Make reasoning levels enforce real token budgets

router/ai_profile_router.py

@@ -1270,6 +1270,15 @@ _REASONING_EFFORT_MAP = {
     "ultra": "xhigh",
 }
 _REASONING_OFF = {"", "none", "off", "disabled", "false"}
+_REASONING_BUDGET_TOKENS = {
+    "minimal": 256,
+    "low": 768,
+    "medium": 2048,
+    "high": 4096,
+    "xhigh": 8192,
+    "max": 8192,
+    "ultra": 8192,
+}
 
 
 def _load_global_system_policy() -> str:
@@ -1330,6 +1339,9 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
     Top-Level-Feld. llama.cpp akzeptiert das Feld zwar, reicht es dort aber
     nicht an das Jinja-Chat-Template weiter. Qwen3.8 erwartet stattdessen
     ``chat_template_kwargs.reasoning_effort`` bzw. ``enable_thinking=false``.
+    Zusätzlich erhält llama.cpp mit ``thinking_budget_tokens`` eine echte,
+    pro Request geltende Obergrenze. Dadurch sind die in Hermes sichtbaren
+    Stufen nicht bloß unterschiedlich formulierte Template-Hinweise.
 
     Die Funktion verändert den übergebenen Request absichtlich in-place und
     entfernt das wirkungslose Top-Level-Feld. Andere Template-Argumente des
@@ -1345,9 +1357,10 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
         # 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 (
+        if "thinking_budget_tokens" in data or (
+                isinstance(existing_kwargs, dict) and (
                 "reasoning_effort" in existing_kwargs
-                or "enable_thinking" 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 ""
@@ -1359,6 +1372,7 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
     if effort in _REASONING_OFF:
         template_kwargs.pop("reasoning_effort", None)
         template_kwargs["enable_thinking"] = False
+        data["thinking_budget_tokens"] = 0
         return data
 
     mapped = _REASONING_EFFORT_MAP.get(effort)
@@ -1373,6 +1387,7 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
 
     template_kwargs["enable_thinking"] = True
     template_kwargs["reasoning_effort"] = mapped
+    data["thinking_budget_tokens"] = _REASONING_BUDGET_TOKENS[effort]
     return data