Reward Hack Guard

RH Guard (rh-guard) is a live hazard gate on coding-agent tools. It sits in hooks for Claude Code, Cursor, Codex, Grok Build, Pi, Amp, Prime Agent, and DeepSeek Harness, and blocks reward-hacking tool use—tampering with graders, hidden tests, or the eval process. Exo is support via ToolRuntime wrap, not drop-in hooks.

Structural detectors deny writes to designated evaluation assets. TypeSafe Jev scores remaining events. Soft judgment is never the sole veto.

Install

Full steps: install the hook pack · README · host matrix

# Skill only (Amp, Codex, Cursor, …). Does not start the sidecar.
npx skills add 24601/rh-guard --skill rh-guard

# Claude Code plugin (skill + hook pack). Sidecar still required.
claude plugin marketplace add 24601/rh-guard
claude plugin install rh-guard@rh-guard

Companion skill: .agents/skills/rh-guard/SKILL.md

Eval integrity

Harbor, jevals, and openevals are practices, not install dependencies. See eval integrity.

Source: 24601/rh-guard

License

MIT. Copyright (c) 2026 Basit Mustafa / 24601. See LICENSE.