hexo-ai/sia ? reverse-engineered prompt

Build me a Python tool called SIA that lets me run a self improving loop for AI agents on benchmark tasks.

I want a simple command line app where I can choose a built in task, set how many generations to run, and then let a meta agent create and improve a target agent over time. It should save each generation in a runs folder with the generated agent code, execution logs, and a short explanation of what changed. I also want a local web page that shows the progress of a run in real time, including the code, logs, improvement notes, and score changes across generations.

Please make it easy to swap models and providers through config profiles, and support running with different agent backends if needed. It should work with the built in tasks mentioned in the README, and also allow pointing at a custom task directory. If anything in the current docs or examples is unclear, please look up the latest docs online and match the intended behavior closely.