dsh-agent-voice 0.1.4 → 0.1.6

package/lib/client.js

@@ -154,21 +154,67 @@ window.__ModuleLoader__.load({
         const [lastPlayed, setLastPlayed] = React.useState(null);
         const [toast, setToast] = React.useState(null);
         const toastTimer = React.useRef(null);
+        const pendingRef = React.useRef(null);
+
+        // Browsers block audio.play() before any user gesture (autoplay
+        // policy). Unlock audio on the first click/keypress, then replay any
+        // announcement that was blocked. This makes Agent Voice audible even
+        // though the tab view never received the gesture.
+        React.useEffect(() => {
+          const unlock = () => {
+            // A silent short audio primes the audio pipeline; subsequent
+            // play() calls are then allowed by the browser.
+            try {
+              const silent = new Audio("data:audio/wav;base64,UklGRiQAAABXQVZFZm10IBAAAAABAAEAQB8AAEAfAAABAAgAZGF0YQAAAAA=");
+              silent.volume = 0.0001;
+              silent.play().catch(() => {});
+            } catch {}
+            if (pendingRef.current) {
+              const p = pendingRef.current;
+              pendingRef.current = null;
+              playAnnouncement(p);
+            }
+          };
+          window.addEventListener("pointerdown", unlock, { once: true });
+          window.addEventListener("keydown", unlock, { once: true });
+          window.addEventListener("pointerup", unlock, { once: true });
+          return () => {
+            window.removeEventListener("pointerdown", unlock);
+            window.removeEventListener("keydown", unlock);
+            window.removeEventListener("pointerup", unlock);
+          };
+        }, []);
+
+        function playAnnouncement(p) {
+          // The WAV is fetched one-shot via the action round-trip (base64),
+          // then played. Settings never carries the audio bytes.
+          const doPlay = (dataUrl) => {
+            const audio = new Audio(dataUrl);
+            audio.volume = Math.max(0, Math.min(1, Number(p.volume ?? 1)));
+            const kindLabel = p.kind === "attention" ? "🔔 Pozor: " : (p.kind === "task" ? "✅ " : "🔊 ");
+            setToast(kindLabel + p.text);
+            if (toastTimer.current) clearTimeout(toastTimer.current);
+            toastTimer.current = setTimeout(() => setToast(null), 6000);
+            audio.play().catch(() => {
+              // Autoplay still blocked: keep the toast and retry on next gesture.
+              pendingRef.current = p;
+            });
+          };
+          if (p.dataUrl) { doPlay(p.dataUrl); return; }
+          // One-shot poll for the audio bytes by announcement id.
+          action({ action: "poll", id: p.id }).then((r) => {
+            if (r && r.ok && r.result && r.result.found && r.result.dataUrl) doPlay(r.result.dataUrl);
+          }).catch(() => {});
+        }
 
         React.useEffect(() => {
           if (!statusJson) return;
           let p = null;
           try { p = JSON.parse(statusJson); } catch { return; }
-          if (!p || !p.dataUrl || !p.id) return;
+          if (!p || !p.id) return;
           if (p.id === lastPlayed) return;
           setLastPlayed(p.id);
-          const audio = new Audio(p.dataUrl);
-          audio.volume = Math.max(0, Math.min(1, Number(p.volume ?? 1)));
-          const kindLabel = p.kind === "attention" ? "🔔 Pozor: " : (p.kind === "task" ? "✅ " : "🔊 ");
-          setToast(kindLabel + p.text);
-          if (toastTimer.current) clearTimeout(toastTimer.current);
-          toastTimer.current = setTimeout(() => setToast(null), 6000);
-          audio.play().catch(() => {});
+          playAnnouncement(p);
         }, [statusJson, lastPlayed]);
 
