Jev Watch — AI skill for Claude Code

jev-watch

[](https://github.com/polaminggkub-debug/jev-watch/actions/workflows/ci.yml)  

**jev-watch is a watchdog for AI coding agents.** It runs Codex, OpenCode or any coding CLI in the background, notices when the agent is **looping on the same error, stalled, or drifting off task**, stops it, and hands a short report to your lead agent (Claude Code, or you). The lead writes a correction and jev-watch **resumes the same Codex/OpenCode session**.

It is built for the "one smart orchestrator, cheaper workers" setup: Claude Code plans and reviews, Codex or OpenCode writes the code. Without a watchdog the orchestrator sits in a `sleep 60; tail log` loop, paying premium tokens to read logs that say "still working", or it only finds out 30 minutes later that the worker spent the whole time fixing the same import.

jev-watch checks every 45 seconds with **free local rules** and one question to [Jev](https://typesafe.ai), TypeSafe's fast decision model (sub-second, a fraction of a cent, available through OpenRouter). Your orchestrator stays idle and costs nothing until something actually needs a decision.

lead agent ──► jev-watch -- codex exec "fix the login page"
                  │  every 45 s: local rules + one Jev question
                  │  healthy  → keep going, lead stays asleep
                  │  problem  → stop worker, print a short report, exit 2
                  ▼
lead reads the report, decides what went wrong
                  ▼
jev-watch --resume  "you are editing the wrong store, fix X in Y"
                  │  same Codex/OpenCode session, watched again

Install

npm install -g github:polaminggkub-debug/jev-watch

Node 20 or newer. Zero runtime dependencies. macOS and Linux.

Set one key: `JEV_WA