diff --git a/CHANGELOG.md b/CHANGELOG.md
index ef83812d..defcb7ee 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -129,6 +129,9 @@ All notable changes to SkillOpt are documented here. This project adheres to
   value is rejected with a clear error.
 
 ### Fixed
+- A flat `--cfg-options` key for an `env` setting (for example `workers=4`)
+  now overrides the YAML value with a structured config instead of being
+  silently replaced by it (#409).
 - Close four ways invalid evidence could reach the SkillOpt-Sleep gate: recall
   no longer brings a content twin of tonight's held-out task into training,
   a measured tool route no longer credits a self-reported `TOOL_CALL:` marker,
diff --git a/skillopt/config.py b/skillopt/config.py
index 12230d31..1b3f13f6 100644
--- a/skillopt/config.py
+++ b/skillopt/config.py
@@ -397,6 +397,14 @@ def apply_overrides(cfg: dict, overrides: list[str]) -> None:
             if structured and dotted:
                 section, subkey = dotted.split(".", 1)
                 cfg.setdefault(section, {})[subkey] = val
+            elif (
+                structured
+                and isinstance(cfg.get("env"), dict)
+                and canonical in cfg["env"]
+            ):
+                # flatten_config copies env keys over top-level ones, so a
+                # top-level copy of an env setting would be ignored.
+                cfg["env"][canonical] = val
             else:
                 cfg[canonical] = val
 
diff --git a/tests/test_cfg_options_flat_env_key.py b/tests/test_cfg_options_flat_env_key.py
new file mode 100644
index 00000000..68e586ae
--- /dev/null
+++ b/tests/test_cfg_options_flat_env_key.py
@@ -0,0 +1,43 @@
+"""Regression: a flat ``--cfg-options`` key for an ``env`` setting must win.
+
+``apply_overrides`` stored an unmapped flat key such as ``workers=4`` at the
+top level of a structured config. ``flatten_config`` copies top-level keys
+first and the ``env`` section's keys after them, so the YAML's ``env.workers``
+silently replaced the command-line value.
+"""
+from __future__ import annotations
+
+from skillopt.config import flatten_config, load_config
+
+
+def _write_config(tmp_path):
+    config = tmp_path / "config.yaml"
+    config.write_text(
+        "train:\n"
+        "  batch_size: 8\n"
+        "env:\n"
+        "  name: searchqa\n"
+        "  workers: 24\n"
+        "  max_turns: 1\n",
+        encoding="utf-8",
+    )
+    return str(config)
+
+
+def test_flat_override_of_env_key_wins_over_yaml(tmp_path):
+    cfg = load_config(_write_config(tmp_path), overrides=["workers=4", "max_turns=3"])
+    flat = flatten_config(cfg)
+    assert flat["workers"] == 4
+    assert flat["max_turns"] == 3
+
+
+def test_flat_and_dotted_env_overrides_agree(tmp_path):
+    path = _write_config(tmp_path)
+    flat_form = flatten_config(load_config(path, overrides=["workers=4"]))
+    dotted_form = flatten_config(load_config(path, overrides=["env.workers=4"]))
+    assert flat_form == dotted_form
+
+
+def test_new_flat_key_still_reaches_the_trainer(tmp_path):
+    cfg = load_config(_write_config(tmp_path), overrides=["limit=16"])
+    assert flatten_config(cfg)["limit"] == 16