update claude (#280)

* add uitars agent code

* improve claude

* improve claude

* improve claude

* improve claude

* improve claude

run_multienv_claude.py

@@ -11,9 +11,9 @@ import sys
 from typing import List, Dict
 import math
 from tqdm import tqdm
-from multiprocessing import Process, Manager
-# import lib_run_single
-# from desktop_env.desktop_env import DesktopEnv
+from multiprocessing import Process, Manager, current_process
+import lib_run_single
+from desktop_env.desktop_env import DesktopEnv
 from mm_agents.anthropic import AnthropicAgent as PromptAgent
 
 import fake_run_single as lib_run_single
@@ -26,41 +26,28 @@ load_dotenv()
 
 #  Logger Configs {{{ #
 logger = logging.getLogger()
-logger.setLevel(logging.DEBUG)
+logger.setLevel(logging.INFO)
 
 datetime_str: str = datetime.datetime.now().strftime("%Y%m%d@%H%M%S")
 
 file_handler = logging.FileHandler(
     os.path.join("logs", "normal-{:}.log".format(datetime_str)), encoding="utf-8"
 )
-debug_handler = logging.FileHandler(
-    os.path.join("logs", "debug-{:}.log".format(datetime_str)), encoding="utf-8"
-)
 stdout_handler = logging.StreamHandler(sys.stdout)
-sdebug_handler = logging.FileHandler(
-    os.path.join("logs", "sdebug-{:}.log".format(datetime_str)), encoding="utf-8"
-)
 
 file_handler.setLevel(logging.INFO)
-debug_handler.setLevel(logging.DEBUG)
 stdout_handler.setLevel(logging.INFO)
-sdebug_handler.setLevel(logging.DEBUG)
 
 formatter = logging.Formatter(
     fmt="\x1b[1;33m[%(asctime)s \x1b[31m%(levelname)s \x1b[32m%(module)s/%(lineno)d-%(processName)s\x1b[1;33m] \x1b[0m%(message)s"
 )
 file_handler.setFormatter(formatter)
-debug_handler.setFormatter(formatter)
 stdout_handler.setFormatter(formatter)
-sdebug_handler.setFormatter(formatter)
 
 stdout_handler.addFilter(logging.Filter("desktopenv"))
-sdebug_handler.addFilter(logging.Filter("desktopenv"))
 
 logger.addHandler(file_handler)
-logger.addHandler(debug_handler)
 logger.addHandler(stdout_handler)
-logger.addHandler(sdebug_handler)
 #  }}} Logger Configs #
 
 logger = logging.getLogger("desktopenv.experiment")
@@ -85,8 +72,18 @@ def config() -> argparse.Namespace:
         default="a11y_tree",
         help="Observation type",
     )
-    parser.add_argument("--screen_width", type=int, default=1920)
-    parser.add_argument("--screen_height", type=int, default=1080)
+    parser.add_argument(
+        "--provider_name", type=str, default="aws", choices=["aws", "virtualbox", "vmware", "docker", "azure"], help="Provider name"
+    )
+    parser.add_argument(
+        "--client_password", type=str, default="", help="Client password"
+    )
+    parser.add_argument(
+        "--screen_width", type=int, default=1920, help="Screen width"
+    )
+    parser.add_argument(
+        "--screen_height", type=int, default=1080, help="Screen height"
+    )
     parser.add_argument("--sleep_after_execution", type=float, default=0.0)
     parser.add_argument("--max_steps", type=int, default=15)
 
@@ -122,7 +119,7 @@ def config() -> argparse.Namespace:
     return args
 
 
-def distribute_tasks(test_all_meta: dict, num_envs: int) -> List[Dict]:
+def distribute_tasks(test_all_meta: dict) -> List[tuple]:
     """Distribute tasks evenly across environments."""
     # Flatten the tasks into a single list
     all_tasks = []
@@ -130,97 +127,35 @@ def distribute_tasks(test_all_meta: dict, num_envs: int) -> List[Dict]:
         for example_id in examples:
             all_tasks.append((domain, example_id))
     
-    # Calculate tasks per environment
-    tasks_per_env = math.ceil(len(all_tasks) / num_envs)
-    
-    # Distribute tasks
-    distributed_tasks = []
-    for i in range(num_envs):
-        env_tasks = {}
-        start_idx = i * tasks_per_env
-        end_idx = min((i + 1) * tasks_per_env, len(all_tasks))
-        
-        for domain, example_id in all_tasks[start_idx:end_idx]:
-            if domain not in env_tasks:
-                env_tasks[domain] = []
-            env_tasks[domain].append(example_id)
-        
-        distributed_tasks.append(env_tasks)
-    
-    return distributed_tasks
+    return all_tasks
 
 
 
