Fix Thinking Off handling

.env.example

@@ -11,7 +11,7 @@ XTTS_IMAGE=ghcr.io/coqui-ai/xtts-streaming-server:latest-cuda121@sha256:f7fb3b1f
 XTTS_CACHE_DIR=/data/models/xtts-v2-cache
 XTTS_GPU_DEVICE=GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b
 AI_DNS=192.168.1.1
-DEFAULT_REASONING_EFFORT=medium
+DEFAULT_REASONING_EFFORT=off
 
 FAST_MODEL_FILE=qwen-mix/Qwen3.8-27B-IQ4-MIX.gguf
 MEDIUM_MODEL_FILE=qwen-pure/qwen3.8-27b-IQ4_XS-pure.gguf

compose.yaml

@@ -576,7 +576,7 @@ services:
       # request limit llama.cpp uses n_predict=-1 and a reasoning loop can
       # consume the complete context before yielding visible output.
       MAX_GENERATION_TOKENS: "8192"
-      DEFAULT_REASONING_EFFORT: "${DEFAULT_REASONING_EFFORT:-medium}"
+      DEFAULT_REASONING_EFFORT: "${DEFAULT_REASONING_EFFORT:-off}"
       IMAGE_DIR: /data/images
       IMAGE_WORKER_URL: http://image-worker:8086
       IMAGE_WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"

config/install.env.example

@@ -72,7 +72,7 @@ EXPERIMENTAL_MODEL_SHA256=40fac4050e940397dbf13087afd50f4734a11805bf9d65ef8ddd74
 VISION_PROJECTOR_FILE=qwen/mmproj-BF16.gguf
 VISION_PROJECTOR_URL=https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/mmproj-BF16.gguf
 VISION_PROJECTOR_SHA256=83ee4f4f205fa514161778c41df1ea14144faa0f713510893b63c2395f5c2d53
-DEFAULT_REASONING_EFFORT=medium
+DEFAULT_REASONING_EFFORT=off
 # All standard profiles use the MTP tensor embedded in their GGUF. A separate
 # draft-model artifact is neither downloaded nor passed to llama-server.
 

dev/test_router_support.py

@@ -129,8 +129,10 @@ class ChatImageInputTests(unittest.TestCase):
 
 
 class LlamaCppReasoningTests(unittest.TestCase):
-    def test_none_really_disables_thinking(self) -> None:
+    def test_disabled_values_really_disable_thinking(self) -> None:
-        request = {"reasoning_effort": "none", "messages": []}
+        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(
@@ -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):

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