.6 may omit this legacy store. Keep the remaining voice controls
     // available while structured-question integration is unavailable.
     const pendingInteractions = ctx.uiSession.pendingInteractions;
@@ -753,7 +779,10 @@ export function apply(ctx) {
       const before = controller.getSnapshot();
       update(next);
       const settings = controller.getSnapshot().settings;
-      if ([...recognitionSettingKeys].some((key) => before.settings[key] !== settings[key])) {
+      if (
+        !entry.parkedInput &&
+        [...recognitionSettingKeys].some((key) => before.settings[key] !== settings[key])
+      ) {
         if (!entry.restoreInput) entry.inputRestoreRequest = entry.request;
         entry.restoreInput ||= Boolean(before.listening || before.starting);
         const revision = (entry.inputSettingsRevision = (entry.inputSettingsRevision || 0) + 1);
@@ -802,6 +831,7 @@ export function apply(ctx) {
       controller[method] = (...args) => {
         // Internal handoffs release resources, not the parked navigation cursor.
         if (method === 'endConversation' && args[0] === true && !entry.speechAttached) {
+          if (entry.parkedInput) return entry.speechRelease ?? Promise.resolve();
           controller.patch({ conversation: false });
           return Promise.all([
             controller.stopListening(),
@@ -823,6 +853,7 @@ export function apply(ctx) {
           }
         }
         entry.request++;
+        if (method !== 'stopSpeech') entry.parkedInput = false;
         if (
           (method === 'endConversation' && args[0] !== true) ||
           method === 'stopConversationInput'
@@ -909,6 +940,7 @@ export function apply(ctx) {
       // Allocation and subscriptions happen only for committed mounts. Suspended
       // or abandoned renders never acquire a session, probe engines, or retain it.
       const current = get(sessionId);
+      current.reconnectChat();
       current.refs++;
       if (kind === 'buttons') current.buttons++;
       setEntry(current);
@@ -937,7 +969,6 @@ export function apply(ctx) {
         entry.composers.delete(token.current);
         if (!entry.composers.size) suspendSessionSpeech(entry);
         if (meetingEntry === entry && !entry.composers.size) {
-          void meeting.attendee.stop();
           meeting.composerReplaced();
           meetingEntry = null;
           meeting.transcript.setActive('microphone', false);
@@ -952,10 +983,7 @@ export function apply(ctx) {
       if (!entry || entry.closed || disposed) return;
       if (!input || typeof props.inputActions?.setDraft !== 'function') return;
       if (meetingEntry !== entry) {
-        if (meetingEntry) {
-          void meeting.attendee.stop();
-          suspendSessionSpeech(meetingEntry);
-        }
+        if (meetingEntry) suspendSessionSpeech(meetingEntry);
         meeting.composerReplaced();
         meetingEntry = entry;
         entry.controller.setExternalSpeechActivity(sharedSpeechActive);
@@ -991,7 +1019,7 @@ export function apply(ctx) {
       const attachSpeech = !entry.speechAttached;
       if (attachSpeech) entry.speechAttached = true;
       const inputReady =
-        voiceModeActive && !entry.controller.getSnapshot().conversation
+        !entry.parkedInput && voiceModeActive && !entry.controller.getSnapshot().conversation
           ? entry.controller.startConversation()
           : Promise.resolve();
       run(

package/src/modules/conversation/models/AttendeeClient.ts

@@ -623,7 +623,9 @@ export class AttendeeClient {
     this.queues.set(source, work);
   }
   async audio(event, generation) {
-    if (this.canDeliver && !this.canDeliver()) throw new Error('no-composer');
+    // Capture is application-owned and can outlive a composer during chat navigation.
+    // MeetingTranscript has no active destination in that gap, so results are dropped
+    // rather than ending the receiver or being written into a stale composer.
     const source = 'attendee:' + event.botId + ':' + event.participantId;
     let row = this.rows.get(source);
     if (!row) {

