/
OS-Worldf593f35from functools import partial
import logging
from typing import Dict, List, Tuple
from mm_agents.os_symphony.agents.memoryer_agent import ReflectionMemoryAgent
from mm_agents.os_symphony.agents.os_aci import OSACI
from mm_agents.os_symphony.core.module import BaseModule
from mm_agents.os_symphony.memory.procedural_memory import PROCEDURAL_MEMORY
from mm_agents.os_symphony.utils.common_utils import (
call_llm_formatted,
extract_coords_from_action_dict,
parse_action_from_string,
parse_code_from_string,
create_pyautogui_code,
)
from mm_agents.os_symphony.utils.formatters import (
SINGLE_ACTION_FORMATTER,
CODE_VALID_FORMATTER,
)
logger = logging.getLogger("desktopenv.agent")
class Worker(BaseModule):
def __init__(
self,
engine_params_for_orchestrator: Dict,
engine_params_for_memoryer: Dict,
os_aci: OSACI,
platform: str,
client_password: str,
max_trajectory_length: int = 8,
enable_reflection: bool = True,
):
"""
Worker receives the main task and generates actions, without the need of hierarchical planning
Args:
worker_engine_params: Dict
Parameters for the worker agent
os_aci: Agent
The grounding agent to use
platform: str
OS platform the agent runs on (darwin, linux, windows)
max_trajectory_length: int
The amount of images turns to keep
enable_reflection: bool
Whether to enable reflection
"""
super().__init__(platform=platform)
self.client_password = client_password
self.temperature = engine_params_for_orchestrator.get("temperature", 0.0)
self.tool_config = engine_params_for_orchestrator.get("tool_config", "")
self.use_thinking = engine_params_for_orchestrator.get("model", "") in [
"claude-opus-4-20250514",
"claude-sonnet-4-20250514",
"claude-3-7-sonnet-20250219",
"claude-sonnet-4-5-20250929",
]
self.engine_params_for_orchestrator = engine_params_for_orchestrator
self.engine_params_for_memoryer = engine_params_for_memoryer
self.os_aci: OSACI = os_aci
self.max_trajectory_length = max_trajectory_length if not self.engine_params_for_orchestrator.get("keep_first_image", False) else max_trajectory_length - 1
self.enable_reflection = enable_reflection
self.reset()
def reset(self):
# set_cell_values only occurs in linux; meanwhile there is no fail option in the other benchmarks
if self.platform in ["windows", "macos"]:
skipped_actions = ["set_cell_values", "fail"]
else:
skipped_actions = []
# Hide code agent action entirely if no env/controller is available
if not getattr(self.os_aci, "env", None) or not getattr(
getattr(self.os_aci, "env", None), "controller", None
):
skipped_actions.append("call_code_agent")
self.orchestrator_sys_prompt = PROCEDURAL_MEMORY.construct_simple_worker_procedural_memory(
agent_class=type(self.os_aci),
skipped_actions=skipped_actions,
tool_config=self.tool_config,
platform=self.platform
).replace("CURRENT_OS", self.platform).replace("CLIENT_PASSWORD", self.client_password)
# Worker contains orchestrator and reflection agent
self.orchestrator_agent = self._create_agent(
engine_params=self.engine_params_for_orchestrator,
system_prompt=self.orchestrator_sys_prompt
)
self.memoryer_agent = ReflectionMemoryAgent(self.engine_params_for_memoryer)
self.instruction = None
self.turn_count = 0
self.worker_history = []
self.coords_history = []
# For loop detection
self.action_dict_history = []
def flush_messages(self):
"""Flush messages based on the model's context limits.
This method ensures that the agent's message history does not exceed the maximum trajectory length.
Side Effects:
- Modifies the messages of generator, reflection, and bon_judge agents to fit within the context limits.
