Update app.py

app.py

@@ -415,7 +415,8 @@ class SlpMultiAgent:
                 result = await loop.run_in_executor(
                     None,
                     lambda: manager_agent.run(f"""
-                    Question: {short_question}
+                    DO NOT output any code blocks, markdown, triple backticks, <code> tags, or print statements. DO NOT output plans, thoughts, or intermediate steps. DO NOT output anything except a single call to final_answer('...').
+                    If you do, you will be penalized and your answer will be rejected.
                     
                     INSTRUCTIONS:
                     - For every question, you MUST call the most appropriate tool to get the answer.
@@ -424,7 +425,19 @@ class SlpMultiAgent:
                     - Only call the tool and then call final_answer('...').
                     - If you cannot answer, call final_answer('I don't know').
                     - STRICTLY FOLLOW THIS FORMAT. DO NOT OUTPUT ANYTHING ELSE.
-
+                    - If you reach max steps, call final_answer('I don't know').
+                    
+                    BAD EXAMPLES (DO NOT DO THIS):
+                    ```python\nfinal_answer('foo')\n```
+                    <code>final_answer('foo')</code>
+                    Thoughts: ...
+                    Plan: ...
+                    
+                    GOOD EXAMPLES (DO THIS):
+                    final_answer('Paris')
+                    final_answer('4')
+                    final_answer('I don't know')
+                    
                     Available tools:
                     - google_search(query): Search the web
                     - math_calculator(expression): Calculate math
@@ -433,7 +446,7 @@ class SlpMultiAgent:
                     - wikipedia_search(query): Search Wikipedia
                     - code_executor(code): Execute Python code
                     - web_scraper(url): Scrape web pages
-
+                    
                     EXAMPLES:
                     Q: What is the capital of France?
                     A: final_answer('Paris')
@@ -451,6 +464,8 @@ class SlpMultiAgent:
                     A: final_answer('avocado')
                     
                     STRICTLY FOLLOW THIS FORMAT. DO NOT OUTPUT ANYTHING ELSE.
+                    
+                    Question: {short_question}
                     """)
                 )
                 break  # Success, exit retry loop
@@ -498,16 +513,24 @@ class SlpMultiAgent:
             return text.strip()
         # Post-process the result to remove code blocks and extract the answer
         if result and isinstance(result, str):
+            # If agent reached max steps, return I don't know
+            if "Reached max steps" in result:
+                return "I don't know"
             cleaned = strip_code_blocks(result)
             answer = extract_final_answer(cleaned)
             if answer:
+                # If the answer is still a code block or not a final_answer, fallback
+                if answer.startswith('```') or answer.startswith('<code>') or 'final_answer' not in result:
+                    return "I don't know"
                 return answer
             # Fallback: try to extract from the original result
             answer = extract_final_answer(result)
             if answer:
+                if answer.startswith('```') or answer.startswith('<code>') or 'final_answer' not in result:
+                    return "I don't know"
                 return answer
         # Return the result from the agent
-        return result if result else "Unable to determine answer."
+        return "I don't know"
     
 def check_reasoning(final_answer, agent_memory):
     # Skip expensive validation to save costs