Spaces:
Running
Running
File size: 16,524 Bytes
0646b18 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 |
import asyncio
import datetime
import re
from cuga.backend.activity_tracker.tracker import ActivityTracker, Step
import os
from typing import Any, List, Optional, Literal, AsyncGenerator, Union
from langchain_core.messages import AIMessage
from pydantic import BaseModel
from cuga.backend.browser_env.browser.gym_obs.http_stream_comm import ChromeExtensionCommunicatorProtocol
from cuga.backend.browser_env.tools.providers import BrowserToolImplProvider
from cuga.backend.cuga_graph.graph import DynamicAgentGraph
from cuga.backend.cuga_graph.nodes.browser.action_agent.tools.tools import setup_tools
from cuga.backend.cuga_graph.utils.event_porcessors.action_agent_event_processor import (
ActionAgentEventProcessor,
)
from cuga.backend.cuga_graph.state.agent_state import AgentState, default_state
from cuga.backend.browser_env.browser.gym_env_async import BrowserEnvGymAsync
from cuga.backend.browser_env.browser.open_ended_async import OpenEndedTaskAsync
from cuga.backend.cuga_graph.utils.agent_loop import AgentLoop, AgentLoopAnswer, StreamEvent
from cuga.config import get_app_name_from_url, settings, PACKAGE_ROOT
from loguru import logger
from langchain_core.messages import ToolCall
try:
from langfuse.langchain import CallbackHandler as LangfuseCallbackHandler
except ImportError:
try:
from langfuse.callback.langchain import LangchainCallbackHandler as LangfuseCallbackHandler
except ImportError:
logger.warning("Langfuse is not installed, LangfuseCallbackHandler will be None")
LangfuseCallbackHandler = None
tracker = ActivityTracker()
class ExperimentResult(BaseModel):
score: float
messages: List[AIMessage]
answer: Optional[str] = ""
number_of_actions: int = 0
steps: Optional[List[Step]] = []
class AgentRunner:
def __init__(self, browser_enabled=True, thread_id: str = "1"):
self.browser_enabled = browser_enabled
self.thread_id = thread_id
self.env = None
self.obs = None
self.info = None
self.agent_loop_obj = None
pass
@staticmethod
async def process_event_async(
tool_calls: Optional[List[ToolCall]],
elements,
page,
tool_provider: BrowserToolImplProvider,
session_id=None,
page_data=None,
communicator: ChromeExtensionCommunicatorProtocol | None = None,
) -> List[Any]:
feedback = []
if tool_calls:
k = ActionAgentEventProcessor(page, tool_handlers=setup_tools())
await k.process_action_agent_async(
page,
elements,
tool_calls,
tool_provider,
session_id=session_id,
page_data=page_data,
communicator=communicator,
)
feedback = feedback + k.feedback_log
# print("Retuned feedback before executor", feedback)
return feedback
async def initialize_webarena_env(self, task_id):
from evaluation.tasks.task import GenericWebArenaTask
self.env = BrowserEnvGymAsync(
GenericWebArenaTask,
headless=settings.eval_config.headless,
resizeable_window=True,
enable_playwright_tracing=True,
feedback=[],
task_kwargs={"task_id": task_id},
enable_nocodeui_pu=True,
pw_extra_args=[
*settings.get("PLAYWRIGHT_ARGS", []),
f"--disable-extensions-except={os.path.join(PACKAGE_ROOT, './cuga/backend/browser_env/browser/nocodeui_obs/prod')}",
f"--load-extension={os.path.join(PACKAGE_ROOT, './cuga/backend/browser_env/browser/nocodeui_obs/prod')}",
],
)
self.obs, self.info = await self.env.reset()
async def setup_page_info(self, state: AgentState, env):
"""Setup page URL, app name, and description from environment."""
