v0.1.0

First release of the Agent Security Contract: a fail-closed gate that checks agent and MCP tool calls against a pinned capability policy before they run.

What's in it:

• YAML capability policy with a pinned capability digest; per-agent tool allowlists, argument constraints, high-impact approval and rate limits
• Evaluator returning ALLOW, BLOCK or REQUIRE_APPROVAL; a malformed policy, digest mismatch, unknown tool, evaluator error or sidecar timeout all give BLOCK
• asc check and asc verify; evidence envelopes with policy SHA-256, trace ID, commit SHA and an integrity digest
• HTTP sidecar plus a fail-closed client, and a reusable GitHub Action (justrunme/agent-security-contract@v0.1.0)

On this corpus of 28 scenarios (14 attack, 14 benign): critical false-ALLOW 0, benign false-BLOCK 0/14. That covers this corpus only, not agent safety in general.

Known limits: it evaluates traces against a declared policy; it isn't a sandbox, a prompt-injection detector or a full ACS Guardian. Envelopes have an integrity digest but aren't signed yet.