dsh-agent-voice 0.1.4 → 0.1.5

package/lib/client.js

@@ -154,21 +154,65 @@ 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(() => {
-          if (!statusJson) return;
-          let p = null;
-          try { p = JSON.parse(statusJson); } catch { return; }
-          if (!p || !p.dataUrl || !p.id) return;
-          if (p.id === lastPlayed) return;
-          setLastPlayed(p.id);
-          const audio = new Audio(p.dataUrl);
+          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) {
+          // Audio is served via HTTP (settings stays small). audioPath is the
+          // absolute file path; the route is the public /dsh-agent-voice/audio/<name>.
+          let url = p.audioUrl;
+          if (!url && p.audioPath) {
+            const name = String(p.audioPath).split("/").pop();
+            url = "/dsh-agent-voice/audio/" + name;
+          }
+          const audio = new Audio(url);
           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(() => {});
+          audio.play().catch((err) => {
+            // Autoplay still blocked: keep the toast and retry on next gesture.
+            pendingRef.current = p;
+          });
+        }
+
+        React.useEffect(() => {
+          if (!statusJson) return;
+          let p = null;
+          try { p = JSON.parse(statusJson); } catch { return; }
+          if (!p || (!p.dataUrl && !p.audioPath && !p.audioUrl) || !p.id) return;
+          if (p.id === lastPlayed) return;
+          setLastPlayed(p.id);
+          playAnnouncement(p);
         }, [statusJson, lastPlayed]);
 
         // Disabled = silent (config read from the same source the host uses).

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";
 
@@ -96,6 +97,30 @@ async function postWorker(path, body) {
 }
 
 function apply(ctx) {
+  // Serve generated audio to the client over HTTP (keeps WAVs out of the
+  // settings document entirely).
+  const ws = ctx.get("webServer");
+  if (ws) {
+    ws.register({
+      kind: "prefix",
+      path: "/dsh-agent-voice/audio/",
+      handler: async (req, res) => {
+        const name = decodeURIComponent(req.url.split("/").pop() || "");
+        if (!/^av-\d+-[a-z0-9]+\.wav$/.test(name)) {
+          res.writeHead(404).end("not found");
+          return;
+        }
+        try {
+          const data = await readFile(join(AUDIO_DIR, name));
+          res.writeHead(200, { "Content-Type": "audio/wav", "Content-Length": data.length });
+          res.end(data);
+        } catch {
+          res.writeHead(404).end("not found");
+        }
+      },
+    });
+  }
+
   let source = () => ({ ...DEFAULTS });
   installSettingsSection(ctx, NS, SCHEMA, DEFAULTS, {
     setSource: (getCurrent) => { source = () => ({ ...DEFAULTS, ...(getCurrent() ?? {}) }); },
@@ -110,30 +135,49 @@ function apply(ctx) {
     response: z.string(),
   }), { base: {} });
 
+  // Audio files are served to the client over a webServer route instead of
+  // being base64-inlined into the settings document. A full WAV in settings
+  // would bloat the shared file by hundreds of KB and stall every settings
+  // round-trip (which is exactly what broke the UI before this fix).
   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");
+    await mkdir(AUDIO_DIR, { recursive: true });
+    const name = `av-${Date.now()}-${Math.random().toString(36).slice(2, 8)}.wav`;
+    const file = join(AUDIO_DIR, name);
+    await writeFile(file, await readFile(res.audio));
+    return file;
   }
 
   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 file = 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,
+      audioPath: file, // the client fetches /dsh-agent-voice/audio/<name>
       at: Date.now(),
     };
     await ctx.settings.replace(STATUS_NS, { json: JSON.stringify(payload) });
+    cleanupOldAudio().catch(() => {});
     return { ok: true, kind, text };
   }
 
+  async function cleanupOldAudio() {
+    const { readdir, rm } = await import("node:fs/promises");
+    let files = [];
+    try { files = (await readdir(AUDIO_DIR)).filter((f) => f.endsWith(".wav")).sort(); } catch { return; }
+    while (files.length > 10) {
+      const old = files.shift();
+      try { await rm(join(AUDIO_DIR, old), { force: true }); } catch {}
+    }
+  }
+
   async function runAction(req) {
     const cfg = source();
     const a = req.action;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-agent-voice",
-  "version": "0.1.4",
+  "version": "0.1.5",
   "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",