knightfolk/open-harness — reverse-engineered prompt

Build me a local desktop app called OpenHarness for people who use a bunch of AI models and want to actually see what is happening behind the scenes. I want one workspace where I can chat, connect different providers, pick which models are available, assign different models to jobs like planner, coder, reviewer, reasoner, summarizer, worker, and title maker, and let an auto router choose the best model for each task.

Make the routing visible instead of hidden. I should be able to watch active runs, tool calls, retries, failures, fallback choices, and the proof or evidence behind the final answer. Add a provider hub for testing connections, checking health, budgets, and rate limits, plus MCP tool support and browser screenshot evidence that can be turned into text for non vision models.

It should support a durable goal command like /goal so longer objectives stay active across turns. Keep state local on my machine, and make it work as a desktop app with a web dev mode too. Look up current docs online if you need to.