What it does

Rung addresses the problem that an AI coding agent can report a change as verified without having run it. The agent runs the real command, server or library through rung, which captures the exact output and stores it with hashes in a bundle file. A check then grades how real the verification was and who did it, the author or an independent reviewer. The gate that reads the bundle is a plain deterministic program: it uses no model and no network, and exits 0 on pass or 30 when blocked. Policies set how much independence a change needs, so higher-risk changes can require a second model or lab to review the recorded output, which is attached with a separate attest command. Rung ships no agent of its own. It includes a markdown skill that you load into an agent such as Claude Code, so the agent drives the real thing under rung as its verify step. It can be installed with pip, pipx, uv or Homebrew, run from a container image, or used as a GitHub Action. It needs no account or API key and is licensed under Apache-2.0.

Capabilities

Runs the real command, server or library and captures what it printed
Offline deterministic gate that passes or blocks a build, with no model or network
Policies that can demand an independent cross-model review
Agent skill in markdown that you load into your own coding agent
Available as a CLI, GitHub Action, container image and Homebrew package
Each verdict records which gate logic and policy produced it

When to use it

Your team merges AI-written changes and wants proof that a real run happened
You build agents and want their self-verification step to leave a re-checkable record
You need higher-risk changes to require an independent review before shipping