Fix Thinking Off handling

router/ai_profile_router.py

@@ -117,7 +117,7 @@ CONNECT_TIMEOUT = float(os.environ.get("CONNECT_TIMEOUT", "10"))    # s, Connect
 POLL_INTERVAL = float(os.environ.get("POLL_INTERVAL", "2"))         # s, Polling-Intervall
 MAX_GENERATION_TOKENS = int(os.environ.get("MAX_GENERATION_TOKENS", "8192"))
 DEFAULT_REASONING_EFFORT = os.environ.get(
-    "DEFAULT_REASONING_EFFORT", "medium").strip().lower()
+    "DEFAULT_REASONING_EFFORT", "off").strip().lower()
 
 # --- Bildgenerierung und Referenzbild-Bearbeitung (FLUX.2 Klein 4B) ---
 LLAMA_SERVICE = os.environ.get("LLAMA_SERVICE", "mike-ai-llama-ui.service")
@@ -1280,7 +1280,9 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
     else:
         # A client may already speak llama.cpp's native template dialect.
         # Preserve that explicit choice; otherwise apply the platform-wide
-        # default so every OpenAI-compatible client behaves consistently.
+        # default.  Hermes deliberately omits reasoning_effort when its UI is
+        # set to Off, so the safe default must remain Off; enabled levels are
+        # sent explicitly by Hermes and other capable clients.
         existing_kwargs = data.get("chat_template_kwargs")
         if isinstance(existing_kwargs, dict) and (
                 "reasoning_effort" in existing_kwargs