OpenHands/opendevin/controller/agent_controller.py
Robert Brennan 82a798990c
refactor remind_iterations (#1760)
* refactor remind_iterations

* regenerate tests

* concatenate iteration message

* fix merge issues

* update integration tests
2024-05-14 08:27:12 -04:00

316 lines
12 KiB
Python

import asyncio
from typing import Optional, Type
from agenthub.codeact_agent.codeact_agent import CodeActAgent
from opendevin.controller.agent import Agent
from opendevin.controller.state.state import State
from opendevin.core.config import config
from opendevin.core.exceptions import (
AgentMalformedActionError,
AgentNoActionError,
LLMOutputError,
MaxCharsExceedError,
)
from opendevin.core.logger import opendevin_logger as logger
from opendevin.core.schema import AgentState
from opendevin.events.action import (
Action,
AddTaskAction,
AgentDelegateAction,
AgentFinishAction,
ChangeAgentStateAction,
MessageAction,
ModifyTaskAction,
NullAction,
)
from opendevin.events.event import Event
from opendevin.events.observation import (
AgentDelegateObservation,
AgentStateChangedObservation,
ErrorObservation,
NullObservation,
Observation,
)
from opendevin.events.stream import EventSource, EventStream, EventStreamSubscriber
from opendevin.runtime import DockerSSHBox, Sandbox
from opendevin.runtime.runtime import Runtime
from opendevin.runtime.server.runtime import ServerRuntime
MAX_ITERATIONS = config.max_iterations
MAX_CHARS = config.llm.max_chars
class AgentController:
id: str
agent: Agent
max_iterations: int
runtime: Runtime
event_stream: EventStream
state: State
agent_task: Optional[asyncio.Task] = None
delegate: 'AgentController | None' = None
_agent_state: AgentState = AgentState.LOADING
_cur_step: int = 0
def __init__(
self,
agent: Agent,
event_stream: EventStream,
sid: str = 'default',
max_iterations: int = MAX_ITERATIONS,
max_chars: int = MAX_CHARS,
inputs: dict | None = None,
sandbox: Optional[Sandbox] = None,
):
"""Initializes a new instance of the AgentController class.
Args:
agent: The agent instance to control.
event_stream: The event stream to publish events to.
sid: The session ID of the agent.
max_iterations: The maximum number of iterations the agent can run.
max_chars: The maximum number of characters the agent can output.
inputs: The initial inputs to the agent.
sandbox: An optional initialized sandbox to run the agent in. If not provided, a default sandbox will be created based on config.
"""
self.id = sid
self.agent = agent
self.state = State(inputs=inputs or {})
self.event_stream = event_stream
self.event_stream.subscribe(
EventStreamSubscriber.AGENT_CONTROLLER, self.on_event
)
self.max_iterations = max_iterations
self.runtime = ServerRuntime(sandbox=sandbox, sid=self.id)
self.max_chars = max_chars
# Initialize agent-required plugins for sandbox (if any)
self.runtime.init_sandbox_plugins(agent.sandbox_plugins)
if isinstance(agent, CodeActAgent) and not isinstance(
self.runtime.sandbox, DockerSSHBox
):
logger.warning(
'CodeActAgent requires DockerSSHBox as sandbox! Using other sandbox that are not stateful (LocalBox, DockerExecBox) will not work properly.'
)
async def close(self):
if self.agent_task is not None:
self.agent_task.cancel()
self.event_stream.unsubscribe(EventStreamSubscriber.AGENT_CONTROLLER)
self.runtime.sandbox.close()
self.runtime.browser.close()
await self.set_agent_state_to(AgentState.STOPPED)
def update_state_for_step(self, i):
self.state.iteration = i
self.state.background_commands_obs = self.runtime.get_background_obs()
def update_state_after_step(self):
self.state.updated_info = []
async def add_error_to_history(self, message: str):
await self.add_history(NullAction(), ErrorObservation(message))
async def add_history(
self, action: Action, observation: Observation, add_to_stream=True
):
if not isinstance(action, Action):
raise TypeError(
f'action must be an instance of Action, got {type(action).__name__} instead'
)
if not isinstance(observation, Observation):
raise TypeError(
f'observation must be an instance of Observation, got {type(observation).__name__} instead'
)
self.state.history.append((action, observation))
self.state.updated_info.append((action, observation))
if add_to_stream:
await self.event_stream.add_event(action, EventSource.AGENT)
await self.event_stream.add_event(observation, EventSource.AGENT)
async def _run(self):
if self._agent_state != AgentState.RUNNING:
raise ValueError('Task is not in running state')
for i in range(self._cur_step, self.max_iterations):
self._cur_step = i
try:
finished = await self.step(i)
if finished:
await self.set_agent_state_to(AgentState.FINISHED)
break
except Exception:
logger.error('Error in loop', exc_info=True)
await self.set_agent_state_to(AgentState.ERROR)
await self.add_error_to_history(
'Oops! Something went wrong while completing your task. You can check the logs for more info.'
)
break
if self._is_stuck():
logger.info('Loop detected, stopping task')
await self.set_agent_state_to(AgentState.ERROR)
await self.add_error_to_history(
'I got stuck into a loop, the task has stopped.'
)
break
await asyncio.sleep(