# Get URL and title
state.url = env.page.url
title = await env.page.title()
url_app_name = get_app_name_from_url(state.url)
# Sanitize title
sanitized_title = re.sub(r'[^\w\s-]', '', title) if title else ""
sanitized_title = re.sub(r'[-\s]+', '_', sanitized_title).strip('_').lower()
# Create app name: url + sanitized title
state.current_app = (
f"{url_app_name}_{sanitized_title}" if sanitized_title else url_app_name or "unknown_app"
)
# Create description
state.current_app_description = f"web application for '{title}' and url '{url_app_name}'"
async def initialize_appworld_env(self):
self.env = BrowserEnvGymAsync(
OpenEndedTaskAsync,
headless=settings.eval_config.headless,
resizeable_window=True,
task_kwargs={"start_url": "https://google.com"},
interface_mode="none",
enable_playwright_tracing=True,
feedback=[],
enable_nocodeui_pu=False,
)
self.obs, self.info = await self.env.reset()
async def initialize_freemode_env(
self, start_url, interface_mode: Literal['browser_only', 'chat_only', 'both'] = 'both'
):
self.env = BrowserEnvGymAsync(
OpenEndedTaskAsync,
headless=False,
interface_mode=interface_mode,
feedback=[],
timeout=15000,
resizeable_window=True,
task_kwargs={"start_url": start_url},
tags_to_mark='all',
enable_nocodeui_pu=False,
)
self.obs, self.info = await self.env.reset()
async def browser_update_state(self, state: AgentState):
if not self.browser_enabled:
return
await self.setup_page_info(state, self.env)
state.url = self.env.get_url()
state.current_app = get_app_name_from_url(state.url)
pu_answer = await self.env.pu_processor.transform(transformer_params={"filter_visible_only": False})
state.elements_as_string = pu_answer.string_representation
state.focused_element_bid = pu_answer.focused_element_bid
tracker.collect_image(pu_answer.img)
state.read_page = pu_answer.page_content
def get_current_state(self) -> AgentState:
"""
Get the current agent state from the graph.
Returns:
AgentState: The current state from the graph
Raises:
RuntimeError: If agent_loop_obj is not initialized
"""
if self.agent_loop_obj is None:
raise RuntimeError("Agent loop not initialized. Call run_task_generic first.")
return AgentState(
**self.agent_loop_obj.graph.get_state({"configurable": {"thread_id": self.thread_id}}).values
)
async def run_task_generic(
self,
eval_mode=False,
goal: str = None,
sites: List[str] = None,
current_datetime: Optional[str] = None,
session_id: str = None,
) -> Optional[ExperimentResult]:
langfuse_handler = None
if settings.advanced_features.langfuse_tracing and LangfuseCallbackHandler is not None:
langfuse_handler = LangfuseCallbackHandler()
logger.debug("Langfuse tracing enabled for agent loop")
agent = DynamicAgentGraph(None, langfuse_handler=langfuse_handler)
await agent.build_graph()
state: AgentState = default_state(
page=self.env.page if self.env else None,
observation=self.obs,
goal=goal if goal else self.obs['goal'],
)
state.sites = sites
await self.browser_update_state(state)
self.agent_loop_obj = AgentLoop(
thread_id=self.thread_id,
langfuse_handler=langfuse_handler,
graph=agent.graph,
env_pointer=self.env,
tracker=tracker,
)
state.current_datetime = current_datetime if current_datetime else datetime.datetime.now().isoformat()
state.pi = tracker.pi
agent_response = await self.agent_loop_obj.run(state=state)
reward = 0.0
i = 0
while True:
if agent_response.has_tools:
i += 1
state = self.get_current_state()
feedback = await AgentRunner.process_event_async(
state.messages[-1].tool_calls,
state.elements,
self.env.page,
self.env.tool_implementation_provider,
session_id=session_id,
tool_provider=self.env.tool_implementation_provider,
)
state.feedback = state.feedback + feedback
if len(feedback) > 0 and feedback[-1]['status'] == "alert":
logger.warning(f"Adding to stm the alert {feedback[-1]['message']}")
state.stm_steps_history.append(
"Response of (ActionAgent): {}".format(feedback[-1]['message'])
)
self.env.messages = state.messages
obs, reward, terminated, truncated, info = await self.env.step("")
if eval_mode and reward == 1.0 or len(tracker.steps) >= settings.evaluation.max_steps:
break
await self.browser_update_state(state)
self.agent_loop_obj.graph.update_state({"configurable": {"thread_id": self.thread_id}}, state)
agent_response = await self.agent_loop_obj.run(state=None)
elif agent_response.end:
tracker.final_answer = agent_response.answer
if self.env:
obs, reward, terminated, truncated, info = await self.env.step("")
break
else:
raise Exception("Agent stopped but no tools or finish.")
if eval_mode:
if self.env.chat:
await self.env.chat.add_message(
role="assistant",
msg="Final answer: {}".format(tracker.final_answer),
)
if len(tracker.steps) >= settings.evaluation.max_steps:
tracker.final_answer = "N/A"
obs, reward, terminated, truncated, info = await self.env.step("")
if sites and len(sites) > 1:
logger.debug("Sleep on finish if multi site")
await asyncio.sleep(15)
obs, reward, terminated, truncated, info = await self.env.step("")
tracker.collect_score(score=reward)
return ExperimentResult(
score=reward,
messages=state.messages,
answer=tracker.final_answer,
number_of_actions=tracker.actions_count,
steps=tracker.steps,
)
else:
return ExperimentResult(
score=0.0,
messages=state.messages,
answer=tracker.final_answer,
number_of_actions=tracker.actions_count,
steps=tracker.steps,
)
async def run_task_generic_yield(
self,
eval_mode=False,
goal: str = None,
sites: List[str] = None,
session_id: str = None,
current_datetime: Optional[str] = None,
chat_messages: List[AIMessage] = None,
) -> AsyncGenerator[Union[ExperimentResult, StreamEvent], None]:
logger.debug(
"Initiated agent with number of chat messages: {}".format(
len(chat_messages) if chat_messages else 0
)
)
langfuse_handler = None
if settings.advanced_features.langfuse_tracing and LangfuseCallbackHandler is not None:
langfuse_handler = LangfuseCallbackHandler()
logger.debug("Langfuse tracing enabled for agent loop")
agent = DynamicAgentGraph(None, langfuse_handler=langfuse_handler)
await agent.build_graph()
state: AgentState = default_state(
page=self.env.page if self.env else None,
observation=self.obs,
goal=goal if goal else self.obs['goal'],
chat_messages=chat_messages if chat_messages else [],
)
state.sites = sites
await self.browser_update_state(state)
self.agent_loop_obj = AgentLoop(
thread_id=self.thread_id,
langfuse_handler=langfuse_handler,
graph=agent.graph,
tracker=tracker,
env_pointer=self.env,
)
state.current_datetime = current_datetime if current_datetime else datetime.datetime.now().isoformat()
state.pi = tracker.pi
reward = 0.0
i = 0
first_time = True
event = None
while True:
agent_response = self.agent_loop_obj.run_stream(state=state if first_time else None)
first_time = False
final_event = None
async for event in agent_response:
final_event = event # Keep track of the last event
if isinstance(event, AgentLoopAnswer):
if event.has_tools:
i += 1
state = self.get_current_state()
feedback = await AgentRunner.process_event_async(
state.messages[-1].tool_calls,
state.elements,
self.env.page,
self.env.tool_implementation_provider,
session_id=session_id,
)
state.feedback = state.feedback + feedback
if len(feedback) > 0 and feedback[-1]['status'] == "alert":
logger.warning(f"Adding to stm the alert {feedback[-1]['message']}")
state.stm_steps_history.append(
"Response of (ActionAgent): {}".format(feedback[-1]['message'])
)
self.env.messages = state.messages
obs, reward, terminated, truncated, info = await self.env.step("")
if eval_mode and reward == 1.0 or len(tracker.steps) >= settings.evaluation.max_steps:
break # Break to handle final result outside the loop
await self.browser_update_state(state)
self.agent_loop_obj.graph.update_state(
{"configurable": {"thread_id": self.thread_id}}, state
)
# Break out of the async for loop to restart with new agent_response
break
elif event.end:
tracker.final_answer = event.answer
if self.env:
obs, reward, terminated, truncated, info = await self.env.step("")
yield ExperimentResult(
score=0.0,
messages=state.messages,
answer=tracker.final_answer,
number_of_actions=tracker.actions_count,
)
return # Exit the entire function
else:
raise Exception("Agent stopped but no tools or finish.")
else:
yield StreamEvent.parse(event)
# Handle the case where we broke out due to max_steps or reward == 1.0
if eval_mode and (reward == 1.0 or len(tracker.steps) >= settings.evaluation.max_steps):
break
# Handle final result outside the loop (matching your original structure)
if final_event and isinstance(final_event, AgentLoopAnswer):
tracker.final_answer = final_event.answer
if self.env:
obs, reward, terminated, truncated, info = await self.env.step("")
yield ExperimentResult(
score=0.0,
messages=state.messages,
answer=tracker.final_answer,
number_of_actions=tracker.actions_count,
)
async def main():
ar = AgentRunner(browser_enabled=False)
await ar.initialize_appworld_env()
await ar.run_task_generic(
eval_mode=False,
goal="Get my accounts top two accounts by revenue from digital sales",
)
if __name__ == '__main__':
asyncio.run(main())
|