Update app.py

app.py

@@ -143,15 +143,18 @@ def create_agent():
 
 def respond(message, history, temperature, max_new_tokens, max_tokens, multi_agent, conversation, max_round):
     updated_history = history + [{"role": "user", "content": message}]
-    response_generator = agent.run_gradio_chat(updated_history, temperature, max_new_tokens, max_tokens, multi_agent, conversation, max_round)
+    response_generator = agent.run_gradio_chat(
+        updated_history, temperature, max_new_tokens, max_tokens, multi_agent, conversation, max_round
+    )
+
     collected = ""
     for chunk in response_generator:
         if isinstance(chunk, dict):
             collected += chunk.get("content", "")
         else:
             collected += str(chunk)
-    updated_history.append({"role": "assistant", "content": collected})
-    return updated_history
+
+    return history + [{"role": "user", "content": message}, {"role": "assistant", "content": collected}]
 
 def create_demo(agent):
     with gr.Blocks(css=chat_css) as demo: