;yes

app.py

@@ -85,7 +85,7 @@ else:
     harmony_encoding = None
 
 # Stop tokens per Harmony spec: <|return|> (200002), <|call|> (200012)
-HARMONY_STOP_IDS = [200002, 200012]
+HARMONY_STOP_IDS = harmony_encoding.stop_tokens_for_assistant_actions() if HARMONY_AVAILABLE else []
 
 # Tokenizer is lightweight; load once
 try:
@@ -214,7 +214,7 @@ def parse_harmony_response(tokens: List[int]) -> Dict[str, str]:
         channel = msg.channel if hasattr(msg, 'channel') else "final"
         if channel not in channels:
             channels[channel] = ""
-        channels[channel] += msg.content
+        channels[channel] += "".join([getattr(part, "text", str(part)) for part in (msg.content if isinstance(msg.content, list) else [msg.content])])
     
     # Ensure we have a final channel
     if "final" not in channels:
@@ -309,17 +309,8 @@ def zerogpu_generate(full_prompt,
             if "attention_mask" not in inputs:
                 inputs["attention_mask"] = torch.ones_like(inputs["input_ids"], dtype=torch.long, device=device)
         # Generate
-        # Build EOS list: Harmony stops + tokenizer EOS
-        eos_ids = []
-        if HARMONY_AVAILABLE:
-            eos_ids.extend(HARMONY_STOP_IDS)
-        tok_eos = tokenizer.eos_token_id
-        if tok_eos is not None:
-            if isinstance(tok_eos, int):
-                eos_ids.append(tok_eos)
-            else:
-                eos_ids.extend(list(tok_eos))
-        eos_ids = list(dict.fromkeys(eos_ids))
+        # Build EOS list: use ONLY Harmony assistant-action stops (per OpenAI docs)
+        eos_ids = HARMONY_STOP_IDS if HARMONY_AVAILABLE else tokenizer.eos_token_id
 
         out_ids = model.generate(
             **inputs,