"""
engine_type = self.engine_params_for_orchestrator.get("engine_type", "")
# Flush strategy for long-context models: keep all text, only keep latest images
if engine_type in ["anthropic", "openai", "gemini", "vllm"]:
max_images = self.max_trajectory_length
# for agent in [self.generator_agent, self.reflection_agent]:
for agent in [self.orchestrator_agent]:
if agent is None:
continue
# keep latest k images
img_count = 0
stop_idx = 1 if self.engine_params_for_orchestrator.get("keep_first_image", False) else -1
for i in range(len(agent.messages) - 1, stop_idx, -1):
# for j in range(len(agent.messages[i]["content"])):
for j in range(len(agent.messages[i]["content"]) - 1, -1, -1):
if "image" in agent.messages[i]["content"][j].get("type", ""):
img_count += 1
if img_count > max_images:
del agent.messages[i]["content"][j]
# Flush strategy for non-long-context models: drop full turns
else:
# generator msgs are alternating [user, assistant], so 2 per round
if len(self.orchestrator_agent.messages) > 2 * self.max_trajectory_length + 1:
self.orchestrator_agent.messages.pop(1)
self.orchestrator_agent.messages.pop(1)
def generate_next_action(self, instruction: str, obs: Dict, is_last_step: bool) -> Tuple[Dict, List]:
"""
Predict the next action(s) based on the current observation.
"""
print("=" * 30, f"Turn {self.turn_count + 1}", "=" * 30)
print("=" * 10)
print(instruction)
print("=" * 10)
self.os_aci.assign_screenshot(obs)
self.os_aci.set_task_instruction(instruction)
generator_message = (
""
if self.turn_count > 0
else "The initial screen is provided. No action has been taken yet."
)
# Load the task into the system prompt
if is_last_step:
# Eager mode: must decide done / fail
prompt_with_instructions = PROCEDURAL_MEMORY.construct_eager_mode_procedural_memory(agent_class=type(self.os_aci)).replace(
"TASK_DESCRIPTION", instruction
).replace(
"CURRENT_OS", self.platform
)
print(f'Eager Mode Started, Instruction: {prompt_with_instructions}')
self.orchestrator_agent.add_system_prompt(prompt_with_instructions)
generator_message += "Note: 'EAGER MODE' is enabled. You must determine whether the task is done or fail in this step!!!"
else:
tutorials = ""
for idx, t in enumerate(self.os_aci.tutorials, start=1):
tutorials += f"### Tutorial {idx}:\n {t}\n"
prompt_with_instructions = self.orchestrator_sys_prompt.replace(
"TASK_DESCRIPTION", instruction
).replace(
"TUTORIAL_PLACEHOLDER", tutorials
)
self.orchestrator_agent.add_system_prompt(prompt_with_instructions)
# print(self.orchestrator_agent.system_prompt)
### Reflection Part
reflection_info = {}
if self.enable_reflection:
# set instruction to memory agent
self.memoryer_agent.add_instruction(instruction)
reflection = None
# Differentiate the operation mode of last step
last_code_summary = ""
mode = "gui"
if (
hasattr(self.os_aci, "last_code_agent_result")
and self.os_aci.last_code_agent_result is not None
):
# If code agent is called last step, we use its execution result as step behavior.
code_result = self.os_aci.last_code_agent_result
mode = "code"
last_code_summary += f"Subtask Instruction: {code_result['task_instruction']}\nSteps Completed: {code_result['steps_executed']}\nCompletion Reason: {code_result['completion_reason']}\nExec Summary: {code_result['summary']}\n"
if (
hasattr(self.os_aci, "last_search_agent_result")
and self.os_aci.last_search_agent_result is not None
):
mode = "search"
