OpenHands/openhands/controller/agent_controller.py
2024-10-29 07:30:50 +01:00

655 lines
27 KiB
Python

import asyncio
import copy
import traceback
from typing import Type
import litellm
from openhands.controller.agent import Agent
from openhands.controller.state.state import State, TrafficControlState
from openhands.controller.stuck import StuckDetector
from openhands.core.config import AgentConfig, LLMConfig
from openhands.core.exceptions import (
LLMMalformedActionError,
LLMNoActionError,
LLMResponseError,
)
from openhands.core.logger import openhands_logger as logger
from openhands.core.schema import AgentState
from openhands.events import EventSource, EventStream, EventStreamSubscriber
from openhands.events.action import (
Action,
ActionConfirmationStatus,
AddTaskAction,
AgentDelegateAction,
AgentFinishAction,
AgentRejectAction,
ChangeAgentStateAction,
CmdRunAction,
IPythonRunCellAction,
MessageAction,
ModifyTaskAction,
NullAction,
)
from openhands.events.event import Event
from openhands.events.observation import (
AgentDelegateObservation,
AgentStateChangedObservation,
CmdOutputObservation,
ErrorObservation,
FatalErrorObservation,
Observation,
)
from openhands.events.serialization.event import truncate_content
from openhands.llm.llm import LLM
from openhands.runtime.utils.shutdown_listener import should_continue
# note: RESUME is only available on web GUI
TRAFFIC_CONTROL_REMINDER = (
"Please click on resume button if you'd like to continue, or start a new task."
)
class AgentController:
id: str
agent: Agent
max_iterations: int
event_stream: EventStream
state: State
confirmation_mode: bool
agent_to_llm_config: dict[str, LLMConfig]
agent_configs: dict[str, AgentConfig]
agent_task: asyncio.Future | None = None
parent: 'AgentController | None' = None
delegate: 'AgentController | None' = None
_pending_action: Action | None = None
def __init__(
self,
agent: Agent,
event_stream: EventStream,
max_iterations: int,
max_budget_per_task: float | None = None,
agent_to_llm_config: dict[str, LLMConfig] | None = None,
agent_configs: dict[str, AgentConfig] | None = None,
sid: str = 'default',
confirmation_mode: bool = False,
initial_state: State | None = None,
is_delegate: bool = False,
headless_mode: bool = True,
):
"""Initializes a new instance of the AgentController class.
Args:
agent: The agent instance to control.
event_stream: The event stream to publish events to.
max_iterations: The maximum number of iterations the agent can run.
max_budget_per_task: The maximum budget (in USD) allowed per task, beyond which the agent will stop.
agent_to_llm_config: A dictionary mapping agent names to LLM configurations in the case that
we delegate to a different agent.
agent_configs: A dictionary mapping agent names to agent configurations in the case that
we delegate to a different agent.
sid: The session ID of the agent.
initial_state: The initial state of the controller.
is_delegate: Whether this controller is a delegate.
headless_mode: Whether the agent is run in headless mode.
"""
self._step_lock = asyncio.Lock()
self.id = sid
self.agent = agent
self.headless_mode = headless_mode
# subscribe to the event stream
self.event_stream = event_stream
self.event_stream.subscribe(
EventStreamSubscriber.AGENT_CONTROLLER, self.on_event, append=is_delegate
)
# state from the previous session, state from a parent agent, or a fresh state
self.set_initial_state(
state=initial_state,
max_iterations=max_iterations,
confirmation_mode=confirmation_mode,
)
self.max_budget_per_task = max_budget_per_task
self.agent_to_llm_config = agent_to_llm_config if agent_to_llm_config else {}
self.agent_configs = agent_configs if agent_configs else {}
self._initial_max_iterations = max_iterations
self._initial_max_budget_per_task = max_budget_per_task
# stuck helper
self._stuck_detector = StuckDetector(self.state)
async def close(self):
"""Closes the agent controller, canceling any ongoing tasks and unsubscribing from the event stream."""
await self.set_agent_state_to(AgentState.STOPPED)
self.event_stream.unsubscribe(EventStreamSubscriber.AGENT_CONTROLLER)
def log(self, level: str, message: str, extra: dict | None = None):
"""Logs a message to the agent controller's logger.
Args:
message (str): The message to log.
