mycelium-labs/mycelium ? reverse-engineered prompt

Build me a Python based reliability layer for AI agents that sits between the agent and its tools and stops bad actions before they happen.

I want it to check tool calls for things like missing or invalid arguments, unsafe destinations, secret leaks, stale authority, duplicate actions, retries, loops, and runaway cost or time. It should also keep a durable record of what happened so the system can recover after a crash or an uncertain result.

Make it work cleanly for a Python app first, with a simple way to configure tools and run a local demo or starter setup. If it helps, also support a small local sidecar so other runtimes can talk to it over HTTP and JSON. Include a basic doctor and verify flow so I can inspect the setup and test failure cases.

Please keep the experience simple enough that I can add it to an existing agent project without rewriting everything, and look up current docs online if you need to.