package/src/modules/conversation/models/meeting.ts

@@ -173,14 +173,17 @@ export class MeetingController {
     for (const input of this.inputs.values()) input.cancelAutoSend();
   }
   composerReplaced() {
-    // Capture follows navigation, but a recognition operation must not deliver
-    // an utterance begun for the previous composer into its replacement.
+    // Recording/capture is application-owned and outlives any one chat composer.
+    // Keep the audio clock and writer, but never deliver a final whose speech began
+    // before the replacement composer was committed.
+    for (const [source, job] of this.jobs) {
+      const state = this.snapshot[source];
+      job.discardResults =
+        (Number.isSafeInteger(state?.pendingTranscriptionRequests)
+          ? state.pendingTranscriptionRequests
+          : state?.pending || 0) || (state?.recognizing ? 1 : 0);
+    }
     this.cancelDelivery();
-    const job = this.jobs.get('shared');
-    if (!job || !job.recognitionAbort) return;
-    job.recognitionAbort.abort();
-    if (this.sharedTrackLive(job)) this.queueRecognitionRecovery('shared', job);
-    else void this.stop('shared').catch(() => {});
   }
   toggleTimestamps() {
     const enabled = !this.snapshot.timestamps;
@@ -206,7 +209,7 @@ export class MeetingController {
     }
     this.patch(null, { active: true });
     const abort = new AbortController();
-    const job = { abort };
+    const job = { abort, discardResults: 0 };
     this.jobs.set(source, job);
     const valid = () => this.jobs.get(source) === job && !abort.signal.aborted;
     this.patch(source, { starting: true, error: null });
@@ -291,6 +294,10 @@ export class MeetingController {
         signal: recognitionAbort.signal,
         onResult: ({ final, interim, startedAt, paragraphBefore }) => {
           if (!valid()) return;
+          if (job.discardResults > 0) {
+            job.discardResults--;
+            return;
+          }
           if (final?.trim()) job.retries = 0;
           const accepted = this.onSpeech?.(source, final || interim) !== false;
           if (accepted)

package/src/modules/core/ARCHITECTURE.md

@@ -2,6 +2,8 @@
 
 `core` owns policy and reusable voice-domain state, not DSH slot/route registration or provider-specific presentation. `coordinator.ts` serializes input transitions, owns conversation state and speech queues, and invalidates stale async work with generations. `ownership.ts` arbitrates microphone and speaker use between sessions and Settings self-tests. `transcript.ts` enforces append-only final recognition text (interim hypotheses never own or rewrite composer content), `microphone.ts` manages capture and activity, `filters.ts` handles voice commands and speech text filtering, and `settings.ts` normalizes persisted values and defaults.
 
+Recording-backed conversation input can be parked by `pauseInputForNavigation()` without aborting recognition or sealing its writer. Final recognition callbacks remain on the original coordinator and are queued while detached; `resumeInputFromNavigation()` delivers them only after composition reacquires hardware ownership and a current composer. Navigation generations reject obsolete resume completions. Stop/End/disposal clears that queue and releases input; native Browser recognition keeps its existing teardown path. Parked capture retains its start-time recognition configuration and does not retry acquisition in the background.
+
 Recording audit emits transition-only assistant wait start/end with bounded reason codes, queue/request counts and delay, playback admission, requested input gating, recognition/meter stop completion or failure, real engine playback start/end, cancellation and errors. Input-stop completion means adapter promises settled successfully, not independent physical microphone proof. Wait reasons never include text, message IDs, provider errors, or audio. Event audit does not change capture/recognition/playback policy.
 
 The `qwen/` subtree contains shared Qwen host configuration and its settings UI; recognition and speaking adapters live in their respective modules. Keep the client/host dependency graphs separate. Preserve the persisted `dsh-live-voice.settings` key and normalization compatibility. `startConversation()` enables input without cancelling playback; `stopConversationInput()` retires input without clearing accepted speech. Speaker-mode starts defer capture until playback finishes, while `takeMicrophone()` retains the explicit teardown-before-capture handoff. Full `endConversation()` remains the global lifecycle/End-command operation. Automatic speech admission depends on the assistant-speech preference, not microphone conversation mode; new segments and accepted queues continue with input closed. Distinguish capture, recognition, playback, and generation; do not conflate pausing with cancellation or resume obsolete speech after a new turn. Consult [voice lifecycle](../../../../../docs/VOICE-LIFECYCLE.md) for expected transitions.