"""
message = f'[Agent Controller {self.id}] {message}'
getattr(logger, level)(message, extra=extra)
def update_state_before_step(self):
self.state.iteration += 1
self.state.local_iteration += 1
async def update_state_after_step(self):
# update metrics especially for cost. Use deepcopy to avoid it being modified by agent.reset()
self.state.local_metrics = copy.deepcopy(self.agent.llm.metrics)
async def report_error(self, message: str, exception: Exception | None = None):
"""Reports an error to the user and sends the exception to the LLM next step, in the hope it can self-correct.
This method should be called for a particular type of errors, which have:
- a user-friendly message, which will be shown in the chat box. This should not be a raw exception message.
- an ErrorObservation that can be sent to the LLM by the user role, with the exception message, so it can self-correct next time.
"""
self.state.last_error = message
if exception:
self.state.last_error += f': {exception}'
detail = str(exception) if exception is not None else ''
if exception is not None and isinstance(exception, litellm.AuthenticationError):
detail = 'Please check your credentials. Is your API key correct?'
self.event_stream.add_event(
ErrorObservation(f'{message}:{detail}'), EventSource.USER
)
async def start_step_loop(self):
"""The main loop for the agent's step-by-step execution."""
self.log('info', 'Starting step loop...')
while should_continue():
try:
await self._step()
except asyncio.CancelledError:
self.log('debug', 'AgentController task was cancelled')
break
except Exception as e:
traceback.print_exc()
self.log('error', f'Error while running the agent: {e}')
self.log('error', traceback.format_exc())
await self.report_error(
'There was an unexpected error while running the agent', exception=e
)
await self.set_agent_state_to(AgentState.ERROR)
break
await asyncio.sleep(0.1)
async def on_event(self, event: Event):
"""Callback from the event stream. Notifies the controller of incoming events.
Args:
event (Event): The incoming event to process.
"""
if hasattr(event, 'hidden') and event.hidden:
return
if isinstance(event, Action):
await self._handle_action(event)
elif isinstance(event, Observation):
await self._handle_observation(event)
async def _handle_action(self, action: Action):
"""Handles actions from the event stream.
Args:
action (Action): The action to handle.
"""
if isinstance(action, ChangeAgentStateAction):
await self.set_agent_state_to(action.agent_state) # type: ignore
elif isinstance(action, MessageAction):
await self._handle_message_action(action)
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.task_id, action.state)
elif isinstance(action, AgentFinishAction):
self.state.outputs = action.outputs
self.state.metrics.merge(self.state.local_metrics)
await self.set_agent_state_to(AgentState.FINISHED)
elif isinstance(action, AgentRejectAction):
self.state.outputs = action.outputs
self.state.metrics.merge(self.state.local_metrics)
await self.set_agent_state_to(AgentState.REJECTED)
async def _handle_observation(self, observation: Observation):
"""Handles observation from the event stream.
Args:
observation (observation): The observation to handle.
"""
if (
self._pending_action
and hasattr(self._pending_action, 'confirmation_state')
and self._pending_action.confirmation_state
== ActionConfirmationStatus.AWAITING_CONFIRMATION
):
return
# Make sure we print the observation in the same way as the LLM sees it
observation_to_print = copy.deepcopy(observation)
if len(observation_to_print.content) > self.agent.llm.config.max_message_chars:
observation_to_print.content = truncate_content(
observation_to_print.content, self.agent.llm.config.max_message_chars
)
self.log('debug', str(observation_to_print), extra={'msg_type': 'OBSERVATION'})
# Merge with the metrics from the LLM - it will to synced to the controller's local metrics in update_state_after_step()
if observation.llm_metrics is not None:
self.agent.llm.metrics.merge(observation.llm_metrics)
if self._pending_action and self._pending_action.id == observation.cause:
self._pending_action = None
if self.state.agent_state == AgentState.USER_CONFIRMED:
await self.set_agent_state_to(AgentState.RUNNING)
if self.state.agent_state == AgentState.USER_REJECTED:
await self.set_agent_state_to(AgentState.AWAITING_USER_INPUT)
return
if isinstance(observation, CmdOutputObservation):
return
elif isinstance(observation, AgentDelegateObservation):
self.state.history.on_event(observation)
elif isinstance(observation, ErrorObservation):
if self.state.agent_state == AgentState.ERROR:
self.state.metrics.merge(self.state.local_metrics)
elif isinstance(observation, FatalErrorObservation):
self.state.last_error = (
f'There was a fatal error during agent execution: {str(observation)}'
)
self.state.metrics.merge(self.state.local_metrics)
await self.set_agent_state_to(AgentState.ERROR)
async def _handle_message_action(self, action: MessageAction):
"""Handles message actions from the event stream.
Args:
action (MessageAction): The message action to handle.
"""
if action.source == EventSource.USER:
self.log(
'debug',
str(action),
extra={'msg_type': 'ACTION', 'event_source': EventSource.USER},
)
if self.get_agent_state() != AgentState.RUNNING:
await self.set_agent_state_to(AgentState.RUNNING)
elif action.source == EventSource.AGENT and action.wait_for_response:
await self.set_agent_state_to(AgentState.AWAITING_USER_INPUT)
def reset_task(self):
"""Resets the agent's task."""
self.almost_stuck = 0
self.agent.reset()
async def set_agent_state_to(self, new_state: AgentState):
"""Updates the agent's state and handles side effects. Can emit events to the event stream.
Args:
new_state (AgentState): The new state to set for the agent.
"""
self.log(
'info',
f'Setting agent({self.agent.name}) state from {self.state.agent_state} to {new_state}',
)
if new_state == self.state.agent_state:
return
if new_state == AgentState.STOPPED or new_state == AgentState.ERROR:
self.reset_task()
elif (
new_state == AgentState.RUNNING
and self.state.agent_state == AgentState.PAUSED
and self.state.traffic_control_state == TrafficControlState.THROTTLING
):
# user intends to interrupt traffic control and let the task resume temporarily
self.state.traffic_control_state = TrafficControlState.PAUSED
# User has chosen to deliberately continue - lets double the max iterations
if (
self.state.iteration is not None
and self.state.max_iterations is not None
and self._initial_max_iterations is not None
):
if self.state.iteration >= self.state.max_iterations:
self.state.max_iterations += self._initial_max_iterations
if (
self.state.metrics.accumulated_cost is not None
and self.max_budget_per_task is not None
and self._initial_max_budget_per_task is not None
):
if self.state.metrics.accumulated_cost >= self.max_budget_per_task:
self.max_budget_per_task += self._initial_max_budget_per_task
elif self._pending_action is not None and (
new_state == AgentState.USER_CONFIRMED
or new_state == AgentState.USER_REJECTED
):
if hasattr(self._pending_action, 'thought'):
self._pending_action.thought = '' # type: ignore[union-attr]
if new_state == AgentState.USER_CONFIRMED:
confirmation_state = ActionConfirmationStatus.CONFIRMED
else:
confirmation_state = ActionConfirmationStatus.REJECTED
self._pending_action.confirmation_state = confirmation_state # type: ignore[attr-defined]
self.event_stream.add_event(self._pending_action, EventSource.AGENT)
self.state.agent_state = new_state
self.event_stream.add_event(
AgentStateChangedObservation('', self.state.agent_state), EventSource.AGENT
)
if new_state == AgentState.INIT and self.state.resume_state:
await self.set_agent_state_to(self.state.resume_state)
self.state.resume_state = None
def get_agent_state(self):
"""Returns the current state of the agent.
Returns:
AgentState: The current state of the agent.
"""
return self.state.agent_state
async def start_delegate(self, action: AgentDelegateAction):
"""Start a delegate agent to handle a subtask.
OpenHands is a multi-agentic system. A `task` is a conversation between
OpenHands (the whole system) and the user, which might involve one or more inputs
from the user. It starts with an initial input (typically a task statement) from
the user, and ends with either an `AgentFinishAction` initiated by the agent, a
stop initiated by the user, or an error.
A `subtask` is a conversation between an agent and the user, or another agent. If a `task`
is conducted by a single agent, then it's also a `subtask`. Otherwise, a `task` consists of
multiple `subtasks`, each executed by one agent.
Args:
action (AgentDelegateAction): The action containing information about the delegate agent to start.