         // Disabled = silent (config read from the same source the host uses).
@@ -193,7 +239,7 @@ window.__ModuleLoader__.load({
         id: "agent-voice-player",
         order: 100,
         label: () => "🔊 Agent Voice přehrávač",
-        inject: () => ({ config, status }),
+        inject: () => ({ action, config, status }),
         children: {},
       }, Player));
 

package/lib/index.js

@@ -24,6 +24,7 @@ const STATUS_NS = settingsNamespace("agent-voice-status");
 const DSH_HOME = process.env.DSH_HOME ?? join(homedir(), ".dsh");
 const SHARED_VENV = join(DSH_HOME, "voice-studio-venv"); // reuse the installed XTTS/Piper venv
 const VENV_DIR = join(DSH_HOME, "agent-voice-venv");
+const AUDIO_DIR = join(DSH_HOME, "agent-voice-outputs");
 const WORKER_URL = "http://127.0.0.1:7863";
 const PYTHON = process.env.PYTHON ?? "/opt/homebrew/opt/python@3.11/bin/python3.11";
 
@@ -47,6 +48,10 @@ const DEFAULTS = {
 
 const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
 
+// One-shot audio cache keyed by announcement id: the client polls action=poll
+// to fetch the WAV bytes (base64) for a given id, then the entry expires.
+const audioCache = new Map();
+
 let workerProcess = null;
 
 async function ensureWorker() {
@@ -110,27 +115,32 @@ function apply(ctx) {
     response: z.string(),
   }), { base: {} });
 
+  // Audio bytes travel to the client via the ACTION round-trip (one-shot,
+  // never persisted): the client calls action=poll with an announcement id and
+  // the host returns the base64 WAV. The STATUS namespace carries only small
+  // metadata (id + text), so the shared settings document stays tiny.
   async function synthesize(text, voice, language) {
     await ensureWorker();
     const res = await postWorker("/tts", { text, voice, language });
-    const bytes = await readFile(res.audio);
-    return "data:audio/wav;base64," + Buffer.from(bytes).toString("base64");
+    return await readFile(res.audio);
   }
 
   async function announce(kind, text) {
     const cfg = source();
     if (!cfg.enabled) return { ok: false, skipped: "disabled" };
-    const dataUrl = await synthesize(text, cfg.voice, cfg.language);
+    const wav = await synthesize(text, cfg.voice, cfg.language);
+    const id = String(Date.now()) + "-" + Math.random().toString(36).slice(2);
     const payload = {
-      id: String(Date.now()) + "-" + Math.random().toString(36).slice(2),
+      id,
       kind,
       text,
       voice: cfg.voice,
       volume: cfg.volume,
-      dataUrl,
       at: Date.now(),
     };
     await ctx.settings.replace(STATUS_NS, { json: JSON.stringify(payload) });
+    audioCache.set(id, wav);
+    setTimeout(() => audioCache.delete(id), 60000).unref?.();
     return { ok: true, kind, text };
   }
 
@@ -170,6 +180,14 @@ function apply(ctx) {
       const text = req.text || "Ahoj, tady agent. Test hlasu proběhl úspěšně.";
       return announce("test", text);
     }
+    if (a === "poll") {
+      // One-shot audio fetch: the client asks for the WAV of an announcement
+      // id and receives it as base64 (never persisted to settings).
+      if (!req.id) throw new Error("poll requires id");
+      const wav = audioCache.get(req.id);
+      if (!wav) return { id: req.id, found: false };
+      return { id: req.id, found: true, dataUrl: "data:audio/wav;base64," + Buffer.from(wav).toString("base64") };
+    }
     throw new Error("unknown action: " + a);
   }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-agent-voice",
-  "version": "0.1.4",
+  "version": "0.1.6",
   "description": "DeepSeek Harness plugin: gives the agent a voice — it reads completed tasks and attention requests aloud (never reasoning or interim reports) via local Piper / XTTS v2, with a settings panel and a curated set of voices.",
   "type": "module",
   "main": "lib/index.js",