fix double experiment directory creation during training

src/holosoma/holosoma/config_types/env.py

@@ -1,7 +1,5 @@
 from __future__ import annotations
 
-from pydantic.dataclasses import dataclass
-
 from holosoma.config_types.action import ActionManagerCfg
 from holosoma.config_types.command import CommandManagerCfg
 from holosoma.config_types.curriculum import CurriculumManagerCfg
@@ -14,6 +12,7 @@
 from holosoma.config_types.simulator import SimulatorConfig
 from holosoma.config_types.termination import TerminationManagerCfg
 from holosoma.config_types.terrain import TerrainManagerCfg
+from pydantic.dataclasses import dataclass
 
 
 @dataclass(frozen=True)
@@ -34,6 +33,7 @@ class EnvConfig:
     robot: RobotConfig
     training: TrainingConfig
     logger: LoggerConfig
+    experiment_dir: str | None = None
 
 
 def get_tyro_env_config(tyro_config: ExperimentConfig) -> EnvConfig:

src/holosoma/holosoma/envs/base_task/base_task.py

@@ -1,7 +1,8 @@
 from __future__ import annotations
 
-import numpy as np
+from pathlib import Path
 
+import numpy as np
 from holosoma.config_types.env import EnvConfig
 from holosoma.config_types.full_sim import FullSimConfig
 from holosoma.managers.action import ActionManager
@@ -76,13 +77,17 @@ def __init__(
         torch._C._jit_set_profiling_mode(False)
         torch._C._jit_set_profiling_executor(False)
 
-        # Compute experiment directory from logger config
-        from holosoma.utils.experiment_paths import get_experiment_dir, get_timestamp
-
-        timestamp = get_timestamp()
-        experiment_dir = get_experiment_dir(
-            tyro_config.logger, tyro_config.training, timestamp, task_name=self._get_task_name()
-        )
+        # Use pre-computed experiment directory if provided (from train_agent.py),
+        # otherwise compute one (for replay, eval, etc.)
+        if tyro_config.experiment_dir is not None:
+            experiment_dir = Path(tyro_config.experiment_dir)
+        else:
+            from holosoma.utils.experiment_paths import get_experiment_dir, get_timestamp  # noqa: PLC0415
+
+            timestamp = get_timestamp()
+            experiment_dir = get_experiment_dir(
+                tyro_config.logger, tyro_config.training, timestamp, task_name=self._get_task_name()
+            )
 
         SimulatorClass = get_class(simulator_config._target_)
         full_sim_config = FullSimConfig(

src/holosoma/holosoma/train_agent.py

@@ -249,6 +249,7 @@ def train(tyro_config: ExperimentConfig, training_context: TrainingContext | Non
         env_target = tyro_config.env_class
 
         tyro_env_config = get_tyro_env_config(tyro_config)
+        tyro_env_config = dataclasses.replace(tyro_env_config, experiment_dir=str(experiment_dir))
         env = get_class(env_target)(tyro_env_config, device=device)
 
         # For manager system, pre-process config AFTER env creation