refactor logging on environment in orch (#1594)

* remove logging

* add per env logging

* log to one file when having multiple worker

* intercept verifier log

* intercept verifier log

* intercept verifier log

* Add worker tag to env worker logger for identification

Co-authored-by: sami <sami@primeintellect.ai>

* Update changelog for logging config changes

Co-authored-by: sami <sami@primeintellect.ai>

* Rename env worker log config flag

Co-authored-by: sami <sami@primeintellect.ai>

---------

Signed-off-by: samsja <55492238+samsja@users.noreply.github.com>
Co-authored-by: sami jaghouar <sami@primeintellect.ai>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>

Author: samsja

CHANGELOG.md

@@ -33,4 +33,6 @@ Documenting changes which affect configuration usage patterns (added/moved/remov
 - **`model.lora.alpha`**: Changed default from 16.0 to 32.0 (2026-01-10)
 - **`orchestrator.env.log`**: Added logging configuration for environment workers. If set, enables logging with `level` (str, default: "warn") and `vf_level` (str, default: "warn") fields. If None (default), logging is disabled (#1561, 2026-01-13)
 - **`eval.watcher`**: Added flag (default `False`) to watch `weights_dir` for newly-created stable checkpoints and evaluate them as they appear (2026-01-14)
+- **`orchestrator.log.env_worker_logs`**: Added flag (default `True`) to write env worker logs to `logs/env_workers/{env_name}.log` (2026-01-15)
+- **`orchestrator.env.log`**: Removed. Use `orchestrator.log` for env worker logging instead (2026-01-15)
 - **`orchestrator.eval.retry.reraise`**: Changed default from `True` to `False`. When `False`, raises `tenacity.RetryError` after retries are exhausted instead of the original exception, allowing failed eval environments to be skipped with a warning (#1586, 2026-01-14)

docs/logging.md

@@ -57,32 +57,28 @@ For RL training, logs are organized by component:
 
 ## Per-Environment Worker Logging
 
-Environment workers run in **separate subprocesses** to isolate event loop lag. By default, they don't write to separate log files. To enable per-env logging, set the `log` field on the environment config:
+Environment workers run in **separate subprocesses** to isolate event loop lag. Worker logging is controlled at the orchestrator level via `orchestrator.log`:
 
 ```toml
-[[orchestrator.env]]
-id = "math-env"
-name = "math"
-log = { level = "debug", vf_level = "info" }
+[orchestrator.log]
+level = "debug"           # Log level for prime-rl logger
+vf_level = "info"         # Log level for verifiers library
+env_worker_logs = true    # Enable file logging for env workers
 ```
 
-This produces:
+When `env_worker_logs = true`, logs are written to:
 ```
 output_dir/
 └── logs/
     └── env_workers/
-        └── math/
-            ├── worker_0.log
-            ├── worker_1.log
-            └── ...
+        ├── {env_name_1}.log
+        ├── {env_name_2}.log
+        └── ...
 ```
 
-### Configuration Options
+All workers for an environment share the same log file.
 
-| Field | Description | Default |
-|-------|-------------|---------|
-| `level` | Log level for prime-rl logger (`debug`, `info`, `warn`, `error`) | `warn` |
-| `vf_level` | Log level for verifiers library logger | `warn` |
+Set `env_worker_logs = false` to disable worker file logging (workers inherit parent process logging).
 
 ## Torchrun
 

src/prime_rl/eval/eval.py

@@ -1,7 +1,5 @@
 import asyncio
 
-import verifiers as vf
-
 from prime_rl.eval.config import OfflineEvalConfig
 from prime_rl.eval.utils import run_evals
 from prime_rl.orchestrator.utils import set_semaphore
@@ -14,7 +12,7 @@
     setup_evals_client,
     update_weights,
 )
-from prime_rl.utils.logger import setup_logger
+from prime_rl.utils.logger import intercept_verifiers_logging, setup_logger
 from prime_rl.utils.monitor import setup_monitor
 from prime_rl.utils.pydantic_config import parse_argv
 from prime_rl.utils.utils import clean_exit, get_env_ids_to_install, get_step_path, install_env
@@ -26,7 +24,7 @@ async def eval(config: OfflineEvalConfig):
     logger = setup_logger(
         config.log.level, log_file=config.output_dir / "logs" / "eval.log" if config.log.file else None
     )
-    vf.setup_logging(level=config.log.vf_level.upper())
+    intercept_verifiers_logging(level=config.log.vf_level)
 
