diff --git a/src/inspect_robots/cli.py b/src/inspect_robots/cli.py
index 383719d..7a1611a 100644
--- a/src/inspect_robots/cli.py
+++ b/src/inspect_robots/cli.py
@@ -71,6 +71,7 @@
     from inspect_robots.rollout import TrialRecord
     from inspect_robots.scene import Scene
     from inspect_robots.spaces import Box
+    from inspect_robots.task import Task
 
 
 def _styled(text: str, code: str) -> str:
@@ -478,6 +479,23 @@ def _resolve_or_exit(
         ) from exc
 
 
+def _apply_epochs_or_exit(task: Task, epochs: int) -> Task:
+    """Apply ``--epochs`` with a guided error instead of a raw traceback (#145).
+
+    ``replace()`` reruns ``Task.__post_init__``, which rejects a count below 1
+    via ``ConfigError`` — the same validation-error class ``_resolve_or_exit``
+    already converts to ``SystemExit`` for registry resolution.
+    """
+    from dataclasses import replace
+
+    from inspect_robots.errors import ConfigError
+
+    try:
+        return replace(task, epochs=epochs)
+    except ConfigError as exc:
+        raise SystemExit(str(exc)) from exc
+
+
 def _build_guardrails(
     space: Box, max_action_delta: float | None
 ) -> tuple[Approver, list[str], list[str]]:
@@ -835,8 +853,6 @@ def _build_and_announce_guardrails(args: argparse.Namespace, action_space: Box)
 
 
 def _cmd_run(args: argparse.Namespace) -> int:
-    from dataclasses import replace
-
     from inspect_robots import eval
     from inspect_robots.logging import JsonLogSink
     from inspect_robots.scene import Scene
@@ -885,7 +901,7 @@ def _cmd_run(args: argparse.Namespace) -> int:
     embodiment = resolved.embodiment
     try:
         if args.epochs is not None:
-            task = replace(task, epochs=args.epochs)
+            task = _apply_epochs_or_exit(task, args.epochs)
 
         _announce_components(resolved)
         approver = _build_and_announce_guardrails(args, embodiment.info.action_space)
@@ -995,8 +1011,6 @@ def _cmd_eval_set(args: argparse.Namespace) -> int:
     the embodiment once per task), the CLI resolves the embodiment exactly
     once for the whole set, so a real robot is not reconnected between tasks.
     """
-    from dataclasses import replace
-
     from inspect_robots import eval_set
 
     _check_shared_run_conflicts(args)
@@ -1005,7 +1019,7 @@ def _cmd_eval_set(args: argparse.Namespace) -> int:
     defaults = load_defaults(os.environ)
     tasks = [_resolve_or_exit("task", name) for name in task_names]
     if args.epochs is not None:
-        tasks = [replace(t, epochs=args.epochs) for t in tasks]
+        tasks = [_apply_epochs_or_exit(t, args.epochs) for t in tasks]
 
     resolved = _resolve_components(args, defaults)
     embodiment = resolved.embodiment
diff --git a/tests/test_registry_cli.py b/tests/test_registry_cli.py
index 0e09a14..67f997e 100644
--- a/tests/test_registry_cli.py
+++ b/tests/test_registry_cli.py
@@ -539,6 +539,29 @@ def test_cli_eval_set_sim_and_embodiment_conflict() -> None:
         main(["eval-set", "cubepick-reach", "--sim", "--embodiment", "cubepick"])
 
 
+def test_cli_eval_set_negative_epochs_is_a_guided_error(tmp_path: Path) -> None:
+    """--epochs -1 must not leak Task.__post_init__'s raw ConfigError (#145)."""
+    _register_task("kb/a")
+    try:
+        with pytest.raises(SystemExit, match="Epochs count must be >= 1, got -1"):
+            main(
+                [
+                    "eval-set",
+                    "kb/a",
+                    "--policy",
+                    "scripted",
+                    "--embodiment",
+                    "cubepick",
+                    "--epochs",
+                    "-1",
+                    "--log-dir",
+                    str(tmp_path),
+                ]
+            )
+    finally:
+        reg._FACTORIES["task"].pop("kb/a", None)
+
+
 def test_cli_eval_set_guardrail_flags_conflict() -> None:
     with pytest.raises(SystemExit, match="drop one"):
         main(
@@ -2462,10 +2485,9 @@ def close(self) -> None:
 def test_cli_run_closes_embodiment_when_validation_raises(
     monkeypatch: pytest.MonkeyPatch, tmp_path: Path
 ) -> None:
-    """A failure between resolving the embodiment and eval() (here: --epochs 0
-    raising ConfigError) must still close the embodiment — otherwise a bad flag
-    leaves real arms energized."""
-    from inspect_robots.errors import ConfigError
+    """A failure between resolving the embodiment and eval() (here: --epochs 0,
+    a guided SystemExit per #145) must still close the embodiment — otherwise a
+    bad flag leaves real arms energized."""
     from inspect_robots.mock import CubePickEmbodiment
 
     closed: list[bool] = []
@@ -2478,7 +2500,7 @@ def close(self) -> None:
     monkeypatch.setitem(reg._FACTORIES["embodiment"], "tracked-cubepick", _Tracked)
     monkeypatch.setenv(ENV_POLICY, "scripted")
     monkeypatch.setenv(ENV_EMBODIMENT, "tracked-cubepick")
-    with pytest.raises(ConfigError):
+    with pytest.raises(SystemExit, match="Epochs count must be >= 1"):
         main(
             [
                 "reach the cube",