"""
agent_cls: Type[Agent] = Agent.get_cls(action.agent)
agent_config = self.agent_configs.get(action.agent, self.agent.config)
llm_config = self.agent_to_llm_config.get(action.agent, self.agent.llm.config)
llm = LLM(config=llm_config)
delegate_agent = agent_cls(llm=llm, config=agent_config)
state = State(
inputs=action.inputs or {},
local_iteration=0,
iteration=self.state.iteration,
max_iterations=self.state.max_iterations,
delegate_level=self.state.delegate_level + 1,
# global metrics should be shared between parent and child
metrics=self.state.metrics,
)
self.log(
'debug',
f'start delegate, creating agent {delegate_agent.name} using LLM {llm}',
)
self.delegate = AgentController(
sid=self.id + '-delegate',
agent=delegate_agent,
event_stream=self.event_stream,
max_iterations=self.state.max_iterations,
max_budget_per_task=self.max_budget_per_task,
agent_to_llm_config=self.agent_to_llm_config,
agent_configs=self.agent_configs,
initial_state=state,
is_delegate=True,
headless_mode=self.headless_mode,
)
await self.delegate.set_agent_state_to(AgentState.RUNNING)
async def _step(self) -> None:
"""Executes a single step of the parent or delegate agent. Detects stuck agents and limits on the number of iterations and the task budget."""
if self.get_agent_state() != AgentState.RUNNING:
await asyncio.sleep(1)
return
if self._pending_action:
await asyncio.sleep(1)
return
# check if agent got stuck before taking any action
if self._is_stuck():
# This need to go BEFORE report_error to sync metrics
self.event_stream.add_event(
FatalErrorObservation('Agent got stuck in a loop'), EventSource.USER
)
return
if self.delegate is not None:
assert self.delegate != self
if self.delegate.get_agent_state() == AgentState.PAUSED:
await asyncio.sleep(1)
else:
await self._delegate_step()
return
self.log(
'info',
f'LEVEL {self.state.delegate_level} LOCAL STEP {self.state.local_iteration} GLOBAL STEP {self.state.iteration}',
extra={'msg_type': 'STEP'},
)
# check if agent hit the resources limit
stop_step = False
if self.state.iteration >= self.state.max_iterations:
stop_step = await self._handle_traffic_control(
'iteration', self.state.iteration, self.state.max_iterations
)
if self.max_budget_per_task is not None:
current_cost = self.state.metrics.accumulated_cost
if current_cost > self.max_budget_per_task:
stop_step = await self._handle_traffic_control(
'budget', current_cost, self.max_budget_per_task
)
if stop_step:
return
self.update_state_before_step()
action: Action = NullAction()
try:
action = self.agent.step(self.state)
if action is None:
raise LLMNoActionError('No action was returned')
except (LLMMalformedActionError, LLMNoActionError, LLMResponseError) as e:
# report to the user
# and send the underlying exception to the LLM for self-correction
await self.report_error(str(e))
return
# FIXME: more graceful handling of litellm.exceptions.ContextWindowExceededError
# e.g. try to condense the memory and try again
except litellm.exceptions.ContextWindowExceededError as e:
self.state.last_error = str(e)
await self.set_agent_state_to(AgentState.ERROR)
return
if action.runnable:
if self.state.confirmation_mode and (
type(action) is CmdRunAction or type(action) is IPythonRunCellAction
):
action.confirmation_state = (
ActionConfirmationStatus.AWAITING_CONFIRMATION
)
self._pending_action = action
if not isinstance(action, NullAction):
if (
hasattr(action, 'confirmation_state')
and action.confirmation_state
== ActionConfirmationStatus.AWAITING_CONFIRMATION
):
await self.set_agent_state_to(AgentState.AWAITING_USER_CONFIRMATION)
self.event_stream.add_event(action, EventSource.AGENT)
await self.update_state_after_step()
self.log('debug', str(action), extra={'msg_type': 'ACTION'})
async def _delegate_step(self):
"""Executes a single step of the delegate agent."""
self.log('debug', 'Delegate not none, awaiting...')