     env_names = [env.name or env.id for env in config.env]
     logger.info(f"Starting evals for {config.model.name} in environments {', '.join(env_names)}")

src/prime_rl/eval/utils.py

@@ -26,7 +26,7 @@
 from prime_rl.synthesize.utils import merge_reasoning_content, save_result
 from prime_rl.utils.client import setup_clients, setup_evals_client
 from prime_rl.utils.config import ClientConfig
-from prime_rl.utils.logger import get_logger, reset_logger, setup_logger
+from prime_rl.utils.logger import get_logger, intercept_verifiers_logging, reset_logger, setup_logger
 from prime_rl.utils.monitor import get_monitor
 from prime_rl.utils.utils import capitalize, get_eval_dir, get_step_path
 from prime_rl.utils.vf import (
@@ -689,6 +689,7 @@ def _run_evals_in_subprocess(
     # Setup logger for subprocess (reset first since we inherit parent's global state when forked)
     reset_logger()
     logger = setup_logger("info")
+    intercept_verifiers_logging(level="info")
     logger.info(f"Eval subprocess started for checkpoint step {ckpt_step}")
 
     # Create fresh clients in subprocess

src/prime_rl/orchestrator/config.py

@@ -297,30 +297,12 @@ class RetryConfig(BaseConfig):
     ] = False
 
 
-class EnvLogConfig(BaseConfig):
-    """Configures logging for an environment worker."""
-
-    level: Annotated[
-        str,
-        Field(description="Log level for prime-rl logger in worker (debug, info, warn, error)."),
-    ] = "warn"
-
-    vf_level: Annotated[
-        str,
-        Field(description="Log level for verifiers logger in worker (debug, info, warn, error)."),
-    ] = "warn"
-
-
 class EnvConfig(BaseConfig):
     """Configures an environment for training."""
 
     id: Annotated[str, Field(description="ID of the environment to use.")] = "reverse-text"
     args: Annotated[dict, Field(description="Arguments to pass to the environment.")] = {}
     name: Annotated[str | None, Field(description="Name of the environment to use.")] = None
-    log: Annotated[
-        EnvLogConfig | None,
-        Field(description="Logging config for this env's workers. If None, logging is disabled."),
-    ] = None
 
 
 class EvalEnvConfig(EnvConfig):

src/prime_rl/orchestrator/env_worker.py

@@ -5,7 +5,6 @@
 """
 
 import asyncio
-import logging
 import queue
 import uuid
 from dataclasses import dataclass
@@ -18,7 +17,7 @@
 
 from prime_rl.utils.client import setup_clients
 from prime_rl.utils.config import ClientConfig
-from prime_rl.utils.logger import reset_logger, setup_logger
+from prime_rl.utils.logger import intercept_verifiers_logging, reset_logger, setup_logger
 
 
 class WorkerDiedError(Exception):
@@ -188,19 +187,14 @@ def worker_main(
     log_level: str,
     vf_log_level: str,
     log_file: str | None,
+    worker_name: str | None = None,
 ):
     """Main entry point for worker process."""
     # Reset logger inherited from parent process, then setup fresh logger for this worker
     if log_file:
         reset_logger()
-        setup_logger(log_level, log_file=Path(log_file))
-        vf.setup_logging(level=vf_log_level.upper())
-        # Redirect verifiers to file instead of inherited stderr
-        vf_logger = logging.getLogger("verifiers")
-        vf_logger.handlers.clear()
-        vf_handler = logging.FileHandler(log_file)
-        vf_handler.setFormatter(logging.Formatter("%(asctime)s %(levelname)7s %(message)s", datefmt="%H:%M:%S"))
-        vf_logger.addHandler(vf_handler)
+        setup_logger(log_level, log_file=Path(log_file), append=True, tag=worker_name)
+        intercept_verifiers_logging(level=vf_log_level)
 
