v0.5.1 — governance as a build step

@metamynd/agentsafe-guard@0.5.1 and @metamynd/agentsafe-mcp-guard@0.1.2 are on npm, MIT, zero dependencies.

New in 0.5.x

agentsafe-guard verify — assert in CI that an agent cannot exceed its mandate, and fail the build if it can.

npx @metamynd/agentsafe-guard verify --require merchants,perTxn

The decision that matters is not which controls it checks. It is that a control the mandate does not set is reported, never passed — held / not-configured / failed, and only the first is a pass. That exists because the alternative shipped: a public sandbox issued with an empty merchant list paid an unapproved supplier, while MERCHANT_NOT_ALLOWED sat documented the whole time. A green check on a control that does not exist is worse than no check.

Evaluation is local and pure — no holds minted, no nonces spent, one GET per run.

Fixed in 0.5.1

verify printed "Every configured control held" and then returned 127, because process.exit() fired while the fetch handle was still closing — a libuv assertion on Windows. It failed the build for a passing agent, which is the one direction a check must never fail in.

Verify it yourself

npm pack @metamynd/agentsafe-guard@0.5.1

The evaluator ships as unminified JavaScript. The protocol is published in full at https://metamynd.ai/developers/spec.