MahdiNavaei/aria ? reverse-engineered prompt

Build me a Python agent platform called ARIA that can plan tasks, look at screenshots or page state, decide what to do, and safely carry out actions through tools like a browser or desktop runner. I want it to feel like a real system, not a demo, with clear separation between reasoning, perception, execution, memory, safety, and event tracking.

It should support human approval for risky actions, keep a full trace of what happened, and make runs easy to replay and audit later. Use a FastAPI backend with a simple operator interface, and wire the workflow with a state based agent style setup. Include support for local LLMs, Redis style state, vector memory, and structured replay records with validation.

Please also add a small but real set of tests around the replay contract, sample docs, and a clean Docker setup so I can run it locally. If you need to check current library docs online, go ahead.