fix(cli): report unresolved config interpolations cleanly in manifest validation

nemo_gym/environment/validation.py

@@ -16,6 +16,7 @@
 
 import orjson
 from omegaconf import DictConfig, OmegaConf
+from omegaconf.errors import InterpolationResolutionError
 from pydantic import BaseModel, ValidationError
 from yaml import YAMLError
 
@@ -35,6 +36,7 @@
     GlobalConfigDictParserConfig,
     dataset_agent_pins,
     resolve_dataset_agent,
+    translate_interpolation_error,
 )
 from nemo_gym.prompt import apply_prompt_to_row, load_prompt_config, validate_prompt_compatibility
 
@@ -171,14 +173,17 @@ def _resolve_manifest_composition(config_path: Path) -> ResolvedComposition:
     )
     parser = GlobalConfigDictParser()
     with _with_component_root(config_path):
-        resolved = parser.parse(
-            GlobalConfigDictParserConfig(
-                initial_global_config_dict=initial,
-                skip_load_from_cli=True,
-                skip_load_from_dotenv=True,
-                offline=True,
+        try:
+            resolved = parser.parse(
+                GlobalConfigDictParserConfig(
+                    initial_global_config_dict=initial,
+                    skip_load_from_cli=True,
+                    skip_load_from_dotenv=True,
+                    offline=True,
+                )
             )
-        )
+        except InterpolationResolutionError as e:
+            raise translate_interpolation_error(e) from e
     servers = parser.filter_for_server_instance_configs(resolved)
     by_instance = {server.name: server for server in servers}
     agents = [server for server in servers if server.SERVER_TYPE == "responses_api_agents"]

nemo_gym/global_config.py

@@ -1382,6 +1382,24 @@ def _apply_verbosity(global_config_dict: DictConfig) -> None:
         logging.getLogger().setLevel(logging.DEBUG)
 
 
+def translate_interpolation_error(e: InterpolationResolutionError) -> ConfigInterpolationError:
+    """Same class of user error as an unset '???' (see raise_on_missing_values), reported the same way
+    instead of letting omegaconf's traceback reach the top level. Covers both a missing `${key}`
+    (InterpolationKeyError) and a failing resolver such as `${oc.env:VAR}`, which carries its own
+    message and so is passed through as-is."""
+    match = re.search(r"Interpolation key '([^']+)' not found", str(e))
+    if not match:
+        return ConfigInterpolationError(str(e))
+    key = match.group(1)
+    return ConfigInterpolationError(
+        f"""Config value '{e.full_key}' references '{key}', which is not set after merging.
+
+Provide it via a CLI override, in env.yaml, or in a config you pass via config_paths.
+For example, on the command line:
+  ++{key}=<value>"""
+    )
+
+
 def set_global_config_dict(
     global_config_dict_parser_config: Optional[GlobalConfigDictParserConfig] = None,
     global_config_dict_parser_cls: Type[GlobalConfigDictParser] = GlobalConfigDictParser,
@@ -1390,21 +1408,7 @@ def set_global_config_dict(
     try:
         global_config_dict = global_config_dict_parser_cls().parse(global_config_dict_parser_config)
     except InterpolationResolutionError as e:
-        # Same class of user error as an unset '???' (see raise_on_missing_values), so report it the same
-        # way instead of letting omegaconf's traceback reach the top level. Covers both a missing `${key}`
-        # (InterpolationKeyError) and a failing resolver such as `${oc.env:VAR}`, which carries its own
-        # message and so is passed through as-is.
-        match = re.search(r"Interpolation key '([^']+)' not found", str(e))
-        if not match:
-            raise ConfigInterpolationError(str(e)) from e
-        key = match.group(1)
-        raise ConfigInterpolationError(
-            f"""Config value '{e.full_key}' references '{key}', which is not set after merging.
-
-Provide it via a CLI override, in env.yaml, or in a config you pass via config_paths.
-For example, on the command line:
-  ++{key}=<value>"""
-        ) from e
+        raise translate_interpolation_error(e) from e
 
     _GLOBAL_CONFIG_DICT = global_config_dict
 

tests/unit_tests/test_environment_validation.py

@@ -7,6 +7,7 @@
 
 import pytest
 
+from nemo_gym.config_types import ConfigInterpolationError
 from nemo_gym.environment.manifest import EnvironmentManifest, dump_manifest, load_manifest
 from nemo_gym.environment.validation import (
     EnvironmentValidationError,
@@ -357,6 +358,21 @@ def test_benchmark_uses_root_prompt_without_executing_prepare(tmp_path: Path) ->
     assert report.datasets[0].prompt_config.endswith("prompts/default.yaml")
 
 
+def test_unresolved_interpolation_is_reported_cleanly_not_as_a_traceback(tmp_path: Path) -> None:
+    manifest_path = _asset(tmp_path)
+    config_path = manifest_path.with_name("config.yaml")
+    config_path.write_text(
+        config_path.read_text(encoding="utf-8").replace(
+            "jsonl_fpath: environments/demo/data/example.jsonl",
+            "jsonl_fpath: environments/demo/data/example.jsonl\n      num_repeats: ${undefined_key}",
+        ),
+        encoding="utf-8",
+    )
+
+    with pytest.raises(ConfigInterpolationError, match="undefined_key"):
+        validate_environment(manifest_path)
+
+
 def test_malformed_benchmark_prompt_is_an_actionable_validation_error(tmp_path: Path) -> None:
     manifest_path = _asset(tmp_path, kind="benchmark")
     manifest_path.parent.joinpath("prompts/default.yaml").write_text("user: [broken\n", encoding="utf-8")