Spaces:
Sleeping
Sleeping
File size: 22,108 Bytes
1ea26af |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 |
#
# the agent
__all__ = [
"register_template", "get_template",
"AgentResult", "ActionResult", "MultiStepAgent"
]
import json
import traceback
import time
from typing import List
from collections import Counter
from .model import LLM
from .session import AgentSession
from .tool import Tool
from .utils import KwargsInitializable, rprint, TemplatedString, parse_response, CodeExecutor, zwarn
TEMPLATES = {}
def register_template(templates):
for k, v in templates.items():
# assert k not in TEMPLATES
if k in TEMPLATES and v != TEMPLATES[k]:
zwarn(f"Overwrite previous templates for k={k}")
TEMPLATES[k] = v
def get_template(key: str):
return TemplatedString(TEMPLATES.get(key))
# --
# storage of the results for an agent call
class AgentResult(KwargsInitializable):
def __init__(self, **kwargs):
self.output = "" # formatted output
self.log = "" # other outputs
self.task = "" # target task
self.repr = None # explicit repr?
super().__init__(_assert_existing=False, **kwargs)
def to_dict(self):
return self.__dict__.copy()
def __contains__(self, item):
return item in self.__dict__
def __getitem__(self, item): # look like a dict
return self.__dict__[item]
def __repr__(self):
if self.repr: # if directly specified
return self.repr
ret = self.output if self.output else "N/A"
if self.log:
ret = f"{ret} ({self.log})"
return ret
class ActionResult(KwargsInitializable):
def __init__(self, action: str, result: str = None, **kwargs):
self.action = action
self.result = result
super().__init__(_assert_existing=False, **kwargs)
def __repr__(self):
return f"Action={self.action}, Result={self.result}"
# --
class StopReasons:
NORMAL_END = "Normal Ending."
MAX_STEP = "Max step exceeded."
MAX_TIME = "Time limit exceeded."
CODE_ERROR_PERFIX = "Code Execution Error:\n"
# --
# a basic class for a multi-step agent
class MultiStepAgent(KwargsInitializable):
def __init__(self, logger=None, **kwargs):
self.name = ""
self.description = ""
# self.sub_agents: List[MultiStepAgent] = [] # sub-agents (sth like advanced tools)
self.sub_agent_names = [] # sub-agent names (able to be found using getattr!)
self.tools: List[Tool] = [] # tools
self.model = LLM(_default_init=True) # main loop's model
self.logger = logger # 诊断日志器
self.templates = {} # template names: plan/action/end
self.max_steps = 10 # maximum steps
self.max_time_limit = 0 # early stop if exceeding this time (in seconds)
self.recent_steps = 5 # feed recent steps
self.store_io = True # whether store the inputs/outputs of the model in session
self.exec_timeout_with_call = 0 # how many seconds to timeout for each exec (0 means no timeout) (with sub-agent call)
self.exec_timeout_wo_call = 0 # how many seconds to timeout for each exec (0 means no timeout) (without sub-agent call)
self.obs_max_token = 8192 # avoid obs that is too long
# --
self.active_functions = [] # note: put active functions here!
# --
super().__init__(**kwargs)
self.templates = {k: get_template(v) for k, v in self.templates.items()} # read real templates from registered ones
# self.python_executor = CodeExecutor() # our own simple python executor (simply recreate it for each run!)
ALL_FUNCTIONS = {z.name: z for z in (self.sub_agents + self.tools)}
assert len(ALL_FUNCTIONS) == len(self.sub_agents + self.tools), "There may be repeated function names of sub-agents and tools."
self.ACTIVE_FUNCTIONS = {k: ALL_FUNCTIONS[k] for k in self.active_functions}
self.final_result = None # to store final result
# --
# repeat-output tracking for minimal prompt nudging
self._last_observation_text = None
self._repeat_count = 0
self._repeat_warning_msg = ""
@property
def sub_agents(self): # obtaining the sub-agents by getattr
return [getattr(self, name) for name in self.sub_agent_names]
# Training/evaluation methods removed - not needed for simple query processing
# get_call_stat(), get_seed(), set_seed() removed as per simplification goals
# called as a managed agent
# note: the communications/APIs between agents should be simple: INPUT={task, **kwargs}, OUTPUT={output(None if error), log}
def __call__(self, task: str, **kwargs):
# task = f"Complete the following task:\n{input_prompt}\n(* Your final answer should follow the format: {output_format})" # note: no longer format it here!
