PerryLink/jevcore ? reverse-engineered prompt

Build me a TypeScript project that adds a simple decision layer for an AI agent, not a chat bot. I want it to answer typed questions like yes or no, choose between options, or give a score, and return calibrated probabilities instead of prose. It should work in a plain Node app, as a DeepSeek Harness plugin, and as an MCP server over stdio.

Please make it offline by default, with a mock provider that never makes network calls unless I explicitly turn on a live provider and give it credentials. I also want clear startup logs that say what is off and what data would be sent if something is enabled. Add basic redaction for obvious secrets before anything leaves the process, and make sure disabled safety checks really do nothing.

If you need current docs for the harness or MCP, look them up online first. Please keep the public surface small and honest, with the core logic separate from the adapters.