fix(cli): guard --max-action-delta against invalid explicit values

CHANGELOG.md

@@ -101,6 +101,16 @@ All notable changes to this project are documented here. The format is based on
 
 ### Fixed
 
+- **An explicit invalid `--max-action-delta` now fails fast instead of silently
+  running with weaker guardrails** (#154). Non-finite or non-positive values
+  were previously caught by `_build_guardrails`'s degrade-per-component path
+  (meant for derived limits an embodiment's space can't support) and
+  downgraded to a stderr warning, so the run proceeded with clamp-only
+  guardrails despite the operator explicitly asking for a tighter limit. Both
+  `run` and `eval-set` now reject a malformed explicit value in the shared
+  conflict check, before anything resolves or energizes. Derived-limit
+  degradation (no explicit flag, an embodiment declaring no bounds) is
+  unaffected — it still warns and continues.
 - **`--epochs 0` or a negative value now exits with a guided error instead of a
   raw traceback** (#145). Both `inspect-robots run` and `inspect-robots eval-set`
   catch the `ConfigError` raised by `Task`'s epoch validation and surface it

src/inspect_robots/cli.py

@@ -745,7 +745,7 @@ class _ResolvedComponents(NamedTuple):
 
 
 def _check_shared_run_conflicts(args: argparse.Namespace) -> None:
-    """Reject flag combinations invalid for both ``run`` and ``eval-set``."""
+    """Reject flag combinations and malformed values invalid for both ``run`` and ``eval-set``."""
     if args.sim and args.embodiment:
         raise SystemExit(
             "--sim selects your configured sim_embodiment; "
@@ -755,6 +755,15 @@ def _check_shared_run_conflicts(args: argparse.Namespace) -> None:
         raise SystemExit(
             "--max-action-delta tunes the guardrails that --disable-guardrails turns off — drop one"
         )
+    if args.max_action_delta is not None and not (
+        math.isfinite(args.max_action_delta) and args.max_action_delta > 0
+    ):
+        # A malformed *explicit* value is a CLI input error, not an embodiment
+        # limitation: fail fast here rather than let _build_guardrails's
+        # degrade-per-component path (meant for derived limits an embodiment's
+        # space can't support) downgrade it to a warning and run with weaker
+        # guardrails than the operator explicitly asked for.
+        raise SystemExit(f"--max-action-delta must be finite and > 0, got {args.max_action_delta}")
 
 
 def _resolve_components(args: argparse.Namespace, defaults: Defaults) -> _ResolvedComponents:

tests/test_registry_cli.py

@@ -2835,6 +2835,36 @@ def test_cli_max_action_delta_conflicts_with_disable(tmp_path: Path) -> None:
         )
 
 
+@pytest.mark.parametrize("value", ["-1", "0", "inf", "nan"])
+def test_cli_run_rejects_invalid_max_action_delta(value: str, tmp_path: Path) -> None:
+    """An explicit malformed --max-action-delta fails fast (#154), not a soft warning:
+    it is a CLI input error, distinct from a derived limit an embodiment's space
+    cannot support (test_cli_degraded_guardrails_warn_but_run stays a warning)."""
+    with pytest.raises(SystemExit, match="finite and > 0"):
+        main(
+            [
+                "run",
+                "--task",
+                "cubepick-reach",
+                "--policy",
+                "scripted",
+                "--embodiment",
+                "cubepick",
+                "--log-dir",
+                str(tmp_path),
+                "--max-action-delta",
+                value,
+            ]
+        )
+
+
+@pytest.mark.parametrize("value", ["-1", "0", "inf", "nan"])
+def test_cli_eval_set_rejects_invalid_max_action_delta(value: str) -> None:
+    """Same guard as run, exercised through eval-set's shared conflict check (#154)."""
+    with pytest.raises(SystemExit, match="finite and > 0"):
+        main(["eval-set", "cubepick-reach", "--max-action-delta", value])
+
+
 def test_cli_degraded_guardrails_warn_but_run(
     tmp_path: Path, capsys: pytest.CaptureFixture[str]
 ) -> None: