GitHub radar
Foreman: autonomous overseer for Codex coding agents
Foreman runs alongside coding agents and independently assesses whether the work is done, the agent is stuck, or human input is needed — then acts on that automatically.
Foreman is a Python runtime that runs alongside Codex coding agents and independently assesses the state of a software job while the agent works. Every 30 seconds it examines the git diff, agent output, and step history, then sends a fast classifier nine yes/no probability questions: is implementation complete, are tests sufficient, is the agent stuck, is work drifting off track, does a human need to step in. Based on those probabilities, a deterministic Python policy decides to continue, steer, stop, retry, launch a verifier, or escalate. The authors call this an architectural experiment, not a production tool — assessment accuracy still needs calibration. Requires Codex CLI and a TypeSafe API key. A fully offline demo needs neither: `foreman demo --repo .`
Why a vibe-coder should care
If you run coding agents on longer tasks and step away, you often come back to a loop or a half-finished result. Foreman adds an independent eye on the work — one that can stop a stuck agent before it wastes more tokens or does something wrong.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Deploy Foreman from https://github.com/thruwire/foreman — follow the README, start with the offline demo `foreman demo --repo .` (no keys needed); ask me for TypeSafe API and Codex credentials if you want to run a real job
Runs on any laptop — needs Python, Codex CLI, and a TypeSafe API key
Open on GitHub▌ More finds