sjh9714/dsh-lean

README excerpt

English 简体中文 dsh-lean dsh sends 8,246 tokens before it reads your prompt. 3,700 of them are tools your session never calls. A cache miss costs 31x a cache hit, and the first request of every session pays the entire tool-schema prefix at the miss rate. On a six-request task, averaged over three runs, paying that prefix once is 46% of the whole bill . That token count does not move with pricing. The money does, and DeepSeek repriced at 2026-08-16 16:00 UTC. Under the previous flat card the same runs put that payment at 52% of the bill and the miss-to-hit ratio at 50x. Every figure on this page is given under the card in force now. Check it on your own session. Nothing is installed and nothing leaves your machine. dsh-lean is the fix: a preset that turns those tool rows off, cutting the prompt prefix by 53%. What that is worth ranges from 2% to 41% of a session bill, and the low end is real. Every number below came out of the DeepSeek API's own usage accounting, and the harness that produced them is in this repository. Measured dsh 0.1.0-rc.6, measured 2026-08-16. Thirty two runs, each starting from a clean copy of the task. The prefix reduction is deterministic. The money is not, so …

Category