     # Load environment
     env = vf.load_environment(env_id, **env_args)
@@ -293,6 +287,7 @@ def start(self):
                 self.log_level,
                 self.vf_log_level,
                 self.log_file,
+                self.worker_name,
             ),
             daemon=True,
         )

src/prime_rl/orchestrator/orchestrator.py

@@ -44,7 +44,7 @@
     update_weights,
 )
 from prime_rl.utils.heartbeat import Heartbeat
-from prime_rl.utils.logger import setup_logger
+from prime_rl.utils.logger import intercept_verifiers_logging, setup_logger
 from prime_rl.utils.monitor import setup_monitor
 from prime_rl.utils.pydantic_config import parse_argv
 from prime_rl.utils.utils import (
@@ -66,7 +66,7 @@ async def orchestrate(config: OrchestratorConfig):
     logger = setup_logger(
         config.log.level, log_file=config.output_dir / "logs" / "orchestrator.log" if config.log.file else None
     )
-    vf.setup_logging(level=config.log.vf_level.upper())
+    intercept_verifiers_logging(level=config.log.vf_level)
     logger.info("Starting orchestrator")
 
     event_loop_lag_monitor = EventLoopLagMonitor()

src/prime_rl/orchestrator/scheduler.py

@@ -19,7 +19,7 @@
 from prime_rl.utils.client import update_weights
 from prime_rl.utils.config import ClientConfig
 from prime_rl.utils.logger import get_logger
-from prime_rl.utils.pathing import get_env_worker_log_dir
+from prime_rl.utils.pathing import get_env_worker_log_file
 from prime_rl.utils.utils import (
     get_broadcast_dir,
     get_latest_ckpt_step,
@@ -92,12 +92,11 @@ def __init__(
             self.env_names.append(env_name)
             self.workers[env_name] = []
 
-            # Setup log directory if logging is enabled for this env
-            env_log = env_config.log
-            env_log_dir = None
-            if env_log is not None and output_dir is not None:
-                env_log_dir = get_env_worker_log_dir(output_dir, env_name)
-                env_log_dir.mkdir(parents=True, exist_ok=True)
+            # Setup log file if env worker file logging is enabled (all workers share one file)
+            env_log_file = None
+            if config.log.env_worker_logs and output_dir is not None:
+                env_log_file = get_env_worker_log_file(output_dir, env_name)
+                env_log_file.parent.mkdir(parents=True, exist_ok=True)
 
             for worker_idx in range(self.workers_per_env):
                 worker = EnvWorker(
@@ -111,9 +110,9 @@ def __init__(
                     example_lookup=self.example_lookups[env_name],
                     sampling_args=self.sampling_args,
                     worker_name=f"{env_name}_{worker_idx}",
-                    log_level=env_log.level if env_log else "warn",
-                    vf_log_level=env_log.vf_level if env_log else "warn",
-                    log_file=str(env_log_dir / f"worker_{worker_idx}.log") if env_log_dir else None,
+                    log_level=config.log.level,
+                    vf_log_level=config.log.vf_level,
+                    log_file=str(env_log_file) if env_log_file else None,
                 )
                 self.workers[env_name].append(worker)
 

src/prime_rl/synthesize/synthesize.py

@@ -1,7 +1,5 @@
 import asyncio
 
-import verifiers as vf
-
 from prime_rl.orchestrator.utils import (
     set_semaphore,
 )
@@ -13,7 +11,7 @@
     setup_admin_clients,
     setup_clients,
 )
-from prime_rl.utils.logger import setup_logger
+from prime_rl.utils.logger import intercept_verifiers_logging, setup_logger
 from prime_rl.utils.pydantic_config import parse_argv
 from prime_rl.utils.utils import clean_exit, get_env_ids_to_install, install_env
 
@@ -24,7 +22,7 @@ async def synthesize(config: SynthesizeConfig):
     logger = setup_logger(
         config.log.level, log_file=config.output_dir / "logs" / "synthesize.log" if config.log.file else None
     )
-    vf.setup_logging(level=config.log.vf_level.upper())
+    intercept_verifiers_logging(level=config.log.vf_level)
 
     env_names = [env.name or env.id for env in config.env]
     logger.info(f"Starting synthetic data generation for {config.model.name} in environments {', '.join(env_names)}")

src/prime_rl/utils/config.py

@@ -73,6 +73,13 @@ class LogConfig(BaseConfig):
         ),
     ] = True
 
+    env_worker_logs: Annotated[
+        bool,
+        Field(
+            description="Whether env workers log to files. If True, workers write to logs/env_workers/{env_name}.log.",
+        ),
+    ] = True
+
     log_data: Annotated[
         bool,
         Field(