Jev cannot write code. TypeSafe says so in the first line of their coding-agent docs. Yet in ten days hundreds of repos have shipped on it, and most of them are plugins for Claude Code, Codex, Cursor and OpenCode. The reason: most of what a coding agent does is not writing. It's deciding. Which file? Is this command safe? Which tests to rerun? Is this tool output still needed? Which model for this turn? Is it actually done? Each of those currently costs a full frontier round trip, seconds and roughly 100k tokens of context. Jev answers each in 100 to 250ms for a thousandth of a cent. So people are moving the decisions and leaving the LLM the writing. Permission prompts: A regex blocklist catches sudo, rm -r and git push. Then Jev answers two questions: does every command stay inside the project and undo with git, and is it a reasonable next step toward the task? Both above 0.9 and it runs. Otherwise you get asked. $0.0000168 per check Model routing: jev-router sits in front of the claude CLI, reads the first message of a turn and picks Haiku, Sonnet, Opus or Fable. Continuations add zero latency. Test selection: jev-test-filter scores every test against the diff and emits the flags vitest, jest, go test or cargo already understand. Real repo, 110 tests, one commit: 1.3 seconds, $0.001, 3 tests chosen. Uncertain means run everything. Seeing the above examples, use Jev wherever an LLM is currently being used as a classifier. That is wasteful and Jev fixes it.