mirror of
https://github.com/OpenHands/OpenHands.git
synced 2026-03-22 13:47:19 +08:00
292 lines
11 KiB
Python
292 lines
11 KiB
Python
from litellm.exceptions import ContextWindowExceededError
|
||
|
||
from agenthub.codeact_agent.action_parser import CodeActResponseParser
|
||
from agenthub.codeact_agent.prompt import (
|
||
COMMAND_DOCS,
|
||
EXAMPLES,
|
||
GITHUB_MESSAGE,
|
||
SYSTEM_PREFIX,
|
||
SYSTEM_SUFFIX,
|
||
)
|
||
from opendevin.controller.agent import Agent
|
||
from opendevin.controller.state.state import State
|
||
from opendevin.core.exceptions import (
|
||
ContextWindowLimitExceededError,
|
||
TokenLimitExceedError,
|
||
)
|
||
from opendevin.core.logger import opendevin_logger as logger
|
||
from opendevin.events.action import (
|
||
Action,
|
||
AgentFinishAction,
|
||
AgentSummarizeAction,
|
||
BrowseInteractiveAction,
|
||
CmdRunAction,
|
||
IPythonRunCellAction,
|
||
MessageAction,
|
||
)
|
||
from opendevin.events.observation import (
|
||
AgentDelegateObservation,
|
||
BrowserOutputObservation,
|
||
CmdOutputObservation,
|
||
IPythonRunCellObservation,
|
||
)
|
||
from opendevin.events.serialization.event import truncate_content
|
||
from opendevin.llm.llm import LLM
|
||
from opendevin.memory.condenser import MemoryCondenser
|
||
from opendevin.runtime.plugins import (
|
||
AgentSkillsRequirement,
|
||
JupyterRequirement,
|
||
PluginRequirement,
|
||
)
|
||
from opendevin.runtime.tools import RuntimeTool
|
||
|
||
ENABLE_GITHUB = True
|
||
|
||
|
||
def action_to_str(action: Action) -> str:
|
||
if isinstance(action, CmdRunAction):
|
||
return f'{action.thought}\n<execute_bash>\n{action.command}\n</execute_bash>'
|
||
elif isinstance(action, IPythonRunCellAction):
|
||
return f'{action.thought}\n<execute_ipython>\n{action.code}\n</execute_ipython>'
|
||
elif isinstance(action, BrowseInteractiveAction):
|
||
return f'{action.thought}\n<execute_browse>\n{action.browser_actions}\n</execute_browse>'
|
||
elif isinstance(action, MessageAction):
|
||
return action.content
|
||
elif isinstance(action, AgentSummarizeAction):
|
||
return action.summary
|
||
return ''
|
||
|
||
|
||
def get_action_message(action: Action) -> dict[str, str] | None:
|
||
if (
|
||
isinstance(action, BrowseInteractiveAction)
|
||
or isinstance(action, CmdRunAction)
|
||
or isinstance(action, IPythonRunCellAction)
|
||
or isinstance(action, MessageAction)
|
||
or isinstance(action, AgentSummarizeAction)
|
||
):
|
||
return {
|
||
'role': 'user' if action.source == 'user' else 'assistant',
|
||
'content': action_to_str(action),
|
||
}
|
||
return None
|
||
|
||
|
||
def get_observation_message(obs) -> dict[str, str] | None:
|
||
if isinstance(obs, CmdOutputObservation):
|
||
content = 'OBSERVATION:\n' + truncate_content(obs.content)
|
||
content += (
|
||
f'\n[Command {obs.command_id} finished with exit code {obs.exit_code}]'
|
||
)
|
||
return {'role': 'user', 'content': content}
|
||
elif isinstance(obs, IPythonRunCellObservation):
|
||
content = 'OBSERVATION:\n' + obs.content
|
||
# replace base64 images with a placeholder
|
||
splitted = content.split('\n')
|
||
for i, line in enumerate(splitted):
|
||
if ' already displayed to user'
|
||
)
|
||
content = '\n'.join(splitted)
|
||
content = truncate_content(content)
|
||
return {'role': 'user', 'content': content}
|
||
elif isinstance(obs, BrowserOutputObservation):
|
||
content = 'OBSERVATION:\n' + truncate_content(obs.content)
|
||
return {'role': 'user', 'content': content}
|
||
elif isinstance(obs, AgentDelegateObservation):
|
||
content = 'OBSERVATION:\n' + truncate_content(str(obs.outputs))
|
||
return {'role': 'user', 'content': content}
|
||
return None
|
||
|
||
|
||
# FIXME: We can tweak these two settings to create MicroAgents specialized toward different area
|
||
def get_system_message() -> str:
|
||
if ENABLE_GITHUB:
|
||
return f'{SYSTEM_PREFIX}\n{GITHUB_MESSAGE}\n\n{COMMAND_DOCS}\n\n{SYSTEM_SUFFIX}'
|
||
else:
|
||
return f'{SYSTEM_PREFIX}\n\n{COMMAND_DOCS}\n\n{SYSTEM_SUFFIX}'
|
||
|
||
|
||
def get_in_context_example() -> str:
|
||
return EXAMPLES
|
||
|
||
|
||
class CodeActAgent(Agent):
|
||
VERSION = '1.7'
|
||
"""
|
||
The Code Act Agent is a minimalist agent.
