Fix Thinking Off handling

dev/test_router_support.py

@@ -129,14 +129,16 @@ class ChatImageInputTests(unittest.TestCase):
 
 
 class LlamaCppReasoningTests(unittest.TestCase):
-    def test_none_really_disables_thinking(self) -> None:
-        request = {"reasoning_effort": "none", "messages": []}
-        normalized = _normalize_llamacpp_reasoning(request)
-        self.assertNotIn("reasoning_effort", normalized)
-        self.assertEqual(
-            normalized["chat_template_kwargs"],
-            {"enable_thinking": False},
-        )
+    def test_disabled_values_really_disable_thinking(self) -> None:
+        for effort in (None, "none", "off", "disabled", False):
+            with self.subTest(effort=effort):
+                request = {"reasoning_effort": effort, "messages": []}
+                normalized = _normalize_llamacpp_reasoning(request)
+                self.assertNotIn("reasoning_effort", normalized)
+                self.assertEqual(
+                    normalized["chat_template_kwargs"],
+                    {"enable_thinking": False},
+                )
 
     def test_low_and_medium_reach_chat_template(self) -> None:
         for effort in ("low", "medium"):
@@ -171,10 +173,13 @@ class LlamaCppReasoningTests(unittest.TestCase):
             "reasoning_effort": "low",
         })
 
-    def test_request_without_effort_is_unchanged(self) -> None:
+    def test_request_without_effort_uses_safe_off_default(self) -> None:
         request = {"messages": []}
         self.assertIs(_normalize_llamacpp_reasoning(request), request)
-        self.assertNotIn("chat_template_kwargs", request)
+        self.assertEqual(
+            request["chat_template_kwargs"],
+            {"enable_thinking": False},
+        )
 
 
 class ChatGenerationLimitTests(unittest.TestCase):