FREE · MIT · NODE.JS 18+

Does your repeat-call guard actually fire?

Agent Run Guard applies repeat-call and budget policies at configured coding-agent hooks. Start with a small, reproducible check: allow two identical synthetic calls, then deny the third.

Get the free package Configure your harness

1. Install the released engine

npm i -g agent-run-guard@0.2.1

Installation makes the CLI available. It does not activate agent hooks. The core has no third-party runtime dependencies and makes no network calls; installing from npm requires a network connection.

2. Reproduce the policy check

Download verify-guard.mjs into an empty folder, inspect it, and run this command from that folder in PowerShell:

node verify-guard.mjs "$(npm root -g)/agent-run-guard/bin/guard.js"

The script sends synthetic hook JSON to the installed CLI. It does not execute the represented file-read operation. Configuration and hash-keyed state are written to a new temporary folder; its location is printed.

Call 1: ALLOW
Call 2: ALLOW
Call 3: DENY
PASS: synthetic CLI mapping, not live-host enforcement.

This output was reproduced on Windows with the npm 0.2.1 artifact on 2026-09-20. The script exits with an error if the expected decisions differ.

3. Check the actual host before relying on it

A CLI test is not proof that your agent invokes its hooks. Configure the correct adapter and verify a harmless allow/block case inside your host. Check its version and the published proof levels.

OpenCode V1: recorded live proof on 1.18.31.
OpenCode V2: unit/dry-run coverage with host limitations; no blanket live-enforcement claim.
Claude Code and other adapters: simulated coverage as documented in the README. The demonstration above does not upgrade that proof level.
Internal errors fail open. This tool is not a security boundary or a guarantee against loss. Dangerous-command and secret-pattern checks default to warnings; repeat and call-budget policies default to deny.

Already have a run to investigate?

Agent Run Inspector is a separate paid, offline diagnostic for OpenCode exports. It reports what happened afterward. It is not required to use Guard.

Open free synthetic Inspector reports · Inspector individual — USD 49 once · Team — USD 149 once

Configuration problem? Report a reproducible issue with your OS, agent version, and a synthetic example. Do not post credentials, private exports, or customer data. No response-time promise.