session = self.run(task, **kwargs) # run the process
final_results = session.get_current_step().get("end", {}).get("final_results", {})
ret = AgentResult(task=task, session=session, **final_results) # a simple wrapper
return ret
def get_function_definition(self, short: bool):
raise NotImplementedError("To be implemented")
# run as the main agent
def run(self, task, stream=False, session=None, max_steps: int = None, **extra_info):
start_pc = time.perf_counter()
# Initialize session
if session is None:
session = AgentSession(task=task, **extra_info)
max_steps = max_steps if max_steps is not None else self.max_steps
# --
if stream: # The steps are returned as they are executed through a generator to iterate on.
ret = self.yield_session_run(session=session, max_steps=max_steps) # return a yielder
else: # Outputs are returned only at the end. We only look at the last step.
for step_info in self.yield_session_run(session=session, max_steps=max_steps):
pass
ret = session
execution_time = time.perf_counter() - start_pc
rprint(f"ZZEnd task for {self.name} [ctime={time.ctime()}, interval={execution_time}]")
return ret
# main running loop
def yield_session_run(self, session, max_steps):
# run them!
start_pc = time.perf_counter()
# reset repeat-tracking per run
self._last_observation_text = None
self._repeat_count = 0
self._repeat_warning_msg = ""
self.init_run(session) # start
progress_state = {} # current state
stop_reason = None
while True:
step_idx = session.num_of_steps()
_error_counts = sum(self.get_obs_str(z['action']).strip().startswith(CODE_ERROR_PERFIX) for z in session.steps)
elapsed_time = time.perf_counter() - start_pc
# 埋点:打印每步的限制检查
print(f"[yield_session_run] Step {step_idx}: error_counts={_error_counts}, elapsed={elapsed_time:.1f}s")
print(f"[yield_session_run] Limits: max_steps={max_steps}, max_time_limit={self.max_time_limit}")
if (step_idx >= max_steps + _error_counts) or (step_idx >= int(max_steps*1.5)): # make up for the errors (but avoid too many steps)
print(f"[yield_session_run] STOP: MAX_STEP reached (step_idx={step_idx}, limit={max_steps + _error_counts} or {int(max_steps*1.5)})")
stop_reason = StopReasons.MAX_STEP # step limit
break
if (self.max_time_limit > 0) and (elapsed_time > self.max_time_limit):
print(f"[yield_session_run] STOP: MAX_TIME reached (elapsed={elapsed_time:.1f}s, limit={self.max_time_limit}s)")
stop_reason = StopReasons.MAX_TIME # time limit
break
rprint(f"# ======\nAgent {self.name} -- Step {step_idx}", timed=True)
_step_info = {"step_idx": step_idx}
session.add_step(_step_info) # simply append before running
yield from self.step(session, progress_state)
if self.step_check_end(session):
stop_reason = StopReasons.NORMAL_END
break
rprint(f"# ======\nAgent {self.name} -- Stop reason={stop_reason}", timed=True)
yield from self.finalize(session, progress_state, stop_reason) # ending!
self.end_run(session)
# --
def step(self, session, state):
_input_kwargs, _extra_kwargs = self.step_prepare(session, state)
_current_step = session.get_current_step()
# planning
has_plan_template = "plan" in self.templates
if has_plan_template: # planning to update state
plan_messages = self.templates["plan"].format(**_input_kwargs)
# 埋点:LLM 规划调用
if hasattr(self, 'logger') and self.logger:
self.logger.info("[WEB_LLM_PLAN] Task: %s", session.task[:200] + "..." if len(session.task) > 200 else session.task)
plan_response = self.step_call(messages=plan_messages, session=session)
plan_res = self._parse_output(plan_response)
# 埋点:LLM 规划结果
if hasattr(self, 'logger') and self.logger:
self.logger.info("[WEB_LLM_PLAN] Response: %s", plan_response[:500] + "..." if len(plan_response) > 500 else plan_response)
self.logger.info("[WEB_LLM_PLAN] Parsed: %s", plan_res)
# state update
if plan_res["code"]:
try:
new_state = eval(plan_res["code"]) # directly eval
except:
new_state = None
if new_state: # note: inplace update!
state.clear()
state.update(new_state)
else:
zwarn("State NOT changed due to empty output!")
else:
# if jailbreak detected, change the experience state by fource.
if plan_res['thought'] == 'Jailbreak or content filter violation detected. Please modify your prompt or stop with N/A.':
if 'experience' in state:
state['experience'].append(f'Jailbreak or content filter violation detected for the action {_input_kwargs["recent_steps_str"].split("Action:")[1]}. Please modify your prompt or stop with N/A.')
