sonnets last attempt

before I cut my losses

app.py

@@ -103,15 +103,17 @@ def is_valid_length(text: str, min_words: int = 3, max_words: int = 50) -> bool:
 def is_verbatim_repetition(
     new_text: str, history: List[Tuple[str, str]], system_prompt: str
 ) -> bool:
-    """Check if text is exact repetition of prior user turn or system prompt (Guardrail 4)."""
+    """Check if text is exact repetition of prior user turn or system prompt (Guardrail 4).
+
+    History format: (assistant_msg, user_msg) - so user messages are in position 1
+    """
     new_text_normalized = new_text.strip().lower()
 
-    # Check against system prompt
     if new_text_normalized == system_prompt.strip().lower():
         return True
 
-    # Check against previous user messages (UserLM's prior outputs)
-    for user_msg, _ in history:
+    # User messages are now in position 1 of the tuple
+    for _, user_msg in history:
         if user_msg and new_text_normalized == user_msg.strip().lower():
             return True
 
@@ -179,32 +181,31 @@ def generate_next_turn(
     top_p: float,
 ):
     """
-    Generate the next user message from UserLM.
-
-    Flow:
-    - If chat_history is empty: Generate first user message
-    - If chat_history exists:
-        1. Add assistant's response to last turn
-        2. Generate next user message
-
-    Tuple structure: (user_message_from_userlm, assistant_response_from_human)
-    - Position 0 (left): UserLM's messages
-    - Position 1 (right): Human's assistant responses
+    Tuple: (human_assistant, userlm_user)
+    Testing if Gradio shows first element on RIGHT not LEFT
     """
 
-    # If we have an assistant response, add it to the last turn
+    # If we have an assistant response, fill in position 0 of last turn
     if assistant_response.strip() and len(chat_history) > 0:
-        last_user_msg, _ = chat_history[-1]
-        chat_history = chat_history[:-1] + [(last_user_msg, assistant_response.strip())]
+        _, last_user_msg = chat_history[-1]
+        chat_history = chat_history[:-1] + [(assistant_response.strip(), last_user_msg)]
+
+    # Build messages for UserLM - user msgs are now in position 1
+    messages = []
+    if system_prompt.strip():
+        messages.append({"role": "system", "content": system_prompt.strip()})
 
-    # Build messages for UserLM
-    messages = build_messages_for_userlm(system_prompt, chat_history)
+    for asst_msg, user_msg in chat_history:
+        if user_msg:
+            messages.append({"role": "user", "content": user_msg})
+        if asst_msg:
+            messages.append({"role": "assistant", "content": asst_msg})
 
     # Generate next user message
     try:
         user_msg = generate_user_message(
             messages,
-            chat_history,
+            [(u, a) for a, u in chat_history],  # Swap for repetition check
             system_prompt,
             max_new_tokens=max_new_tokens,
             temperature=temperature,
@@ -213,14 +214,10 @@ def generate_next_turn(
     except Exception as e:
         user_msg = f"(Generation error: {e})"
 
-    # Add new user message to history (with empty assistant slot)
-    new_history = chat_history + [(user_msg, "")]
-
-    # Determine button text for next action
-    needs_assistant_response = True
-    button_text = "Generate Next User Message"
+    # Add: (empty_assistant, new_user_msg)
+    new_history = chat_history + [("", user_msg)]
 
-    return "", new_history, button_text
+    return "", new_history, "Generate Next User Message"
 
 
 def clear_conversation():
@@ -256,8 +253,8 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
     chatbot = gr.Chatbot(
         height=420,
         label="Conversation",
-        type="tuples",
-        # Left side = UserLM (simulated user), Right side = You (playing assistant)
+        type="messages",  # Changed from tuples to have more control
+        # Will manually format messages with role attribute
     )
 
     with gr.Row():