Vermahash/Orchestrator ? reverse-engineered prompt

Build me a local first autonomous agent runtime for my laptop that can take a job, break it into a typed task graph, run through plan, build, judge, test, and fix loops, and keep going until the work is actually done. It should prefer open source local models by default, manage limited VRAM carefully, and only escalate to outside models if I explicitly allow it.

I want it to work from a simple command line flow, with a daemon that can pick up jobs when the machine is idle, store progress safely in SQLite, and recover cleanly after restart. It should use real evidence like git diffs and tests to decide whether a step passed, not just trust the model saying it is finished.

If useful, include a small dashboard or chat surface for submitting jobs, plus a way to plug in tools like Python scripts or browser automation. Look up current docs online if you need to.