Jev Eye

Jev Eye is an ultra-lean, zero-dependency supervisor for pi, the pi-coding-agent. It blocks destructive commands, prevents credential leaks, enforces test verification, routes models per turn and runs a semantic gate to catch code slop before files are written.

It blocks wipes such as rm -rf / and DROP DATABASE, force pushes to protected branches, and hardcoded secrets like sk-... keys, ghp-... tokens and SSH keys, at 0 ms and zero tokens.

It evaluates diffs of 10 lines or more through TypeSafe Jev using the has_slop question, rejecting placeholder stubs and unfulfilled TODOs before the write reaches disk. Live cache misses make synchronous network requests, so the README says it fits evaluation and targeted workflows.

Three: TypeSafe Jev, OpenRouter, or Cloudflare Workers AI (Clef and clef-flash). The active backend is picked in /jev-eye models, and the fallback is tried only after a non-OK response, a rejected envelope, a timeout or a network error; /jev-eye status counts how often it was needed.

Clef runs inside Cloudflare's free daily allowance of 10,000 neurons, so the gate costs nothing for most sessions. The TypeSafe Jev and OpenRouter backends require their own credentials, stored through the /jev-eye Accounts screen.

It warns when the agent claims a task is done after editing files without running verification suites such as npm test, cargo test or pytest, so unverified completions are flagged before you trust them.

It is a guardrail layer rather than a full reviewer: the regex gate only catches patterns it lists, and the README publishes no accuracy numbers for the semantic gate. The digest's README was also truncated mid-section, so some routing details come from metadata and topics.