Message property sent to llm (#1273)

* Add to_memory for observations

* Don't send 'message' property to the llm from observation; fix other agents

agenthub/SWE_agent/agent.py

@@ -38,7 +38,7 @@ class SWEAgent(Agent):
 
     def _remember(self, action: Action, observation: Observation) -> None:
         """Agent has a limited memory of the few steps implemented as a queue"""
-        memory = MEMORY_FORMAT(action.to_dict(), observation.to_dict())
+        memory = MEMORY_FORMAT(action.to_memory(), observation.to_memory())
         self.running_memory.append(memory)
 
     def _think_act(self, messages: List[dict]) -> tuple[Action, str]:

agenthub/monologue_agent/agent.py

@@ -162,7 +162,7 @@ class MonologueAgent(Agent):
                     observation = BrowserOutputObservation(
                         content=thought, url='', screenshot=''
                     )
-                self._add_event(observation.to_dict())
+                self._add_event(observation.to_memory())
                 output_type = ''
             else:
                 action: Action = NullAction()
@@ -205,7 +205,7 @@ class MonologueAgent(Agent):
         self._initialize(state.plan.main_goal)
         for prev_action, obs in state.updated_info:
             self._add_event(prev_action.to_memory())
-            self._add_event(obs.to_dict())
+            self._add_event(obs.to_memory())
 
         state.updated_info = []
 

agenthub/planner_agent/prompt.py

@@ -171,7 +171,7 @@ def get_prompt(plan: Plan, history: List[Tuple[Action, Observation]]) -> str:
             history_dicts.append(action.to_memory())
             latest_action = action
         if not isinstance(observation, NullObservation):
-            observation_dict = observation.to_dict()
+            observation_dict = observation.to_memory()
             if (
                 'extras' in observation_dict
                 and 'screenshot' in observation_dict['extras']

opendevin/observation/base.py

@@ -15,6 +15,12 @@ class Observation:
         return self.content
 
     def to_dict(self) -> dict:
+        """Converts the observation to a dictionary and adds user message."""
+        memory_dict = self.to_memory()
+        memory_dict['message'] = self.message
+        return memory_dict
+
+    def to_memory(self) -> dict:
         """Converts the observation to a dictionary."""
         extras = copy.deepcopy(self.__dict__)
         content = extras.pop('content', '')
@@ -23,7 +29,6 @@ class Observation:
             'observation': observation,
             'content': content,
             'extras': extras,
-            'message': self.message,
         }
 
     @property

tests/test_observation_serialization.py

@@ -1,6 +1,6 @@
-import pytest
 from opendevin.observation import observation_from_dict, Observation, CmdOutputObservation
 
+
 def test_observation_serialization_deserialization():
     original_observation_dict = {
         'observation': 'run',
@@ -12,6 +12,9 @@ def test_observation_serialization_deserialization():
     assert isinstance(observation_instance, Observation), 'The observation instance should be an instance of Action.'
     assert isinstance(observation_instance, CmdOutputObservation), 'The observation instance should be an instance of AgentThinkAction.'
     serialized_observation_dict = observation_instance.to_dict()
+    serialized_observation_memory = observation_instance.to_memory()
     assert serialized_observation_dict == original_observation_dict, 'The serialized observation should match the original observation dict.'
+    original_observation_dict.pop('message')
+    assert serialized_observation_memory == original_observation_dict, 'The serialized observation in memory should match the original observation dict.'
 
 # Additional tests for various observation subclasses can be included here