Add editable LLM penalties and preserve sampling through backup restore

api_compat.py

@@ -3,7 +3,9 @@
 Effort is a template hint, not a guaranteed compute budget. llama.cpp forwards
 positive levels to the model template; it handles 'none' as thinking disabled.
 """
-CHAT_FIELDS=frozenset({'model','messages','stream','stream_options','temperature','top_p','top_k','max_tokens','max_completion_tokens','stop','seed','tools','tool_choice','parallel_tool_calls','response_format','presence_penalty','frequency_penalty','logprobs','top_logprobs','user','n','reasoning_effort'})
+import math
+
+CHAT_FIELDS=frozenset({'model','messages','stream','stream_options','temperature','top_p','top_k','max_tokens','max_completion_tokens','stop','seed','tools','tool_choice','parallel_tool_calls','response_format','repeat_penalty','presence_penalty','frequency_penalty','logprobs','top_logprobs','user','n','reasoning_effort'})
 LLAMA_EFFORTS=frozenset({'none','low','medium','high','xhigh'})
 # The installed llama.cpp server rejects minimal and max. Use nearest supported hints.
 EFFORT_ALIASES={'minimal':'low','max':'xhigh','ultra':'xhigh'}
@@ -14,7 +16,13 @@ def normalize_chat(data):
  """Return a new request and non-sensitive translation metadata; never mutate input."""
  unknown=set(data)-CHAT_FIELDS
  if unknown:raise CompatibilityError('Nicht unterstützte Chat-Felder: '+', '.join(sorted(unknown)))
- body=dict(data);requested=body.get('reasoning_effort')
+ body=dict(data)
+ for key,lo,hi in [('repeat_penalty',0,2),('presence_penalty',-2,2),('frequency_penalty',-2,2)]:
+  if key in body:
+   value=body[key]
+   if isinstance(value,bool) or not isinstance(value,(int,float)) or not math.isfinite(value) or not lo<=value<=hi:
+    raise CompatibilityError('Ungültiger '+key+'-Wert.')
+ requested=body.get('reasoning_effort')
  if requested is None:
   body.pop('reasoning_effort',None)
   return body,{}