https://gitlab.synchro.net/main/sbbs/-/commit/a7718428415a389396c1ec21  
  Modified Files:  
   exec/chat_llm.js  
  Log Message:  
  chat_llm: treat a repeated question as a failed answer  
    
  Reply-text similarity cannot detect a stuck conversation.  A user asked  
  the same question three times running and got three answers that meant  
  the same thing but scored only 0.33-0.58 against each other, because  
  the model rewords itself at this temperature.  Genuinely new answers on  
  a single topic score 0.42-0.62, so the ranges overlap and no threshold  
  separates them.  
    
  The question itself is the reliable signal: someone retyping what they  
  just asked is telling you the previous answer did not land, whatever  
  its wording.  When the incoming question matches one of that speaker's  
  own recent questions, the previous answer is treated as a repeat  
  outright, bypassing the similarity test, and the existing regenerate-  
  then-deflect path runs.  
    
  Only the speaker's own turns are compared -- the bot restating a  
  question back is not the user asking twice.  Similarity still governs  
  whether the REGENERATED answer counts as another repeat, at a looser  
  threshold, since the strict one would essentially never fire there.  
  --- SBBSecho 3.37-Linux  
   * Origin: Vertrauen - [vert/cvs/bbs].synchro.net (1:103/705)