XAGI-Lab/melra ? reverse-engineered prompt

Build me a local first autonomy kernel called MELRA that sits between an LLM and the real world, and safely runs approved actions with durable recovery.

I want it to manage a loop of tasks and workflows, keep local state in SQLite, survive crashes and restarts, and prove what happened instead of just saying it succeeded. It should support a small MCP style tool surface, stdio and local HTTP, with a simple read only console for viewing runs, approvals, receipts, and evidence. Make it so the system can grant limited capabilities instead of handing out raw credentials, and make sure every important action is logged, resumable, and verifiable.

Please include a clean TypeScript implementation, a Python SDK if that fits the repo, Docker support, and some realistic examples that show a model asking for an effect, the kernel approving it, running it, and confirming the result. Keep the default setup safe, local, and narrow in scope. Look up current docs online if you need to.