exoharness/exo ? reverse-engineered prompt

Build me a local AI agent harness called Exo that can run as a long lived assistant from the command line and also through a simple browser chat. I want the agent to use a Docker sandbox for tools and experiments, keep a durable event log of messages and tool calls, and let me reconnect later without losing state. It should support basic commands like starting, stopping, listing conversations, and starting fresh.

The main goal is that the agent can safely improve itself over time, including editing its own prompts, tools, memory, and other parts of its setup, while keeping an immutable history so it does not get stuck in loops. Please include a setup script that makes first time install easy, and a small web interface for chat access from anywhere. If helpful, look up current docs online as you build it. Use a clean, practical default experience first, then make the self improvement and recursion parts work in a safe, incremental way.