-def run_env_tasks(env_idx: int, env: DesktopEnv, agent: PromptAgent, env_tasks: dict, args: argparse.Namespace, shared_scores: list):
+def run_env_tasks(task_queue, args, shared_scores):
     """Run tasks for a single environment."""
-    logger.info(f"Executing tasks in environment {env_idx + 1}/{args.num_envs}")
-    
-    for domain in tqdm(env_tasks, desc=f"Env{env_idx+1}-Domain"):
-        for example_id in tqdm(env_tasks[domain], desc="Example", leave=False):
-            config_file = os.path.join(
-                args.test_config_base_dir, f"examples/{domain}/{example_id}.json"
-            )
-            with open(config_file, "r", encoding="utf-8") as f:
-                example = json.load(f)
-
-            logger.info(f"[Env {env_idx+1}][Domain]: {domain}")
-            logger.info(f"[Env {env_idx+1}][Example ID]: {example_id}")
-            logger.info(f"[Env {env_idx+1}][Instruction]: {example['instruction']}")
-            
-            example_result_dir = os.path.join(
-                args.result_dir,
-                args.action_space,
-                args.observation_type,
-                args.model,
-                domain,
-                example_id,
-            )
-            os.makedirs(example_result_dir, exist_ok=True)
-
-            try:
-                lib_run_single.run_single_example(
-                    agent,
-                    env,
-                    example,
-                    args.max_steps,
-                    example["instruction"],
-                    args,
-                    example_result_dir,
-                    shared_scores,
-                )
-            except Exception as e:
-                import traceback
-                # logger traceback
-                logger.error(f"Exception in Env{env_idx+1} {domain}/{example_id}: {e}")
-                logger.error(traceback.format_exc())
-                env.controller.end_recording(
-                    os.path.join(example_result_dir, "recording.mp4")
-                )
-                with open(os.path.join(example_result_dir, "traj.jsonl"), "a") as f:
-                    f.write(
-                        json.dumps(
-                            {"Error": f"Time limit exceeded in {domain}/{example_id}"}
-                        )
-                    )
-                    f.write("\n")
-    
-    env.close()
-
-
-def test(args: argparse.Namespace, test_all_meta: dict) -> None:
-    logger.info("Args: %s", args)
-    
-    distributed_tasks = distribute_tasks(test_all_meta, args.num_envs)
-    
-    # First, set up all environments
-    logger.info("Setting up all environments...")
-    envs = []
-    agents = []
-    
-    for env_idx in range(args.num_envs):
-        logger.info(f"Setting up environment {env_idx + 1}/{args.num_envs}")
-        
+    active_environments = []
+    env = None
+    try:
+        from desktop_env.providers.aws.manager import IMAGE_ID_MAP
+        REGION = args.region
+        screen_size = (args.screen_width, args.screen_height)
+        ami_id = IMAGE_ID_MAP[REGION].get(screen_size, IMAGE_ID_MAP[REGION][(1920, 1080)])
+        env = DesktopEnv(
+            path_to_vm=args.path_to_vm,
+            action_space=args.action_space,
+            provider_name=args.provider_name,
+            region=REGION,
+            snapshot_name=ami_id,
+            screen_size=screen_size,
+            headless=args.headless,
+            os_type="Ubuntu",
+            require_a11y_tree=args.observation_type in ["a11y_tree", "screenshot_a11y_tree", "som"],
+            enable_proxy=True,
+            client_password=args.client_password
+        )
+        active_environments.append(env)
         agent = PromptAgent(
+            env=env,
             model=args.model,
             max_tokens=args.max_tokens,
             top_p=args.top_p,
@@ -228,50 +163,193 @@ def test(args: argparse.Namespace, test_all_meta: dict) -> None:
             action_space=args.action_space,
             observation_type=args.observation_type,
             max_trajectory_length=args.max_trajectory_length,
-            screen_size=(args.screen_width, args.screen_height),
+            client_password=args.client_password,
+            provider_name=args.provider_name,
+            screen_width=args.screen_width,
+            screen_height=args.screen_height
         )
-        agents.append(agent)
-        
-        from desktop_env.providers.aws.manager import IMAGE_ID_MAP
-        REGION = "us-east-1"
-        env = DesktopEnv(
-            path_to_vm=args.path_to_vm,
-            action_space=agent.action_space,
+        logger.info(f"Process {current_process().name} started.")
+        while True:
+            try:
+                item = task_queue.get(timeout=5)
+            except Exception:
+                break
+            domain, example_id = item
+            try:
+                config_file = os.path.join(
+                    args.test_config_base_dir, f"examples/{domain}/{example_id}.json"
+                )
+                with open(config_file, "r", encoding="utf-8") as f:
+                    example = json.load(f)
+                logger.info(f"[{current_process().name}][Domain]: {domain}")
+                logger.info(f"[{current_process().name}][Example ID]: {example_id}")
+                logger.info(f"[{current_process().name}][Instruction]: {example['instruction']}")
+                example_result_dir = os.path.join(
+                    args.result_dir,
+                    args.action_space,
+                    args.observation_type,
+                    args.model,
+                    domain,
+                    example_id,
+                )
+                os.makedirs(example_result_dir, exist_ok=True)
+                try:
+                    lib_run_single.run_single_example(
+                        agent,
+                        env,
+                        example,
+                        args.max_steps,
+                        example["instruction"],
+                        args,
+                        example_result_dir,
+                        shared_scores,
+                    )
+                except Exception as e:
+                    import traceback
+                    logger.error(f"Exception in {current_process().name} {domain}/{example_id}: {e}")
+                    logger.error(traceback.format_exc())
+                    try:
+                        env.controller.end_recording(
+                            os.path.join(example_result_dir, "recording.mp4")
+                        )
+                    except Exception as rec_e:
+                        logger.error(f"Failed to end recording: {rec_e}")
+                    with open(os.path.join(example_result_dir, "traj.jsonl"), "a") as f:
+                        f.write(
+                            json.dumps(
+                                {"Error": f"{domain}/{example_id} - {e}"}
+                            )
+                        )
+                        f.write("\n")
+            except Exception as e:
+                logger.error(f"Task-level error in {current_process().name}: {e}")
+                import traceback
+                logger.error(traceback.format_exc())
+    except Exception as e:
+        logger.error(f"Process-level error in {current_process().name}: {e}")
+        import traceback
+        logger.error(traceback.format_exc())
+    finally:
+        logger.info(f"{current_process().name} cleaning up environment...")
+        try:
+            if env:
+                env.close()
+                logger.info(f"{current_process().name} environment closed successfully")
+        except Exception as e:
+            logger.error(f"{current_process().name} error during environment cleanup: {e}")
 