|
||
The agent works by passing the model a list of action-observation pairs and prompting the model to take the next step.
|
||
|
||
### Overview
|
||
|
||
This agent implements the CodeAct idea ([paper](https://arxiv.org/abs/2402.13463), [tweet](https://twitter.com/xingyaow_/status/1754556835703751087)) that consolidates LLM agents’ **act**ions into a unified **code** action space for both *simplicity* and *performance* (see paper for more details).
|
||
|
||
The conceptual idea is illustrated below. At each turn, the agent can:
|
||
|
||
1. **Converse**: Communicate with humans in natural language to ask for clarification, confirmation, etc.
|
||
2. **CodeAct**: Choose to perform the task by executing code
|
||
- Execute any valid Linux `bash` command
|
||
- Execute any valid `Python` code with [an interactive Python interpreter](https://ipython.org/). This is simulated through `bash` command, see plugin system below for more details.
|
||
|
||
|
||
|
||
### Plugin System
|
||
|
||
To make the CodeAct agent more powerful with only access to `bash` action space, CodeAct agent leverages OpenDevin's plugin system:
|
||
- [Jupyter plugin](https://github.com/OpenDevin/OpenDevin/tree/main/opendevin/runtime/plugins/jupyter): for IPython execution via bash command
|
||
- [SWE-agent tool plugin](https://github.com/OpenDevin/OpenDevin/tree/main/opendevin/runtime/plugins/swe_agent_commands): Powerful bash command line tools for software development tasks introduced by [swe-agent](https://github.com/princeton-nlp/swe-agent).
|
||
### Demo
|
||
|
||
https://github.com/OpenDevin/OpenDevin/assets/38853559/f592a192-e86c-4f48-ad31-d69282d5f6ac
|
||
|
||
*Example of CodeActAgent with `gpt-4-turbo-2024-04-09` performing a data science task (linear regression)*
|
||
|
||
### Work-in-progress & Next step
|
||
|
||
[] Support web-browsing
|
||
[] Complete the workflow for CodeAct agent to submit Github PRs
|
||
|
||
"""
|
||
|
||
sandbox_plugins: list[PluginRequirement] = [
|
||
# NOTE: AgentSkillsRequirement need to go before JupyterRequirement, since
|
||
# AgentSkillsRequirement provides a lot of Python functions
|
||
# and it need to be initialized before Jupyter for Jupyter to use those functions.
|
||
AgentSkillsRequirement(),
|
||
JupyterRequirement(),
|
||
]
|
||
runtime_tools: list[RuntimeTool] = [RuntimeTool.BROWSER]
|
||
|
||
system_message: str = get_system_message()
|
||
in_context_example: str = f"Here is an example of how you can interact with the environment for task solving:\n{get_in_context_example()}\n\nNOW, LET'S START!"
|
||
|
||
attempts_to_condense: int = 10
|
||
|
||
action_parser = CodeActResponseParser()
|
||
|
||
def __init__(
|
||
self,
|
||
llm: LLM,
|
||
attempts_to_condense: int = 2,
|
||
) -> None:
|
||
"""Initializes a new instance of the CodeActAgent class.
