feat: add no_upgrade config option for environment installation

Add a `no_upgrade` field to EnvConfig that controls whether the
`--no-upgrade` flag is passed to `prime env install`. This prevents
the install command from upgrading existing packages, preserving
locked dependencies (e.g., from uv.lock).

Changes:
- Add `no_upgrade: bool = False` to EnvConfig
- Update install_env() to accept and pass --no-upgrade flag
- Update get_env_ids_to_install() to return dict with no_upgrade settings
- Update all callers (eval, orchestrator, synthesize) to use new API

Author: Maxusmusti

CHANGELOG.md

@@ -37,3 +37,4 @@ Documenting changes which affect configuration usage patterns (added/moved/remov
 - **`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)
 - **`model.ep`**: Expert parallelism now supported (with auto/custom impl only), changed from the old behaviour when `ep>1` was a no-op to a proper parallelization of the MoE layers. (#1595, 2026-01-15)
+- **`env.no_upgrade`**: Added flag (default `False`) to pass `--no-upgrade` to `prime env install`, preserving locked dependencies instead of upgrading packages (2026-01-19)

src/prime_rl/eval/eval.py

@@ -32,8 +32,8 @@ async def eval(config: OfflineEvalConfig):
 
     # Install environments
     env_ids_to_install = get_env_ids_to_install(config.env)
-    for env_id in env_ids_to_install:
-        install_env(env_id)
+    for env_id, no_upgrade in env_ids_to_install.items():
+        install_env(env_id, no_upgrade=no_upgrade)
 
     # Initialize the monitor
     logger.info(f"Initializing monitor ({config.wandb})")

src/prime_rl/orchestrator/config.py

@@ -303,6 +303,10 @@ class EnvConfig(BaseConfig):
     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
+    no_upgrade: Annotated[
+        bool,
+        Field(description="If True, don't upgrade existing packages during install. Useful with locked dependencies."),
+    ] = False
 
 
 class EvalEnvConfig(EnvConfig):

src/prime_rl/orchestrator/orchestrator.py

@@ -83,13 +83,17 @@ async def orchestrate(config: OrchestratorConfig):
         tomli_w.dump(config.model_dump(exclude_none=True, mode="json"), f)
 
     # Install environments
-    env_ids_to_install = set()
-    env_ids_to_install.update(get_env_ids_to_install(config.env))
+    env_ids_to_install: dict[str, bool] = {}
+    for env_id, no_upgrade in get_env_ids_to_install(config.env).items():
+        existing = env_ids_to_install.get(env_id, False)
+        env_ids_to_install[env_id] = existing or no_upgrade
     if config.eval is not None:
-        env_ids_to_install.update(get_env_ids_to_install(config.eval.env))
+        for env_id, no_upgrade in get_env_ids_to_install(config.eval.env).items():
+            existing = env_ids_to_install.get(env_id, False)
+            env_ids_to_install[env_id] = existing or no_upgrade
 
-    for env_id in env_ids_to_install:
-        install_env(env_id)
+    for env_id, no_upgrade in env_ids_to_install.items():
+        install_env(env_id, no_upgrade=no_upgrade)
 
     # Setup client
     logger.info(

src/prime_rl/synthesize/synthesize.py

@@ -30,8 +30,8 @@ async def synthesize(config: SynthesizeConfig):
 
     # Install environments
     env_ids_to_install = get_env_ids_to_install(config.env)
-    for env_id in env_ids_to_install:
-        install_env(env_id)
+    for env_id, no_upgrade in env_ids_to_install.items():
+        install_env(env_id, no_upgrade=no_upgrade)
 
     # Setup clients
     logger.info(

src/prime_rl/utils/utils.py

@@ -281,21 +281,35 @@ def default_dtype(dtype):
         torch.set_default_dtype(prev)
 
 
-def install_env(env_id: str) -> None:
-    """Install an environment in subprocess."""
+def install_env(env_id: str, no_upgrade: bool = False) -> None:
+    """Install an environment in subprocess.
+
+    Args:
+        env_id: Environment ID to install.
+        no_upgrade: If True, don't upgrade existing packages during install.
+    """
     logger = get_logger()
     logger.info(f"Installing environment {env_id}")
     install_cmd = ["uv", "run", "--no-sync", "prime", "env", "install", env_id]
+    if no_upgrade:
+        install_cmd.append("--no-upgrade")
     result = subprocess.run(install_cmd, capture_output=True, text=True)
     if result.returncode != 0:
         raise RuntimeError(f"Failed to install environment {env_id} (stdout={result.stdout}, stderr={result.stderr})")
     logger.info(f"Successfully installed environment {env_id}")
 
 
-def get_env_ids_to_install(env_configs: list[EnvConfig] | list[EvalEnvConfig]) -> set[str]:
-    """Get the list of environment IDs to install."""
-    env_ids_to_install = set()
+def get_env_ids_to_install(env_configs: list[EnvConfig] | list[EvalEnvConfig]) -> dict[str, bool]:
+    """Get the environments to install with their no_upgrade settings.
+
+    Returns:
+        Dict mapping env_id -> no_upgrade. If the same env appears multiple times
+        with different settings, no_upgrade=True takes precedence.
+    """
+    env_ids_to_install: dict[str, bool] = {}
     for env_config in env_configs:
         if "/" in env_config.id:
-            env_ids_to_install.add(env_config.id)
+            # If env already exists, preserve no_upgrade=True if either has it
+            existing = env_ids_to_install.get(env_config.id, False)
+            env_ids_to_install[env_config.id] = existing or env_config.no_upgrade
     return env_ids_to_install