nodejs>=20.4.*
-
AdaptClaude: (false)/true
- tries to make human/assistant prompts uniform between endpoints
- almost useless with streaming on... for now ;)
- effective with streaming off
- Human->H
- Human<-H
-
AntiStall: (false)/1/2
- no effect when using streaming
- 1 sends whatever was last when exceeding size (might send empty reply)
- 2 keeps going until it finds something usable or hitting size limit
-
ClearFlags: (false)/true
- possibly snake-oil
-
RecycleChats: (false)/true
- false is less likely to get caught in a censorship loop
-
StripAssistant: (false)/true
- might be good if your prompt/jailbreak itself ends with Assistant:
-
StripHuman: (false)/true
- bad idea without RecycleChats, sends only your very last message
I get empty replies
They send broken replies VERY often. If you have AntiStall set to false/1, once clewd receives ALL/hits the limit of claude's fever dream, it will send that shit to SillyTavern, IF it's broken, SillyTavern will error out and there will be no message
I keep getting H:
This can be a character issue, a main prompt issue, a jailbreak issue or even claude itself, AdaptClaude is there to try to make shit more consistent
What about context/temperature/top_k/top_p None of it is sent, some experimental shit will come out soon™ though
What about the flags
Only consumer_restricted_mode is the real bad one from what I've seen. If you get the yellow warning though, you'll know your account is possibly toast