-            provider_name="aws",
-            region="us-east-1",
-            snapshot_name=IMAGE_ID_MAP[REGION],
-            screen_size=(args.screen_width, args.screen_height),
-            headless=args.headless,
-            os_type="Ubuntu",
-            require_a11y_tree=args.observation_type
-            in ["a11y_tree", "screenshot_a11y_tree", "som"],
-        )
-        envs.append(env)
-    
-    logger.info("All environments are ready. Starting parallel task execution...")
-    
-    # Create a shared list for scores across processes
+
+def process_signal_handler(signum, frame, env_idx):
+    logger.info(f"Process {env_idx + 1} received signal {signum}. Shutting down...")
+    local_vars = frame.f_locals
+    active_environments = local_vars.get('active_environments', [])
+    for env in active_environments:
+        if env is not None:
+            try:
+                logger.info(f"Process {env_idx + 1} closing environment...")
+                env.close()
+                logger.info(f"Process {env_idx + 1} environment closed successfully")
+            except Exception as e:
+                logger.error(f"Process {env_idx + 1} error closing environment: {e}")
+    logger.info(f"Process {env_idx + 1} shutdown complete. Exiting.")
+    sys.exit(0)
+
+
+def signal_handler(signum, frame):
+    global is_terminating, active_environments, processes
+    if is_terminating:
+        return
+    is_terminating = True
+    logger.info(f"Received signal {signum}. Gracefully shutting down...")
+    for env in active_environments:
+        try:
+            logger.info(f"Closing environment...")
+            env.close()
+            logger.info(f"Environment closed successfully")
+        except Exception as e:
+            logger.error(f"Error closing environment: {e}")
+    for p in processes:
+        if p.is_alive():
+            try:
+                logger.info(f"Sending termination signal to process {p.name}...")
+                p.terminate()
+            except Exception as e:
+                logger.error(f"Error sending termination signal to process: {e}")
+    time.sleep(1)
+    for p in processes:
+        if p.is_alive():
+            try:
+                logger.info(f"Forcefully terminating process {p.name}...")
+                import signal as sig
+                os.kill(p.pid, sig.SIGKILL)
+            except Exception as e:
+                logger.error(f"Error forcefully terminating process: {e}")
+    logger.info("Shutdown complete. Exiting.")
+    sys.exit(0)
+
+
+def test(args: argparse.Namespace, test_all_meta: dict) -> None:
+    global processes
+    logger.info("Args: %s", args)
+    all_tasks = distribute_tasks(test_all_meta)
+    logger.info(f"Total tasks: {len(all_tasks)}")
     with Manager() as manager:
         shared_scores = manager.list()
-        
-        # Create and start processes for each environment
+        task_queue = manager.Queue()
+        for item in all_tasks:
+            task_queue.put(item)
+        num_envs = args.num_envs
         processes = []
-        for env_idx, (env, agent, env_tasks) in enumerate(zip(envs, agents, distributed_tasks)):
+        for i in range(num_envs):
             p = Process(
                 target=run_env_tasks,
-                args=(env_idx, env, agent, env_tasks, args, shared_scores)
+                args=(task_queue, args, shared_scores),
+                name=f"EnvProcess-{i+1}"
             )
-            processes.append(p)
+            p.daemon = True
             p.start()
-        
-        # Wait for all processes to complete
-        for p in processes:
-            p.join()
-        
-        # Convert shared list to regular list
+            processes.append(p)
+            logger.info(f"Started process {p.name} with PID {p.pid}")
+        try:
+            while True:
+                alive_count = 0
+                for idx, p in enumerate(processes):
+                    if not p.is_alive():
+                        logger.warning(f"Process {p.name} died, restarting...")
+                        new_p = Process(
+                            target=run_env_tasks,
+                            args=(task_queue, args, shared_scores),
+                            name=f"EnvProcess-Restart-{idx+1}"
+                        )
+                        new_p.daemon = True
+                        new_p.start()
+                        processes[idx] = new_p
+                        logger.info(f"Restarted process {new_p.name} with PID {new_p.pid}")
+                    else:
+                        alive_count += 1
+                if task_queue.empty():
+                    logger.info("All tasks finished.")
+                    break
+                if alive_count == 0:
+                    logger.error("All processes died, exiting.")
+                    break
+                time.sleep(5)
+            for p in processes:
+                p.join()
+        except KeyboardInterrupt:
+            logger.info("Main process received KeyboardInterrupt. Initiating graceful shutdown...")
+            raise
+        except Exception as e:
+            logger.error(f"Unexpected error while waiting for processes: {e}", exc_info=True)
+            for p in processes:
+                if p.is_alive():
+                    try:
+                        logger.info(f"Terminating process {p.name} due to error...")
+                        p.terminate()
+                    except Exception as term_e:
+                        logger.error(f"Error terminating process {p.name}: {term_e}")
+            raise
         scores = list(shared_scores)
-    
     logger.info(f"Average score: {sum(scores) / len(scores) if scores else 0}")
 