else:
state['experience'] = []
# hardcode here: disable the current visual_content if jailbreaking. This is because most jailbreak happens for images.
_input_kwargs['visual_content'] = None
# update session step
_current_step["plan"] = plan_res
plan_res["state"] = state.copy() # after updating the progress state (make a copy)
if self.store_io: # further storage
plan_res.update({"llm_input": plan_messages, "llm_output": plan_response})
yield {"type": "plan", "step_info": _current_step}
# predict action
_action_input_kwargs = _input_kwargs.copy()
_action_input_kwargs["state"] = json.dumps(state, ensure_ascii=False, indent=2) # there can be state updates
action_messages = self.templates["action"].format(**_action_input_kwargs)
# Inject minimal repeat-warning hint for NEXT step if previous outputs repeated
if getattr(self, "_repeat_warning_msg", ""):
if isinstance(action_messages, list):
action_messages = list(action_messages)
action_messages.append({"role": "user", "content": self._repeat_warning_msg})
# 埋点:LLM 动作调用
if hasattr(self, 'logger') and self.logger:
current_url = "unknown"
if "web_page" in _action_input_kwargs:
# 尝试从 accessibility tree 中提取 URL
web_page = _action_input_kwargs["web_page"]
if "RootWebArea" in web_page:
lines = web_page.split('\n')
for line in lines:
if "RootWebArea" in line and "'" in line:
current_url = line.split("'")[1] if "'" in line else "unknown"
break
self.logger.info("[WEB_LLM_ACTION] Browser_State: %s", current_url)
action_response = self.step_call(messages=action_messages, session=session)
action_res = self._parse_output(action_response)
# 埋点:LLM 动作结果
if hasattr(self, 'logger') and self.logger:
self.logger.info("[WEB_LLM_ACTION] Response: %s", action_response[:500] + "..." if len(action_response) > 500 else action_response)
self.logger.info("[WEB_LLM_ACTION] Actions: %s", action_res.get('code', 'No code generated'))
# perform action
step_res = self.step_action(action_res, _action_input_kwargs, **_extra_kwargs)
# update session info
_current_step["action"] = action_res
action_res["observation"] = step_res # after executing the step
# update repeat-tracking for next step
_obs_txt = self._normalize_observation(step_res)
if _obs_txt and _obs_txt == self._last_observation_text:
self._repeat_count += 1
else:
self._repeat_count = 0
self._last_observation_text = _obs_txt
if self._repeat_count > 0 and _obs_txt:
self._repeat_warning_msg = (
f"Notice: The last step produced the exact same output as before (repeated {self._repeat_count + 1} times): {_obs_txt}\n"
"If the task is complete, call stop(output=<YOUR_FINAL_ANSWER>, log='...') NOW to finalize.\n"
"Otherwise, investigate why the result repeated (e.g., state not updated, code had no effect) BEFORE continuing.\n"
"Good cases:\n"
"- stop(output=<YOUR_FINAL_ANSWER>, log='Answer verified; finalizing')\n"
"- Update progress state (e.g., add a completed note) and produce a DIFFERENT next action.\n"
"Bad cases:\n"
"- Printing the same output again without any change.\n"
"- Continuing without calling stop when the result is already final."
)
else:
self._repeat_warning_msg = ""
if self.store_io: # further storage
action_res.update({"llm_input": action_messages, "llm_output": action_response})
yield {"type": "action", "step_info": _current_step}
# --
def finalize(self, session, state, stop_reason: str):
has_end_template = "end" in self.templates
has_final_result = self.has_final_result()
final_results = self.get_final_result() if has_final_result else None
if has_end_template: # we have an ending module to further specify final results
_input_kwargs, _extra_kwargs = self.step_prepare(session, state)
# --
# special ask_llm if not normal ending
if stop_reason != StopReasons.NORMAL_END and hasattr(self, "tool_ask_llm"):
ask_llm_output = self.tool_ask_llm(session.task) # directly ask it
_input_kwargs["ask_llm_output"] = ask_llm_output
# --
if final_results:
stop_reason = f"{stop_reason} (with the result of {final_results})"
_input_kwargs["stop_reason"] = stop_reason
end_messages = self.templates["end"].format(**_input_kwargs)
end_response = self.step_call(messages=end_messages, session=session)
end_res = self._parse_output(end_response)
if self.store_io: # further storage
end_res.update({"llm_input": end_messages, "llm_output": end_response})
else: # no end module
end_res = {}
# no need to execute anything and simply prepare final outputs
_current_step = session.get_current_step()
if has_end_template or final_results is None: # try to get final results, end_module can override final_results
try:
final_results = eval(end_res["code"])
assert isinstance(final_results, dict) and "output" in final_results and "log" in final_results
except Exception as e: # use the final step's observation as the result!
