Viktor.comAn AI coworker that actually does the work
CacheCatch
Agent Cache Audit
9 followers
Agent Cache Audit
9 followers
Audit LLM token costs and prompt-cache efficiency across LangSmith, Langfuse, Braintrust, and local IDE agents. CacheCatch detects cache breakers, estimates wasted spend, and gives exact prompt-layout fixes to cut AI costs up to 90%.
How does it actually identify a cache breaker in practice, like is it pattern matching on prompt structure or something more dynamic?
@polatetyl its multilayer.
CacheCatch first finds where stable prefixes diverge using character/token-level prefix diffing. Then it classifies likely causes with rules, and finally compares patterns across traces/sessions to catch structural cache breakers like dynamic fields appearing too early. So it’s not just regex, but regex is part of naming the issue
If you tried it i would love to hear your feedback.
A Slack or Teams alert when wasted spend crosses a daily threshold would be really useful, so I don't have to remember to check the dashboard. Bonus if it links straight to the specific prompt that triggered the spike.
@zkansevgzgzkj noted. thanks for trying and giving feedback. doing that requires connecting the report to the network and running on cron jobs, the reason i did not do that is i purposefully kept the report locally run.
So people don't have any concerns on sending their data over network.
But i will see how can we work around that, that is a really good feature to have.
Caught a sneaky cache breaker in one of our prompts that was silently doubling our spend — the layout fix recommendation was specific and worked on the first try. Wish I'd found this months ago.
@kezban55261 sounds great! happy it was useful to you. did you do the local IDE Agent report or platfrom report?