 
@@ -350,46 +428,43 @@ def get_result(action_space, use_model, observation_type, result_dir, total_file
 if __name__ == "__main__":
     ####### The complete version of the list of examples #######
     os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    import signal
+    import time
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+    try:
+        args = config()
 
-    args = config()
-    # save args to json in result_dir/action_space/observation_type/model/args.json
-    path_to_args = os.path.join(
-        args.result_dir,
-        args.action_space,
-        args.observation_type,
-        args.model,
-        "args.json",
-    )
-    os.makedirs(os.path.dirname(path_to_args), exist_ok=True)
-    with open(path_to_args, "w", encoding="utf-8") as f:
-        json.dump(vars(args), f, indent=4)
+        with open(args.test_all_meta_path, "r", encoding="utf-8") as f:
+            test_all_meta = json.load(f)
 
-    with open(args.test_all_meta_path, "r", encoding="utf-8") as f:
-        test_all_meta = json.load(f)
+        if args.domain != "all":
+            test_all_meta = {args.domain: test_all_meta[args.domain]}
 
-    if args.domain != "all":
-        test_all_meta = {args.domain: test_all_meta[args.domain]}
+        test_file_list = get_unfinished(
+            args.action_space,
+            args.model,
+            args.observation_type,
+            args.result_dir,
+            test_all_meta,
+        )
+        left_info = ""
+        for domain in test_file_list:
+            left_info += f"{domain}: {len(test_file_list[domain])}\n"
+        logger.info(f"Left tasks:\n{left_info}")
 
-    test_file_list = get_unfinished(
-        args.action_space,
-        args.model,
-        args.observation_type,
-        args.result_dir,
-        test_all_meta,
-    )
-    left_info = ""
-    for domain in test_file_list:
-        left_info += f"{domain}: {len(test_file_list[domain])}\n"
-    logger.info(f"Left tasks:\n{left_info}")
-
-    get_result(
-        args.action_space,
-        args.model,
-        args.observation_type,
-        args.result_dir,
-        test_all_meta,
-    )
-    test(args, test_file_list)
-
-
-# path_to_vm can be a list["xxx","xxx"]
+        get_result(
+            args.action_space,
+            args.model,
+            args.observation_type,
+            args.result_dir,
+            test_all_meta,
+        )
+        test(args, test_file_list)
+    except KeyboardInterrupt:
+        logger.info("Main process received KeyboardInterrupt.")
+    except Exception as e:
+        logger.error(f"Unexpected error in main process: {e}", exc_info=True)
+        signal_handler(signal.SIGTERM, None)
+    finally:
+        logger.info("Main process final cleanup...")