DeepSeek caches tokens. I built a CLI that caches solutions. 2nd time: 0 calls.

I'm a former hotel chef, building alone. I made a terminal agent that runs on DeepSeek, and I want to share some numbers because this sub actually cares about cost. DeepSeek's context cache already makes repeated tokens cheap. I wanted to go one level up: don't repeat the work. So before Virgil calls the model, it checks a local library of things it already solved (debug, code, UI, speech): - known answer: 0 calls - known pattern: 1 call to adapt it - new: full chain, then the result gets stored What I measured: - "build me a clock": 1st folder 1 call, 7 s. 2nd folder: 0 calls, 1 s. - small talk ("how late is it"): 1st time 0.02 cents, 2nd time 0. - repairing the same file again: 1st run 8 model requests, 2nd run 0. - 55 real bugs from HumanEvalFix: 1st pass 65 model calls. The same bugs renamed and reformatted so they don't match literally: 14 calls. What I did NOT measure yet: whether it transfers to different bugs of the same kind in other people's projects. That's what I need you for. Pricing is pay per use, 1 credit = 1 cent, 15 free to start, no account. A hard self-correcting task is around 3 credits. You can also plug in any OpenAI-compatible endpoint with your own key. npx virgil-cli Honest question for this sub: has anyone tried caching at the solution level instead of the token level? Where did it break for you?