# 埋点:finalizing step 错误详情
if hasattr(self, 'logger') and self.logger:
self.logger.error("[WEB_FINALIZING_ERROR] Function: finalize | Line: 302")
self.logger.error("[WEB_FINALIZING_ERROR] Error: %s", str(e))
self.logger.error("[WEB_FINALIZING_ERROR] End_Response: %s", end_response if 'end_response' in locals() else "No end_response")
self.logger.error("[WEB_FINALIZING_ERROR] End_Code: %s", end_res.get("code", "No code in end_res"))
self.logger.error("[WEB_FINALIZING_ERROR] Stop_Reason: %s", stop_reason if 'stop_reason' in locals() else "Unknown")
_log = "We are returning the final step's answer since there are some problems in the finalizing step." if has_end_template else ""
final_results = {"output": self.get_obs_str(_current_step), "log": _log}
end_res["final_results"] = final_results
# --
_current_step["end"] = end_res
yield {"type": "end", "step_info": _current_step}
# --
# --
# other helpers
def _normalize_observation(self, obs):
if isinstance(obs, (list, tuple)):
if not obs:
return ""
return str(obs[0]).strip()
return str(obs).strip() if obs is not None else ""
def get_obs_str(self, action, obs=None, add_seq_enum=True):
if obs is None:
obs = action.get("observation", "None")
if isinstance(obs, (list, tuple)): # list them
ret = "\n".join([(f"- Result {ii}: {zz}" if add_seq_enum else str(zz)) for ii, zz in enumerate(obs)])
else:
ret = str(obs)
# --
if len(ret) > self.obs_max_token:
ret = f"{ret[:self.obs_max_token]} ... (observation string truncated: exceeded {self.obs_max_token} characters)"
return ret
# common preparations of inputs
def _prepare_common_input_kwargs(self, session, state):
# previous steps
_recent_steps = session.get_latest_steps(count=self.recent_steps) # no including the last which is simply empty
_recent_steps_str = "\n\n".join([f"### Step {ss['step_idx']}\nThought: {ss['action']['thought']}\nAction: ```\n{ss['action']['code']}```\nObservation: {self.get_obs_str(ss['action'])}" for ii, ss in enumerate(_recent_steps)])
_current_step = session.get_current_step()
_current_step_action = _current_step.get("action", {})
_current_step_str = f"Thought: {_current_step_action.get('thought')}\nAction: ```\n{_current_step_action.get('code')}```\nObservation: {self.get_obs_str(_current_step_action)}"
# tools and sub-agents
ret = {
"task": session.task, "state": json.dumps(state, ensure_ascii=False, indent=2),
"recent_steps": _recent_steps, "recent_steps_str": _recent_steps_str,
"current_step": _current_step, "current_step_str": _current_step_str,
}
for short in [True, False]:
_subagent_str = "## Sub-Agent Functions\n" + "\n".join([z.get_function_definition(short) for z in self.sub_agents])
_tool_str = "## Tool Functions\n" + "\n".join([z.get_function_definition(short) for z in self.tools])
_subagent_tool_str = f"{_subagent_str}\n\n{_tool_str}"
_kkk = "subagent_tool_str_short" if short else "subagent_tool_str_long"
ret[_kkk] = _subagent_tool_str
# --
return ret
def _parse_output(self, output: str):
_target_list = ["Thought:", "Code:"]
if (output is None) or (output.strip() == ""):
output = "Thought: Model returns empty output. There might be a connection error or your input is too complex. Consider simplifying your query." # error without any output
_parsed_output = parse_response(output, _target_list, return_dict=True)
_res = {k[:-1].lower(): _parsed_output[k] for k in _target_list}
# parse code
_res["code"] = CodeExecutor.extract_code(output)
return _res
# --
# an explicit mechanism for ending
def has_final_result(self):
return self.final_result is not None
def put_final_result(self, final_result):
self.final_result = final_result
def get_final_result(self, clear=True):
ret = self.final_result
if clear:
self.final_result = None
return ret
# --
# --
# to be implemented in sub-classes
def init_run(self, session):
pass
def end_run(self, session):
pass
def step_call(self, messages, session, model=None):
if model is None:
model = self.model
response = model(messages)
return response
def step_prepare(self, session, state):
_input_kwargs = self._prepare_common_input_kwargs(session, state)
_extra_kwargs = {}
return _input_kwargs, _extra_kwargs
def step_action(self, action_res, action_input_kwargs, **kwargs):
python_executor = CodeExecutor()
python_executor.add_global_vars(**self.ACTIVE_FUNCTIONS) # to avoid that things might get re-defined at some place ...
_exec_timeout = self.exec_timeout_with_call if any((z in action_res["code"]) for z in self.sub_agent_names) else self.exec_timeout_wo_call # choose timeout value
python_executor.run(action_res["code"], catch_exception=True, timeout=_exec_timeout) # handle err inside!
ret = python_executor.get_print_results() # currently return a list of printed results
rprint(f"Obtain action res = {ret}", style="white on yellow")
return ret # return a result str
def step_check_end(self, session):
return self.has_final_result()
|