0.001
) # Give back control for a tick, so other async stuff can run
final_state = self.get_agent_state()
if final_state == AgentState.RUNNING:
await self.set_agent_state_to(AgentState.PAUSED)
async def on_event(self, event: Event):
if isinstance(event, ChangeAgentStateAction):
await self.set_agent_state_to(event.agent_state) # type: ignore
elif isinstance(event, MessageAction) and event.source == EventSource.USER:
await self.add_history(event, NullObservation(''), add_to_stream=False)
if self.get_agent_state() != AgentState.RUNNING:
await self.set_agent_state_to(AgentState.RUNNING)
async def reset_task(self):
if self.agent_task is not None:
self.agent_task.cancel()
self.state = State()
self._cur_step = 0
self.agent.reset()
async def set_agent_state_to(self, new_state: AgentState):
logger.info(
f'Setting agent({type(self.agent).__name__}) state from {self._agent_state} to {new_state}'
)
if new_state == self._agent_state:
return
self._agent_state = new_state
if new_state == AgentState.RUNNING:
self.agent_task = asyncio.create_task(self._run())
elif (
new_state == AgentState.PAUSED
or new_state == AgentState.AWAITING_USER_INPUT
):
self._cur_step += 1
if self.agent_task is not None:
self.agent_task.cancel()
elif new_state == AgentState.STOPPED or new_state == AgentState.ERROR:
await self.reset_task()
await self.event_stream.add_event(
AgentStateChangedObservation('', self._agent_state), EventSource.AGENT
)
def get_agent_state(self):
"""Returns the current state of the agent task."""
return self._agent_state
async def start_delegate(self, action: AgentDelegateAction):
AgentCls: Type[Agent] = Agent.get_cls(action.agent)
agent = AgentCls(llm=self.agent.llm)
self.delegate = AgentController(
sid=self.id + '-delegate',
agent=agent,
event_stream=self.event_stream,
max_iterations=self.max_iterations,
max_chars=self.max_chars,
inputs=action.inputs,
)
async def step(self, i: int) -> bool:
if self.delegate is not None:
delegate_done = await self.delegate.step(i)
if delegate_done:
outputs = self.delegate.state.outputs if self.delegate.state else {}
obs: Observation = AgentDelegateObservation(content='', outputs=outputs)
await self.add_history(NullAction(), obs)
self.delegate = None
self.delegateAction = None
return False
logger.info(f'STEP {i}', extra={'msg_type': 'STEP'})
if self.state.num_of_chars > self.max_chars:
raise MaxCharsExceedError(self.state.num_of_chars, self.max_chars)
log_obs = self.runtime.get_background_obs()
for obs in log_obs:
await self.add_history(NullAction(), obs)
logger.info(obs, extra={'msg_type': 'BACKGROUND LOG'})
self.update_state_for_step(i)
action: Action = NullAction()
observation: Observation = NullObservation('')
try:
action = self.agent.step(self.state)
if action is None:
raise AgentNoActionError('No action was returned')
except (AgentMalformedActionError, AgentNoActionError, LLMOutputError) as e:
observation = ErrorObservation(str(e))
logger.info(action, extra={'msg_type': 'ACTION'})
self.update_state_after_step()
if isinstance(action, AgentFinishAction):
self.state.outputs = action.outputs # type: ignore[attr-defined]
logger.info(action, extra={'msg_type': 'INFO'})
await self.add_history(action, NullObservation(''))
return True
elif isinstance(action, MessageAction) and action.wait_for_response:
# FIXME: remove this once history is managed outside the agent controller
await self.add_history(action, NullObservation(''))
await self.set_agent_state_to(AgentState.AWAITING_USER_INPUT)
return False
elif isinstance(action, AgentDelegateAction):
await self.start_delegate(action)
elif isinstance(action, AddTaskAction):
self.state.root_task.add_subtask(
action.parent, action.goal, action.subtasks
)
elif isinstance(action, ModifyTaskAction):
self.state.root_task.set_subtask_state(action.id, action.state)
elif not isinstance(observation, ErrorObservation):
observation = await self.runtime.run_action(action)
if not isinstance(observation, NullObservation):
logger.info(observation, extra={'msg_type': 'OBSERVATION'})
await self.add_history(action, observation)
return False
def get_state(self):
return self.state
def _is_stuck(self):
# check if delegate stuck
if self.delegate and self.delegate._is_stuck():
return True
if len(self.state.history) < 3:
return False
# if the last three (Action, Observation) tuples are too repetitive
# the agent got stuck in a loop
if all(
[
self.state.history[-i][0] == self.state.history[-3][0]
for i in range(1, 3)
]
):
# it repeats same action, give it a chance, but not if:
if all(
isinstance(self.state.history[-i][1], NullObservation)
for i in range(1, 4)
):
# same (Action, NullObservation): like 'think' the same thought over and over
logger.debug('Action, NullObservation loop detected')
return True
elif all(
isinstance(self.state.history[-i][1], ErrorObservation)
for i in range(1, 4)
):
# (NullAction, ErrorObservation): errors coming from an exception
# (Action, ErrorObservation): the same action getting an error, even if not necessarily the same error
logger.debug('Action, ErrorObservation loop detected')
return True
return False