What it does: • 2PC (Two-Phase Commit) for AI agent operations • Risk-based approval: LOW (auto) / MEDIUM (verify) / HIGH (human) • Sandboxed execution (no shell injection) • Fail-closed: when in doubt, ask a human

100% on-premise, Apache 2.0, Python 3.10+

Would love feedback from the HN community on the architecture and security model.