await self.delegate._step() # type: ignore[union-attr]
self.log('debug', 'Delegate step done')
assert self.delegate is not None
delegate_state = self.delegate.get_agent_state()
self.log('debug', f'Delegate state: {delegate_state}')
if delegate_state == AgentState.ERROR:
# update iteration that shall be shared across agents
self.state.iteration = self.delegate.state.iteration
# close the delegate upon error
await self.delegate.close()
self.delegate = None
self.delegateAction = None
await self.report_error('Delegator agent encountered an error')
elif delegate_state in (AgentState.FINISHED, AgentState.REJECTED):
self.log('debug', 'Delegate agent has finished execution')
# retrieve delegate result
outputs = self.delegate.state.outputs if self.delegate.state else {}
# update iteration that shall be shared across agents
self.state.iteration = self.delegate.state.iteration
# close delegate controller: we must close the delegate controller before adding new events
await self.delegate.close()
# update delegate result observation
# TODO: replace this with AI-generated summary (#2395)
formatted_output = ', '.join(
f'{key}: {value}' for key, value in outputs.items()
)
content = (
f'{self.delegate.agent.name} finishes task with {formatted_output}'
)
obs: Observation = AgentDelegateObservation(
outputs=outputs, content=content
)
# clean up delegate status
self.delegate = None
self.delegateAction = None
self.event_stream.add_event(obs, EventSource.AGENT)
return
async def _handle_traffic_control(
self, limit_type: str, current_value: float, max_value: float
):
"""Handles agent state after hitting the traffic control limit.
Args:
limit_type (str): The type of limit that was hit.
current_value (float): The current value of the limit.
max_value (float): The maximum value of the limit.
"""
stop_step = False
if self.state.traffic_control_state == TrafficControlState.PAUSED:
self.log(
'debug', 'Hitting traffic control, temporarily resume upon user request'
)
self.state.traffic_control_state = TrafficControlState.NORMAL
else:
self.state.traffic_control_state = TrafficControlState.THROTTLING
if self.headless_mode:
# This need to go BEFORE report_error to sync metrics
await self.set_agent_state_to(AgentState.ERROR)
# set to ERROR state if running in headless mode
# since user cannot resume on the web interface
await self.report_error(
f'Agent reached maximum {limit_type} in headless mode, task stopped. '
f'Current {limit_type}: {current_value:.2f}, max {limit_type}: {max_value:.2f}'
)
else:
await self.set_agent_state_to(AgentState.PAUSED)
await self.report_error(
f'Agent reached maximum {limit_type}, task paused. '
f'Current {limit_type}: {current_value:.2f}, max {limit_type}: {max_value:.2f}. '
f'{TRAFFIC_CONTROL_REMINDER}'
)
stop_step = True
return stop_step
def get_state(self):
"""Returns the current running state object.
Returns:
State: The current state object.
"""
return self.state
def set_initial_state(
self,
state: State | None,
max_iterations: int,
confirmation_mode: bool = False,
):
"""Sets the initial state for the agent, either from the previous session, or from a parent agent, or by creating a new one.
Args:
state: The state to initialize with, or None to create a new state.
max_iterations: The maximum number of iterations allowed for the task.
confirmation_mode: Whether to enable confirmation mode.
"""
# state from the previous session, state from a parent agent, or a new state
# note that this is called twice when restoring a previous session, first with state=None
if state is None:
self.state = State(
inputs={},
max_iterations=max_iterations,
confirmation_mode=confirmation_mode,
)
else:
self.state = state
# when restored from a previous session, the State object will have history, start_id, and end_id
# connect it to the event stream
self.state.history.set_event_stream(self.event_stream)
# if start_id was not set in State, we're starting fresh, at the top of the stream
start_id = self.state.start_id
if start_id == -1:
start_id = self.event_stream.get_latest_event_id() + 1
else:
self.log(
'debug', f'AgentController {self.id} restoring from event {start_id}'
)
# make sure history is in sync
self.state.start_id = start_id
self.state.history.start_id = start_id
# if there was an end_id saved in State, set it in history
# currently not used, later useful for delegates
if self.state.end_id > -1:
self.state.history.end_id = self.state.end_id
def _is_stuck(self):
"""Checks if the agent or its delegate is stuck in a loop.
Returns:
bool: True if the agent is stuck, False otherwise.
"""
# check if delegate stuck
if self.delegate and self.delegate._is_stuck():
return True
return self._stuck_detector.is_stuck()
def __repr__(self):
return (
f'AgentController(id={self.id}, agent={self.agent!r}, '
f'event_stream={self.event_stream!r}, '
f'state={self.state!r}, agent_task={self.agent_task!r}, '
f'delegate={self.delegate!r}, _pending_action={self._pending_action!r})'
)