Map Hermes reasoning controls into Qwen templates

dev/test_router_support.py

@@ -27,6 +27,7 @@ from ai_profile_router import (  # noqa: E402
     _context_matches,
     _normalize_chat_image,
     _normalize_chat_images,
+    _normalize_llamacpp_reasoning,
     _request_has_image,
 )
 
@@ -126,6 +127,55 @@ class ChatImageInputTests(unittest.TestCase):
         self.assertEqual(request, normalized)
 
 
+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_low_and_medium_reach_chat_template(self) -> None:
+        for effort in ("low", "medium"):
+            with self.subTest(effort=effort):
+                request = {"reasoning_effort": effort, "messages": []}
+                normalized = _normalize_llamacpp_reasoning(request)
+                self.assertEqual(normalized["chat_template_kwargs"], {
+                    "enable_thinking": True,
+                    "reasoning_effort": effort,
+                })
+
+    def test_unsupported_high_levels_are_clamped_to_xhigh(self) -> None:
+        for effort in ("high", "xhigh", "max", "ultra"):
+            with self.subTest(effort=effort):
+                request = {"reasoning_effort": effort, "messages": []}
+                normalized = _normalize_llamacpp_reasoning(request)
+                self.assertEqual(
+                    normalized["chat_template_kwargs"]["reasoning_effort"],
+                    "xhigh",
+                )
+
+    def test_existing_template_kwargs_are_preserved(self) -> None:
+        request = {
+            "reasoning_effort": "low",
+            "chat_template_kwargs": {"preserve_thinking": True},
+            "messages": [],
+        }
+        normalized = _normalize_llamacpp_reasoning(request)
+        self.assertEqual(normalized["chat_template_kwargs"], {
+            "preserve_thinking": True,
+            "enable_thinking": True,
+            "reasoning_effort": "low",
+        })
+
+    def test_request_without_effort_is_unchanged(self) -> None:
+        request = {"messages": []}
+        self.assertIs(_normalize_llamacpp_reasoning(request), request)
+        self.assertNotIn("chat_template_kwargs", request)
+
+
 class RetentionTests(unittest.TestCase):
     def test_oldest_pairs_are_removed(self) -> None:
         with tempfile.TemporaryDirectory() as temp: