Map Hermes reasoning controls into Qwen templates

dev/mock_upstream.py

@@ -89,6 +89,8 @@ class Handler(BaseHTTPRequestHandler):
                       "total_tokens": 15},
             "mock_ctx": current_ctx(),
             "mock_authorization": self.headers.get("Authorization"),
+            "mock_reasoning_effort": body.get("reasoning_effort"),
+            "mock_chat_template_kwargs": body.get("chat_template_kwargs"),
         }
 
     def _stream(self, body: dict) -> None:

dev/test_local.sh

@@ -156,6 +156,25 @@ assert "Mock-Antwort" in d["choices"][0]["message"]["content"], d
 assert d.get("mock_authorization") is None, d
 ' && ok "Request wurde weitergeleitet, Modell ersetzt" || bad "Forwarding"
 
+echo "== Test 3b: Hermes-Reasoning erreicht das llama.cpp-Chat-Template"
+RESP=$(curl -sf "$BASE/v1/chat/completions" -H "Content-Type: application/json" \
+  -d '{"model":"qwen-fast","reasoning_effort":"none","messages":[{"role":"user","content":"Hallo"}]}')
+echo "$RESP" | python3 -c '
+import json,sys
+d=json.load(sys.stdin)
+assert d.get("mock_reasoning_effort") is None, d
+assert d.get("mock_chat_template_kwargs") == {"enable_thinking": False}, d
+' && ok "none wird als enable_thinking=false weitergegeben" || bad "Reasoning none"
+RESP=$(curl -sf "$BASE/v1/chat/completions" -H "Content-Type: application/json" \
+  -d '{"model":"qwen-fast","reasoning_effort":"medium","messages":[{"role":"user","content":"Hallo"}]}')
+echo "$RESP" | python3 -c '
+import json,sys
+d=json.load(sys.stdin)
+assert d.get("mock_reasoning_effort") is None, d
+assert d.get("mock_chat_template_kwargs") == {
+    "enable_thinking": True, "reasoning_effort": "medium"}, d
+' && ok "medium erreicht chat_template_kwargs" || bad "Reasoning medium"
+
 # --- 4. Streaming ----------------------------------------------------------------
 echo "== Test 4: Streaming (SSE)"
 RESP=$(curl -sfN "$BASE/v1/chat/completions" -H "Content-Type: application/json" \

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: