Default reasoning effort to medium

.env.example

@@ -11,6 +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
 
 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

@@ -570,6 +570,7 @@ 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: "${DEFAULT_REASONING_EFFORT:-medium}"
       IMAGE_DIR: /data/images
       IMAGE_WORKER_URL: http://image-worker:8086
       IMAGE_WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"

config/install.env.example

@@ -72,6 +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
 # All standard profiles use the MTP tensor embedded in their GGUF. A separate
 # draft-model artifact is neither downloaded nor passed to llama-server.
 

router/ai_profile_router.py

@@ -116,6 +116,8 @@ REQUEST_TIMEOUT = float(os.environ.get("REQUEST_TIMEOUT", "600"))   # s, Read-Ti
 CONNECT_TIMEOUT = float(os.environ.get("CONNECT_TIMEOUT", "10"))    # s, Connect-Timeout
 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()
 
 # --- Bildgenerierung und Referenzbild-Bearbeitung (FLUX.2 Klein 4B) ---
 LLAMA_SERVICE = os.environ.get("LLAMA_SERVICE", "mike-ai-llama-ui.service")
@@ -1272,10 +1274,19 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
     entfernt das wirkungslose Top-Level-Feld. Andere Template-Argumente des
     Clients bleiben erhalten.
     """
-    if "reasoning_effort" not in data:
-        return data
-
-    raw_effort = data.pop("reasoning_effort")
+    explicit_effort = "reasoning_effort" in data
+    if explicit_effort:
+        raw_effort = data.pop("reasoning_effort")
+    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.
+        existing_kwargs = data.get("chat_template_kwargs")
+        if isinstance(existing_kwargs, dict) and (
+                "reasoning_effort" in existing_kwargs
+                or "enable_thinking" in existing_kwargs):
+            return data
+        raw_effort = DEFAULT_REASONING_EFFORT
     effort = str(raw_effort).strip().lower() if raw_effort is not None else ""
     template_kwargs = data.get("chat_template_kwargs")
     if not isinstance(template_kwargs, dict):