Article
Jevgrep lets Coding Agent you spend less time finding code yourself:SWE-bench lowering the main model's cost 29%Dongcha Beating AI news: developer David open-sourced a research tool called Jevgrep based on Jev, specifically for Coding Agent finding code.
Users only need to ask a question such as “Where is the login validation written?” It uses Jev, a TypeSafe decision model, to search the codebase layer by layer, identify relevant files and source snippets, and then hand them to Agents such as Claude Code and Codex for further modification and testing.
It mainly addresses the code-finding stage, which consumes a lot of Coding Agent token. Jevgrep does not first feed the entire repository to the model, nor does it perform only a single semantic search. It first determines which directories are worth continuing to search, then checks relevant files and code declarations, and finally returns source snippets, line numbers, and clues for further reading. The repository also provides a Skill that tells the Agent when to call `jg` to collect context.
The latest SWE-bench experiment used 10 Python tasks. Both using and not using Jevgrep completed 8 tasks, but GPT-5.6 Sol's total cost fell from 7.62 USD to 5.44 USD, a reduction of 28.63%. The author initially wrote 40%on X, and the repository subsequently updated the experimental results; it now says “approximately 30%”.
However, this figure counts only Sol's costs and does not include Jev. The Jev invocation costs confirmed in the experiment logs already amount to at least 1.57 USD. Some invocations lack complete billing records, so the actual total is unknown. In addition, this experiment included only 10 tasks, and each question was run only once.
Original link https://m.theblockbeats.info/flash/369364