Default reasoning effort to medium

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):