# retrieve reflection!!!
reflection_info = self.memoryer_agent.get_reflection(
cur_obs=obs,
# only use the string after "(next action)" in orchestrator's output
generator_output=parse_action_from_string(self.worker_history[-1]) if self.turn_count != 0 else "",
coordinates=self.coords_history[-1] if self.turn_count != 0 else [],
mode=mode,
code_exec_summary=last_code_summary,
action_dict=self.action_dict_history[-1] if self.turn_count != 0 else {}
)
reflection = reflection_info['reflection']
logger.info(f'[Reflection]: {reflection}')
if reflection:
generator_message += f"REFLECTION: You MUST use this reflection on the latest action:\n{reflection}\n"
else:
generator_message += "You should go on with your plan.\n"
else:
generator_message += "You should go on with your plan.\n"
# Add code agent result from previous step if available (from full task or subtask execution)
if (
hasattr(self.os_aci, "last_code_agent_result")
and self.os_aci.last_code_agent_result is not None
):
code_result = self.os_aci.last_code_agent_result
generator_message += f"\nCODE AGENT RESULT:\n"
generator_message += (
f"Task/Subtask Instruction: {code_result['task_instruction']}\n"
)
generator_message += f"Steps Completed: {code_result['steps_executed']}\n"
generator_message += f"Max Steps: {code_result['budget']}\n"
generator_message += (
f"Completion Reason: {code_result['completion_reason']}\n"
)
generator_message += f"Summary: {code_result['summary']}\n"
generator_message += "\n"
# Reset the code agent result after adding it to context
self.os_aci.last_code_agent_result = None
if (
hasattr(self.os_aci, "last_search_agent_result")
and self.os_aci.last_search_agent_result is not None
):
# Retrieve the result dictionary
search_result = self.os_aci.last_search_agent_result
# Add a clear, distinct header for this section in the prompt
generator_message += f"\nSEARCH AGENT RESULT:\n"
# Add contextual metadata from the search task
generator_message += f"Search Query: {search_result['query']}\n"
generator_message += f"Search Completion Reason: {search_result['completion_reason']}\n"
generator_message += "Search Result: "
# Add the most important part: the tutorial found by the agent.
# This is given a prominent sub-header so the LLM knows to pay close attention.
if search_result["completion_reason"] == "DONE":
generator_message += f'Search is completed, the tutorial it found has been already added to your system prompt.\n'
elif search_result["completion_reason"] == "FAIL":
generator_message += f"Search is fail, the failure reason or the hint is as follow: {search_result['final_answer']}\n"
# CRITICAL: Reset the search agent result after adding it to the context.
# This prevents it from being added to the prompt again in the next turn.
self.os_aci.last_search_agent_result = None
# Finalize the generator message
self.orchestrator_agent.add_message(
generator_message, image_content=obs["screenshot"], role="user", put_text_last=True
)
# Generate the plan and next action
format_checkers = [
SINGLE_ACTION_FORMATTER,
partial(CODE_VALID_FORMATTER, self.tool_config),
]
plan = call_llm_formatted(
self.orchestrator_agent,
format_checkers,
temperature=self.engine_params_for_orchestrator.get("temperture", 0.1),
use_thinking=self.use_thinking,
)
self.worker_history.append(plan)
self.orchestrator_agent.add_message(plan, role="assistant")
logger.info("PLAN:\n %s", plan)
# Extract the next action from the plan
# 此时的plan code e.g. agent.click('xxxxx', 1)
plan_code = parse_code_from_string(plan)
action_dict, coordinates = None, None
try:
assert plan_code, "Plan code should not be empty"
# exec_code e.g. import pyautogui; pyautogui.click(1, 2);
exec_code, action_dict = create_pyautogui_code(self.os_aci, plan_code, obs)
coordinates = extract_coords_from_action_dict(action_dict)
except Exception as e:
logger.error(
f"Could not evaluate the following plan code:\n{plan_code}\nError: {e}"
)
exec_code, action_dict = self.os_aci.wait(
1.333
) # Skip a turn if the code cannot be evaluated
self.action_dict_history.append(action_dict)
executor_info = {
"refined_instruction": self.instruction,
"plan": plan,
"plan_code": plan_code,
"exec_code": exec_code,
"coordinates": coordinates,
"reflection": reflection_info,
"code_agent_output": (
self.os_aci.last_code_agent_result
if hasattr(self.os_aci, "last_code_agent_result")
and self.os_aci.last_code_agent_result is not None
else None
),
"search_agent_output": (
self.os_aci.last_search_agent_result
if hasattr(self.os_aci, "last_search_agent_result")
and self.os_aci.last_search_agent_result is not None
else None
)
}
self.turn_count += 1
self.coords_history.append(coordinates)
self.flush_messages()
return executor_info, [exec_code]