|
||
|
||
Parameters:
|
||
- llm (LLM): The llm to be used by this agent
|
||
"""
|
||
super().__init__(llm)
|
||
self.memory_condenser = MemoryCondenser(llm)
|
||
self.attempts_to_condense = attempts_to_condense
|
||
self.reset()
|
||
|
||
def reset(self) -> None:
|
||
"""
|
||
Resets the CodeAct Agent.
|
||
"""
|
||
super().reset()
|
||
|
||
def step(self, state: State) -> Action:
|
||
"""
|
||
Run the agent for one step.
|
||
|
||
Parameters:
|
||
- state: The current state of the environment.
|
||
|
||
Returns:
|
||
- CmdRunAction(command) - bash command to run
|
||
- IPythonRunCellAction(code) - IPython code to run
|
||
- AgentDelegateAction(agent, inputs) - delegate action for (sub)task
|
||
- MessageAction(content) - Message action to run (e.g. ask for clarification)
|
||
- AgentFinishAction() - end the interaction
|
||
"""
|
||
logger.info(f'Running CodeActAgent v{self.VERSION}')
|
||
|
||
# if we're done, go back
|
||
latest_user_message = state.history.get_latest_user_message()
|
||
if latest_user_message and latest_user_message.strip() == '/exit':
|
||
return AgentFinishAction()
|
||
|
||
# prepare what we want to send to the LLM
|
||
messages: list[dict[str, str]] = self._get_messages(state)
|
||
|
||
response = None
|
||
# give it multiple chances to get a response
|
||
# if it fails, we'll try to condense memory
|
||
attempt = 0
|
||
while not response and attempt < self.attempts_to_condense:
|
||
try:
|
||
response = self.llm.completion(
|
||
messages=messages,
|
||
stop=[
|
||
'</execute_ipython>',
|
||
'</execute_bash>',
|
||
'</execute_browse>',
|
||
],
|
||
temperature=0.0,
|
||
)
|
||
except (ContextWindowExceededError, TokenLimitExceedError):
|
||
logger.warning(
|
||
'Context window exceeded or token limit exceeded. Condensing memory, attempt %d',
|
||
attempt,
|
||
)
|
||
|
||
# Retry processing events with condensed memory
|
||
summary_action = self.memory_condenser.condense(state.history)
|
||
attempt += 1
|
||
if summary_action:
|
||
# update the messages, so we get the new summary
|
||
messages = self._get_messages(state)
|
||
continue
|
||
|
||
if not response:
|
||
raise ContextWindowLimitExceededError(
|
||
'Context window limit exceeded. Unable to condense memory.'
|
||
)
|
||
|
||
state.num_of_chars += sum(
|
||
len(message['content']) for message in messages
|
||
) + len(response.choices[0].message.content)
|
||
return self.action_parser.parse(response)
|
||
|
||
def search_memory(self, query: str) -> list[str]:
|
||
raise NotImplementedError('Implement this abstract method')
|
||
|
||
def _get_messages(self, state: State) -> list[dict[str, str]]:
|
||
messages = [
|
||
{'role': 'system', 'content': self.system_message},
|
||
{
|
||
'role': 'user',
|
||
'content': self.in_context_example,
|
||
},
|
||
]
|
||
|
||
for event in state.history.get_events():
|
||
message = (
|
||
get_action_message(event)
|
||
if isinstance(event, Action)
|
||
else get_observation_message(event)
|
||
)
|
||
if message:
|
||
messages.append(message)
|
||
|
||
latest_user_message = next(
|
||
(m for m in reversed(messages) if m['role'] == 'user'), None
|
||
)
|
||
|
||
if latest_user_message:
|
||
if state.almost_stuck == 1:
|
||
latest_user_message['content'] += (
|
||
'\n\nENVIRONMENT REMINDER: You are almost stuck repeating the same action. You have only 1 iteration left and you must change your approach. Now.'
|
||
)
|
||
elif state.almost_stuck == 2:
|
||
latest_user_message['content'] += (
|
||
'\n\nENVIRONMENT REMINDER: You are almost stuck repeating the same action. You have only 2 iterations left and you must change your approach.'
|
||
)
|
||
else:
|
||
latest_user_message['content'] += (
|
||
f'\n\nENVIRONMENT REMINDER: You have {state.max_iterations - state.iteration} turns left to complete the task.'
|
||
)
|
||
|
||
return messages
|