How is this different from an allowlist like jevgate?

An allowlist proves what may run and judges only the rest, with soft judgment never the sole veto. rh-guard instead watches live tool use for reward-hacking patterns and denies the gaming edit while allowing harmless commands.

Does it work with Exo or only the listed harnesses?

Partially. Exo is supported through ToolRuntime wrapping rather than the drop-in hooks the other harnesses use, so expect integration work instead of a one-line hook install.

Why keep Augustus and rh-guard as separate repos?

They judge different things: Augustus covers design judgment for where typed judgment belongs, while this repo is the live hazard gate on agent tools. Merging them would mix design advice with runtime enforcement.

Can this replace a scored eval harness?

No. Harbor tasksets with harness, runtime, independent validator, and evidence loops stay the scoring substrate; these hooks are structural and judgment gates inside a harness.

Which coding agents does it install into?

Claude Code, Cursor, Codex, Grok Build, Pi, Amp, Prime Agent and DeepSeek Harness, each through that agent's hook mechanism. Exo is the exception, going through ToolRuntime wrapping instead of a drop-in hook, which is the integration cost to expect before relying on it.

What behaviour is it actually watching for?

Three named patterns: grader tampering, edits to hidden tests, and interference with the eval process itself. The stated goal is to steer toward checks that cannot be gamed, which is why it pairs with an independent validator rather than trying to be one. It gates the harness around a scored taskset rather than replacing that substrate.

How does it relate to JevLint, GLiGuard and Abide?

It stays separate from them, along with jevgate: those are sidecars in their own right rather than features folded in here. The one relationship it does claim is a pairing with Augustus, where Augustus supplies design judgment about where a typed judgment belongs and this repo enforces the live hazard gate on agent tools.

CommunityJev Practices & Patterns

Rh Guard

Hook-level gate for coding agents that watches tool use for grader tampering, hidden-test edits, and eval-process interference, steering toward checks that cannot be gamed.

See highlights

Open github.com

Stars: 4
Language: TypeScript
License: MIT
Published by: Community
Author: 24601
Last verified: Oct 2026

24601/rh-guard

User repository on GitHub

View on GitHub

Reward-hack radar for coding agents: structural denies + TypeSafe Jev System One sidecar for Claude Code & Cursor hooks

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Installs into Claude Code, Cursor, Codex, Grok Build, Pi, Amp, Prime Agent, and DeepSeek Harness hooks.
Denies edits to protected assertions while allowing harmless commands, shown in the recorded side-by-side demo.
Pairs with Augustus for design judgment and stays separate from JevLint, jevgate, GLiGuard, and Abide sidecars.
Treats Harbor tasksets and independent validators as the scoring substrate rather than replacing them.

Watch out

Exo works through ToolRuntime wrapping rather than drop-in hooks, and hooks gate the harness without replacing a scored taskset. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-agent-authorization

Starter kit putting Kinde identity plus Jev judgment in front of every agent tool call, with allow, hold-for-approval, or deny decided in code and written to a ledger first.