Community · E2 · artifact verified

Run Jev judgments from the shell before an agent acts

jev-axi is a CLI that puts a half-second Jev opinion in front of every command an agent runs: pick, rate, check, rank, triage, and guard, plus a PreToolUse hook that blocks risky Bash calls before they execute.

01 · Role in the system

What Jev does here

The safety hook receives a tool name and its arguments, and one request asks Jev for hazard probabilities - destructive, exfiltration, remote code execution, weakened security, outside-project - plus a 0-2 risk score, returning deny or allow with a written reason. Routine commands are decided by a local list and cost nothing; only the rest pays for a call. The same judgment screens text an agent fetched, where the danger is instructions hidden in the page rather than the command: the README shows a real run flagging a buried curl-pipe-sh with injection at 0.99, decided in 375ms for four thousandths of a cent. Every transcript in the README is regenerated from the same commands by scripts in the repository.

02 · Control boundary

Where Jev sits

Tool calls and fetched text become fitted state for hazard Nouls and a risk score; local rules short-circuit routine input, and only the rest reaches Jev.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • A second opinion on text, not a sandbox: it judges what a command says, and the local routine-command short-circuit is a heuristic list.
  • Latency and cost figures are author-reported transcripts from the author's own runs.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.