dsh: Option 1 gateway — home-shaped chat requests (announce, HA, shopping, calendar, music, messaging) route raw to Jervis; normal chat untouched

dsh/app.py

@@ -42,13 +42,7 @@ 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.\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.",
+    "secrets. If asked something unsafe, decline politely.",
 )
 
 # --- vision (image ingest) -> OpenRouter gpt-5 (works; OmniRoute auto/best-vision
@@ -147,62 +141,43 @@ _SUM_HEAD = (
 )
 
 
-_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:
+def _looks_homeish(msg: str) -> bool:
+    """Broad gate: hand the message to Jervis when it talks about the home
+    (announcements, HA devices, shopping/calendar/timers, music, messaging,
+    fish, time/weather). Jervis decides answer-vs-act and speaks confirmations."""
     m = (msg or "").lower()
     return any(h in m for h in (
+        # speak / announce
         "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",
+        "say on ", "say over ", "say through ", "on the speakers",
+        "over the speakers", "through the speakers",
+        # home assistant actions
+        "turn on ", "turn off ", "switch on ", "switch off ",
+        "lights", "lamp", "dim ", "scene", "thermostat", "climate",
+        "aircon", "heating", "curtains", "blinds", "garage",
+        "feed the fish",
+        # lists / calendar / reminders
+        "shopping", "bunnings", "add to my calendar", "calendar",
+        "remind me", "reminder", "set a timer", "timer for ",
+        # music
+        "play some music", "play music", "play a song", "play ",
+        "pause the music", "music ", "spotify", "mopidy", "volume",
+        # messaging
+        "text ", "whatsapp", "message ", "notify ", "email ",
     ))
 
 
-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}
+async def _jervis_act(text: str) -> str | None:
+    """Forward a raw home command to Jervis (/voice). Returns Jervis's spoken
+    reply for the chat bubble, or None if Jervis is unreachable/errored."""
     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"]
+            r = await client.post(VOICE_AGENT_URL, json={"text": (text or "")[:500]})
+        if r.status_code != 200:
+            return None
+        return (r.json().get("reply") or "").strip()[:500] or None
     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)
@@ -370,10 +345,11 @@ async def chat(message: Annotated[str, Form()],
             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)
+            # Home pass: announce/HA/shopping/calendar/music/messaging requests
+            # go straight to Jervis (full HA tool loop, speaks confirmations).
+            # Jervis decides answer-vs-act; dsh just relays its reply.
+            if _looks_homeish(message):
+                spoke = await _jervis_act(message)
                 if spoke:
                     hist.append({"role": "assistant", "content": spoke})
                     save_session(sid, hist, session.get("model"))