Jev in production › Agent guardrails and approvals
Jev triages Claude Code requests, sets effort level, blocks destructive commands and checks task completion; $0.07 over six days of use (author).
Claude says "done". Jev checks the diff. ClauJev is a Claude Code plugin that puts a second model, from another company, on Claude's work: Jev, TypeSafe AI's judgment model. Before the reply reaches you, Jev compares the real git diff with what Claude says it did. If something is half done, Claude goes back and finishes.
- โ Final check by an independent model: not Claude grading Claude. It reads the turn's diff (shell edits included), the commands and the final reply, and stops "done!" when the work looks unfinished or your tests fail. - ๐๏ธ Effort per request: Jev estimates the size and depth of each request and asks whether you want low, high or xhigh before Claude starts. You decide. - ๐ It measures itself: report shows how many times the...
For the project's own README, linking back here: