dsh: speak tool — announce over house speakers via Jervis (function-calling gate, zero latency for normal chat)

dsh/app.py

@@ -42,7 +42,13 @@ SYSTEM_PROMPT = os.getenv(
     "DSH_SYSTEM_PROMPT",
     f"You are {USER_NAME}'s helpful DeepSeek assistant on the family home lab. "
     "Be clear, safe and concise. Never expose server file contents or system "
-    "secrets. If asked something unsafe, decline politely.",
+    "secrets. If asked something unsafe, decline politely.\n\n"
+    "You can speak aloud through the house speakers using the speak function "
+    "when the user asks to announce/say something over them (e.g. 'announce "
+    "X', 'tell Finn that...', 'say dinner is ready', or requests a voice "
+    "profile like Donald Trump or HAL 9000). Pass the exact message as text "
+    "and the requested voice when given. Only fall back to writing the message "
+    "in chat if the tool is unavailable.",
 )
 
 # --- vision (image ingest) -> OpenRouter gpt-5 (works; OmniRoute auto/best-vision
@@ -52,6 +58,9 @@ VISION_MODEL = os.getenv("DSH_VISION_MODEL", "openai/gpt-5")
 
 # --- existing speech->text (voice_whisper on .13:5000, faster-whisper). Do NOT edit that service ---
 WHISPER_URL = os.getenv("DSH_WHISPER_URL", "http://192.168.20.13:5000/transcribe")
+
+# --- Jervis voice agent (.13:8501) — local agent that acts + speaks via Snapcast ---
+VOICE_AGENT_URL = os.getenv("DSH_VOICE_AGENT_URL", "http://192.168.20.13:8501/voice")
 VISION_KEY = os.getenv("DSH_VISION_KEY", "")
 IMAGE_MODEL = os.getenv("DSH_IMAGE_MODEL", "openai/gpt-image-1")
 VIDEO_MODEL = os.getenv("DSH_VIDEO_MODEL", "z-ai/glm-5v-turbo")
@@ -138,7 +147,63 @@ _SUM_HEAD = (
 )
 
