Running through the OpenHands proxy (

<https://llm-proxy.app.all-hands.dev>

) and it looks like requests are not being cached anymore. Before I would get something like (from April 6):

{ "prompt_tokens": 6787, "cache_read_tokens": 1888, "cache_write_tokens": 0 }
{ "prompt_tokens": 6948, "cache_read_tokens": 6752, "cache_write_tokens": 0 }

whereas now I get this, no caching at all:

{ "prompt_tokens": 3155, "cache_read_tokens": 0, "cache_write_tokens": 0 }                                                          
{ "prompt_tokens": 3484, "cache_read_tokens": 0, "cache_write_tokens": 0 }

Has there been a change to the proxy?

which model are you using?

👀

I tried a couple of models through the OH proxy, minimax m2.5 and claude sonnet 4.6, both had zero cache read tokens. Before, I had non-zero cache read tokens. This was over multiple turns back to back, well within the 5 min windows. I also had the issue with minimax, so I guess that wouldn’t be affected by the Anthropic changes, right? Still no cache. I’m a bit lost in those screenshots. What can I try to troubleshoot this, please?

would appreciate if you have an exported conversation (eg events, base state) - if you are using the UI, click Export Conversation -- and share those with us to help the debugging process..

Sure, there are a couple of examples with non-zero and zero read cache, both with MiniMax M2.5 through the OpenHands proxy with dates 04/06 and 04/17 respectively.

Thanks! I think OH found smth, asking it to send a fix

ah, that makes sense, I did switch to streaming the response between those 2 trajectories

at least it doesn't break the actual caching, just the display

Awesome, thanks for the great support!