Upload app.py with huggingface_hub

app.py

@@ -471,8 +471,8 @@ if ON_HF:
         )
         _hf_model.eval()
 
-        @spaces.GPU(duration=30)
-        def _hf_gpu_generate(messages, max_new_tokens=150):
+        @spaces.GPU(duration=25)
+        def _hf_gpu_generate(messages, max_new_tokens=96):
             model = _hf_model.to("cuda")
             try:
                 # Ask the model not to emit chain-of-thought, if it supports it
@@ -779,7 +779,7 @@ class Joe:
         self.running = False
         self.history = deque(maxlen=30)
         self.current_data = {}
-        self.current_message = ("Joe", "Starting...", "", "")
+        self.current_message = ("Hi, I am Joe", "Press Start", "to wake me up", "")
         self.current_ascii = ""
         self.current_context = ""
         self.current_thinking = ""
@@ -1243,12 +1243,19 @@ How do I feel right now?"""
                 frags = re.split(r"[.!?\n|]+", text)
                 lines = [f.strip() for f in frags if len(f.strip()) > 2][:6]
 
+            # Normalize to a flat list of strings (model JSON can be messy:
+            # bools, ints, nested values, or a bare string instead of a list).
+            if isinstance(lines, (str, bytes)):
+                lines = [lines]
+            elif not isinstance(lines, list):
+                lines = []
+
             if lines:
                 # Deduplicate and clean
                 seen = set()
                 clean = []
                 for l in lines:
-                    l = l.strip().strip('.')
+                    l = str(l).strip().strip('.')
                     if l and l.lower() not in seen and len(l) > 2:
                         seen.add(l.lower())
                         clean.append(l)
@@ -1449,7 +1456,7 @@ How do I feel right now?"""
             data.get("minute", 0) // 5,  # Refresh every 5 min
             data.get("user_pattern", ""),
             data.get("audio", {}).get("type", ""),
-            int(time.time()) // 30,  # Force refresh every 30s
+            int(time.time()) // 90,  # Throttle: at most one new generation ~90s
         ]
         return hashlib.md5(str(key_parts).encode()).hexdigest()[:8]
 
@@ -1545,7 +1552,10 @@ def connect():
 
 def start():
     agent.running = True
-    threading.Thread(target=agent.agent_loop, daemon=True).start()
+    # On HF the Gradio Timer drives the agent (ZeroGPU needs a request context);
+    # only spin a background loop for the local hardware build.
+    if not ON_HF:
+        threading.Thread(target=agent.agent_loop, daemon=True).start()
     return f"Agent started ({LLM_MODE})"
 
 def stop():
@@ -1626,7 +1636,10 @@ def hf_step():
     """One agent iteration driven by the Gradio Timer. On HF Spaces the LLM
     must be called from a request context (not a background thread) so ZeroGPU
     can allocate a GPU — that's what this provides. Returns dashboard outputs."""
-    agent.running = True
+    # Only consume GPU when explicitly running (Start button). Idle/abandoned
+    # browser tabs just refresh the display and cost zero GPU.
+    if not agent.running:
+        return refresh_all()
     try:
         data = agent.collect_data()
         data["audio"] = agent.audio.get_context()