Handle error observations in codeact (#3383)

* Handle error observations in codeact

* Remove comments

agenthub/codeact_agent/codeact_agent.py

@@ -22,6 +22,7 @@ from opendevin.events.observation import (
     CmdOutputObservation,
     IPythonRunCellObservation,
 )
+from opendevin.events.observation.error import ErrorObservation
 from opendevin.events.observation.observation import Observation
 from opendevin.events.serialization.event import truncate_content
 from opendevin.llm.llm import LLM
@@ -169,7 +170,14 @@ class CodeActAgent(Agent):
                 str(obs.outputs), max_message_chars
             )
             return Message(role='user', content=[TextContent(text=text)])
-        return None
+        elif isinstance(obs, ErrorObservation):
+            text = 'OBSERVATION:\n' + truncate_content(obs.content, max_message_chars)
+            text += '\n[Error occurred in processing last action]'
+            return Message(role='user', content=[TextContent(text=text)])
+        else:
+            # If an observation message is not returned, it will cause an error
+            # when the LLM tries to return the next message
+            raise ValueError(f'Unknown observation type: {type(obs)}')
 
     def reset(self) -> None:
         """Resets the CodeAct Agent."""

opendevin/controller/agent_controller.py

@@ -176,11 +176,11 @@ class AgentController:
         elif isinstance(event, ModifyTaskAction):
             self.state.root_task.set_subtask_state(event.task_id, event.state)
         elif isinstance(event, AgentFinishAction):
-            self.state.outputs = event.outputs  # type: ignore[attr-defined]
+            self.state.outputs = event.outputs
             self.state.metrics.merge(self.state.local_metrics)
             await self.set_agent_state_to(AgentState.FINISHED)
         elif isinstance(event, AgentRejectAction):
-            self.state.outputs = event.outputs  # type: ignore[attr-defined]
+            self.state.outputs = event.outputs
             self.state.metrics.merge(self.state.local_metrics)
             await self.set_agent_state_to(AgentState.REJECTED)
         elif isinstance(event, Observation):

opendevin/events/action/agent.py

@@ -1,4 +1,5 @@
 from dataclasses import dataclass, field
+from typing import Any
 
 from opendevin.core.schema import ActionType
 
@@ -35,7 +36,15 @@ class AgentSummarizeAction(Action):
 
 @dataclass
 class AgentFinishAction(Action):
-    outputs: dict = field(default_factory=dict)
+    """An action where the agent finishes the task.
+
+    Attributes:
+        outputs (dict): The outputs of the agent, for instance "content".
+        thought (str): The agent's explanation of its actions.
+        action (str): The action type, namely ActionType.FINISH.
+    """
+
+    outputs: dict[str, Any] = field(default_factory=dict)
     thought: str = ''
     action: str = ActionType.FINISH
 

opendevin/events/observation/delegate.py

@@ -7,7 +7,13 @@ from .observation import Observation
 
 @dataclass
 class AgentDelegateObservation(Observation):
-    """This data class represents the result from delegating to another agent"""
+    """This data class represents the result from delegating to another agent.
+
+    Attributes:
+        content (str): The content of the observation.
+        outputs (dict): The outputs of the delegated agent.
+        observation (str): The type of observation.
+    """
 
     outputs: dict
     observation: str = ObservationType.DELEGATE

tests/unit/test_codeact_agent.py

@@ -0,0 +1,97 @@
+from unittest.mock import Mock
+
+import pytest
+
+from agenthub.codeact_agent.codeact_agent import CodeActAgent
+from opendevin.core.config import LLMConfig
+from opendevin.core.message import TextContent
+from opendevin.events.observation.commands import (
+    CmdOutputObservation,
+    IPythonRunCellObservation,
+)
+from opendevin.events.observation.delegate import AgentDelegateObservation
+from opendevin.events.observation.error import ErrorObservation
+from opendevin.llm.llm import LLM
+
+
+@pytest.fixture
+def agent() -> CodeActAgent:
+    agent = CodeActAgent(llm=LLM(LLMConfig()))
+    agent.llm = Mock()
+    agent.llm.config = Mock()
+    agent.llm.config.max_message_chars = 100
+    return agent
+
+
+def test_cmd_output_observation_message(agent: CodeActAgent):
+    obs = CmdOutputObservation(
+        command='echo hello', content='Command output', command_id=1, exit_code=0
+    )
+
+    result = agent.get_observation_message(obs)
+
+    assert result is not None
+    assert result.role == 'user'
+    assert len(result.content) == 1
+    assert isinstance(result.content[0], TextContent)
+    assert 'OBSERVATION:' in result.content[0].text
+    assert 'Command output' in result.content[0].text
+    assert 'Command 1 finished with exit code 0' in result.content[0].text
+
+
+def test_ipython_run_cell_observation_message(agent: CodeActAgent):
+    obs = IPythonRunCellObservation(
+        code='plt.plot()',
+        content='IPython output\n![image](data:image/png;base64,ABC123)',
+    )
+
+    result = agent.get_observation_message(obs)
+
+    assert result is not None
+    assert result.role == 'user'
+    assert len(result.content) == 1
+    assert isinstance(result.content[0], TextContent)
+    assert 'OBSERVATION:' in result.content[0].text
+    assert 'IPython output' in result.content[0].text
+    assert (
+        '![image](data:image/png;base64, ...) already displayed to user'
+        in result.content[0].text
+    )
+    assert 'ABC123' not in result.content[0].text
+
+
+def test_agent_delegate_observation_message(agent: CodeActAgent):
+    obs = AgentDelegateObservation(
+        content='Content', outputs={'content': 'Delegated agent output'}
+    )
+
+    result = agent.get_observation_message(obs)
+
+    assert result is not None
+    assert result.role == 'user'
+    assert len(result.content) == 1
+    assert isinstance(result.content[0], TextContent)
+    assert 'OBSERVATION:' in result.content[0].text
+    assert 'Delegated agent output' in result.content[0].text
+
+
+def test_error_observation_message(agent: CodeActAgent):
+    obs = ErrorObservation('Error message')
+
+    result = agent.get_observation_message(obs)
+
+    assert result is not None
+    assert result.role == 'user'
+    assert len(result.content) == 1
+    assert isinstance(result.content[0], TextContent)
+    assert 'OBSERVATION:' in result.content[0].text
+    assert 'Error message' in result.content[0].text
+    assert 'Error occurred in processing last action' in result.content[0].text
+
+
+def test_unknown_observation_message(agent: CodeActAgent):
+    obs = Mock()
+
+    with pytest.raises(ValueError) as excinfo:
+        agent.get_observation_message(obs)
+        assert 'Unknown observation type:' in str(excinfo.value)