 
-def _strip_html(s: str) -> str:
+_SPEAK_TOOLS = [{"type": "function", "function": {
+    "name": "speak",
+    "description": "Speak text aloud through the house speakers via the Jervis voice agent (Snapcast). Use when the user wants an announcement, wants someone told something out loud, or asks for a voice profile.",
+    "parameters": {"type": "object", "properties": {
+        "text": {"type": "string", "description": "The exact text to say"},
+        "voice": {"type": "string", "description": "Optional voice profile name e.g. 'Donald Trump', 'HAL 9000', 'Picard'. Omit for the default voice."}
+    }, "required": ["text"]}
+}}]
+
+
+def _looks_speakish(msg: str) -> bool:
+    m = (msg or "").lower()
+    return any(h in m for h in (
+        "announce", "speak", "speaker", "shout", "out loud",
+        "say on ", "say over ", "say through ", "over the speakers",
+        "on the speakers", "through the speakers",
+        "tell finn", "tell jo", "tell harry", "tell sam", "tell the family",
+    ))
+
+
+async def _tool_speak(hist: list, model: str, headers: dict) -> str | None:
+    """Tool-decide pass: offer the model a 'speak' function. Returns a display
+    string if an announcement was made, else None (no speak requested)."""
+    payload = {"model": model, "stream": False, "temperature": 0.4,
+               "tools": _SPEAK_TOOLS, "tool_choice": "auto",
+               "messages": [{"role": "system", "content": SYSTEM_PROMPT}] + hist}
+    try:
+        async with httpx.AsyncClient(timeout=120) as client:
+            r = await client.post(f"{LLM_BASE.rstrip('/')}/chat/completions",
+                                  json=payload, headers=headers)
+            r.raise_for_status()
+            msg = r.json()["choices"][0]["message"]
+    except Exception:
+        return None
+    for c in msg.get("tool_calls") or []:
+        fn = c.get("function") or {}
+        if fn.get("name") != "speak":
+            continue
+        try:
+            args = json.loads(fn.get("arguments") or "{}")
+        except Exception:
+            args = {}
+        text = (args.get("text") or "").strip()[:400]
+        if not text:
+            continue
+        voice = (args.get("voice") or "").strip()
+        command = (f"Speak in the {voice} voice: {text}"
+                   if voice and voice.lower() not in ("default",) else text)
+        try:
+            async with httpx.AsyncClient(timeout=90) as client:
+                rr = await client.post(VOICE_AGENT_URL, json={"text": command})
+                ok = rr.status_code == 200
+        except Exception as exc:
+            return f"⚠️ Couldn't reach Jervis to speak: {exc}"
+        return f"🔊 {text}" if ok else "⚠️ Jervis couldn't speak it just now."
+    return None
+
     s = re.sub(r"(?is)<(script|style|head|noscript|svg)[^>]*>.*?</\1>", " ", s)
     s = re.sub(r"(?s)<[^>]+>", " ", s)
     return _html.unescape(re.sub(r"\s+", " ", s)).strip()
@@ -301,6 +366,19 @@ async def chat(message: Annotated[str, Form()],
             active_model = sel or LLM_MODEL
             if active_model not in MODEL_CHOICES:
                 active_model = LLM_MODEL
+            headers = {"Authorization": f"Bearer {key}"} if key else {}
+            if base == LLM_BASE:
+                # OmniRoute requires a session id for opencode-go models (else 400 "MissingSessionID")
+                headers["x-opencode-session"] = f"dsh-{os.getpid()}-{uuid4().hex[:8]}"
+            # Speak pass: only when the user sounds like they want an announcement.
+            # Non-streaming tool-decide, then stream a short confirmation.
+            if _looks_speakish(message):
+                spoke = await _tool_speak(hist, active_model, headers)
+                if spoke:
+                    hist.append({"role": "assistant", "content": spoke})
+                    save_session(sid, hist, session.get("model"))
+                    yield f"data: {json.dumps({'c': spoke})}\n\n"
+                    return
             payload = {
                 "model": active_model,
                 "messages": [{"role": "system", "content": SYSTEM_PROMPT}] + hist,
@@ -573,6 +651,33 @@ async def generate_image(prompt: Annotated[str, Form()]) -> dict:
         return {"ok": False, "error": f"{type(exc).__name__}: {exc}"}
 
 
+@app.post("/api/tool/speak")
+async def speak_text(text: Annotated[str, Form()], voice: Annotated[str, Form()] = "") -> dict:
+    """Speak text through the house speakers (Jervis voice agent on .13:8501).
+    Acts and speaks via Snapcast. Optional 'voice' names a profile e.g.
+    'Donald Trump', 'HAL 9000', 'Picard'."""
+    text = (text or "").strip()[:500]
+    if not text:
+        return {"ok": False, "error": "No text to speak."}
+    command = text
+    if voice and voice.strip().lower() not in ("default", ""):
+        command = f"Speak in the {voice.strip()} voice: {text}"
+    try:
+        async with httpx.AsyncClient(timeout=60) as client:
+            r = await client.post(VOICE_AGENT_URL, json={"text": command})
+            reply = ""
+            if r.status_code == 200:
+                try:
+                    reply = (r.json().get("reply") or "")[:300]
+                except Exception:  # noqa: BLE001
+                    reply = ""
+            else:
+                return {"ok": False, "error": f"Jervis returned HTTP {r.status_code}"}
+        return {"ok": True, "spoken": command[:200], "reply": reply}
+    except Exception as exc:  # noqa: BLE001
+        return {"ok": False, "error": f"Speak failed: {exc}"}
+
+
 
 MODEL_CHOICES = ["auto/best-chat",
                  "deepseek/deepseek-v4-flash",