diff --git a/cheetahclaws/skill/tools.py b/cheetahclaws/skill/tools.py
index e63ad651..89ae4347 100644
--- a/cheetahclaws/skill/tools.py
+++ b/cheetahclaws/skill/tools.py
@@ -39,10 +39,50 @@
 }
 
 
+def _normalize_args(args: object) -> tuple[bool, str]:
+    """Coerce model-supplied `args` into a plain string, or say why not.
+
+    The schema declares `args` as a string, but small/open models on custom
+    providers often wrap every tool input in an object (they see
+    `input: {...}` for all tools). Before the check, `substitute_arguments`
+    did `prompt.replace("$ARGUMENTS", args)` and crashed with a raw
+    `TypeError: replace() argument 2 must be str, not dict` — an error the
+    model cannot correct, so it retried the same call and gave up (#182).
+
+    Accepted shapes:
+    - str: unchanged.
+    - dict with a single string value (the observed `{"args": "..."}`): unwrap.
+    - list of str: join with spaces (the observed `["--env", "prod"]`).
+    Anything else returns (False, explanation) so the caller can hand the
+    model a concrete instruction instead of a traceback.
+    """
+    if isinstance(args, str):
+        return True, args
+    if isinstance(args, dict):
+        values = [v for v in args.values() if isinstance(v, str)]
+        if len(args) == 1 and len(values) == 1:
+            return True, values[0]
+        return False, (
+            f"Skill 'args' must be a string, got an object with keys "
+            f"{sorted(args)}. Pass args as a single string, e.g. "
+            f"\"--corpus X --k 5\"."
+        )
+    if isinstance(args, list):
+        if all(isinstance(v, str) for v in args):
+            return True, " ".join(args)
+        return False, "Skill 'args' list must contain only strings."
+    return False, (
+        f"Skill 'args' must be a string, got {type(args).__name__}. "
+        f"Pass args as a single string."
+    )
+
+
 def _skill_tool(params: dict, config: dict) -> str:
     """Execute a skill by name and return its output."""
     skill_name = params.get("name", "").strip()
-    args = params.get("args", "")
+    args_ok, args = _normalize_args(params.get("args", ""))
+    if not args_ok:
+        return f"Error: {args}"
 
     # Look up by name first, then by trigger
     skill = None
diff --git a/tests/test_skills.py b/tests/test_skills.py
index b7c0f6c6..0d49142f 100644
--- a/tests/test_skills.py
+++ b/tests/test_skills.py
@@ -48,6 +48,7 @@ def skill_dir(tmp_path, monkeypatch):
     skills_dir.mkdir()
     (skills_dir / "commit.md").write_text(COMMIT_MD, encoding="utf-8")
     (skills_dir / "review.md").write_text(REVIEW_MD, encoding="utf-8")
+    (skills_dir / "deploy.md").write_text(ARGS_MD, encoding="utf-8")
 
     monkeypatch.setattr(_loader, "_get_skill_paths", lambda: [skills_dir])
     # Also patch the builtin list to be empty so tests are predictable
@@ -133,9 +134,9 @@ def test_parse_skill_file_allowed_tools(tmp_path):
 
 def test_load_skills(skill_dir):
     skills = load_skills()
-    assert len(skills) == 2
+    assert len(skills) == 3
     names = {s.name for s in skills}
-    assert names == {"commit", "review"}
+    assert names == {"commit", "review", "deploy"}
 
 
 def test_load_skills_empty_dir(tmp_path, monkeypatch):
@@ -278,3 +279,54 @@ def test_load_skills_finds_nested(tmp_path, monkeypatch):
     monkeypatch.setattr(_loader, "_BUILTIN_SKILLS", [])
     skills = load_skills()
     assert any(s.name == "myskill" for s in skills)
+
+
+# ------------------------------------------------------------------
+# _skill_tool args normalization (#182)
+# ------------------------------------------------------------------
+
+from cheetahclaws.skill.tools import _skill_tool
+
+
+def _fake_run(*args, **kwargs):
+    class _Ev:
+        text = "deployed to prod"
+    yield _Ev()
+
+
+def _call_skill(params):
+    """Run _skill_tool against the deploy skill. A regression that raises will
+    fail the test outright — that is the point of the suite."""
+    return _skill_tool(params, {})
+
+
+def test_skill_tool_coerces_dict_wrapped_single_value_args(skill_dir, monkeypatch):
+    """#182: a model that wraps args in an object (`{"args": {...}}`) must not
+    crash the Skill tool — take the single string value and proceed."""
+    monkeypatch.setattr("cheetahclaws.agent.run", _fake_run)
+    out = _call_skill({"name": "deploy", "args": {"args": "--env prod"}})
+    assert "CRASH" not in out
+    assert "deployed to prod" in out
+
+
+def test_skill_tool_joins_list_args(skill_dir, monkeypatch):
+    monkeypatch.setattr("cheetahclaws.agent.run", _fake_run)
+    out = _call_skill({"name": "deploy", "args": ["--env", "prod"]})
+    assert "CRASH" not in out
+    assert "deployed to prod" in out
+
+
+def test_skill_tool_rejects_multivalue_dict_with_clear_error(skill_dir, monkeypatch):
+    """Multi-key dicts have no unambiguous string form — return an error the
+    model can act on instead of a raw TypeError."""
+    monkeypatch.setattr("cheetahclaws.agent.run", _fake_run)
+    out = _call_skill({"name": "deploy", "args": {"env": "prod", "version": "1"}})
+    assert "CRASH" not in out
+    assert "must be a string" in out
+
+
+def test_skill_tool_rejects_non_string_args(skill_dir, monkeypatch):
+    monkeypatch.setattr("cheetahclaws.agent.run", _fake_run)
+    out = _call_skill({"name": "deploy", "args": 123})
+    assert "CRASH" not in out
+    assert "must be a string" in out