# swarm-verify

A command-line tool that independently reruns a repository's own test commands after an AI coding change, then reports what actually ran and what remains unverified.

swarm-verify checks what an AI-written code change actually did, rather than trusting a model's own claim that it finished the work. Run inside a repository whose dependencies are already installed, it discovers the project's declared test command from its manifests, runs it the way a continuous-integration job would, and reports five things separately: whether the command ran, what the checks found, how the run was contained, whether any requirement was actually judged, and whether any check was challenged. It writes nothing into the repository and needs no model, API key, or extra configuration.

A regression-only pass is reported as exactly that, a pass on the repository's own existing checks, not proof that the requested work was completed; only a separate requirement contract can establish that. It also ships as a GitHub Action that fetches a pull request's head and base commits into an isolated, network-disabled checkout, signs the verdict as a GitHub artifact attestation, and posts a single comment on the pull request describing what was measured.

It requires Node 22 or newer and git, plus uv for a Python project, and full command execution is supported on Linux and macOS, with bundle verification only on Windows.

Website: https://github.com/moonrunnerkc/swarm-orchestrator