Fix issue #2237: Properly handle LLM output with both Action and Final Answer

Co-Authored-By: Joe Moura <joao@crewai.com>

src/crewai/agents/crew_agent_executor.py

@@ -232,7 +232,14 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                 self._format_answer(answer)
             except OutputParserException as e:
                 if FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE in e.error:
-                    answer = answer.split("Observation:")[0].strip()
+                    # If both Action and Final Answer are present, prioritize the Action
+                    # by removing the Final Answer part
+                    if "Final Answer:" in answer:
+                        parts = answer.split("Final Answer:")
+                        answer = parts[0].strip()
+                    # If that doesn't work, try splitting at Observation
+                    elif "Observation:" in answer:
+                        answer = answer.split("Observation:")[0].strip()
 
         return self._format_answer(answer)
 

tests/agents/test_action_final_answer_error.py

@@ -0,0 +1,78 @@
+import pytest
+from unittest.mock import MagicMock
+
+from crewai.agents.crew_agent_executor import CrewAgentExecutor
+from crewai.agents.parser import AgentAction, AgentFinish, OutputParserException, FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE
+
+
+def test_process_llm_response_with_action_and_final_answer():
+    """Test that _process_llm_response correctly handles outputs with both Action and Final Answer."""
+    # Create a mock LLM
+    mock_llm = MagicMock()
+    mock_llm.supports_stop_words.return_value = False
+    
+    # Create a mock agent
+    mock_agent = MagicMock()
+    
+    # Create a CrewAgentExecutor instance
+    executor = CrewAgentExecutor(
+        llm=mock_llm,
+        task=MagicMock(),
+        crew=MagicMock(),
+        agent=mock_agent,
+        prompt={},
+        max_iter=5,
+        tools=[],
+        tools_names="",
+        stop_words=[],
+        tools_description="",
+        tools_handler=MagicMock(),
+    )
+    
+    # Test case 1: Output with both Action and Final Answer, with Final Answer after Action
+    output_with_both = """
+    Thought: I need to search for information and then provide an answer.
+    Action: search
+    Action Input: what is the temperature in SF?
+    Final Answer: The temperature is 100 degrees
+    """
+    
+    # Mock the _format_answer method to first raise an exception and then return a valid result
+    format_answer_mock = MagicMock()
+    format_answer_mock.side_effect = [
+        OutputParserException(FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE),
+        AgentAction(thought="", tool="search", tool_input="what is the temperature in SF?", text=""),
+    ]
+    executor._format_answer = format_answer_mock
+    
+    # Process the response
+    result = executor._process_llm_response(output_with_both)
+    
+    # Verify that the result is an AgentAction
+    assert isinstance(result, AgentAction)
+    assert result.tool == "search"
+    assert result.tool_input == "what is the temperature in SF?"
+    
+    # Test case 2: Output with both Action and Final Answer, with Observation in between
+    output_with_observation = """
+    Thought: I need to search for information.
+    Action: search
+    Action Input: what is the temperature in SF?
+    Observation: The temperature in SF is 100 degrees.
+    Final Answer: The temperature is 100 degrees
+    """
+    
+    # Reset the mock
+    format_answer_mock.reset_mock()
+    format_answer_mock.side_effect = [
+        OutputParserException(FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE),
+        AgentAction(thought="", tool="search", tool_input="what is the temperature in SF?", text=""),
+    ]
+    
+    # Process the response
+    result = executor._process_llm_response(output_with_observation)
+    
+    # Verify that the result is an AgentAction
+    assert isinstance(result, AgentAction)
+    assert result.tool == "search"
+    assert result.tool_input == "what is the temperature in SF?"