lzw12w/swarm-harness-demo — reverse-engineered prompt

Build me a Python demo framework called Swarm of Thought that uses multiple AI agents to solve coding tasks by splitting big requests into smaller subtasks, solving those pieces, then merging the results into one final Python file.

I want it to feel usable from the terminal. Include a normal CLI where I can run a task, compare a single agent answer against the swarm approach, inspect the last task tree, and view logs. Also make a simple rich terminal UI so I can watch the split, solve, merge flow happen live.

It should work with OpenAI compatible LLM APIs using env variables for the key, base URL, and model. Please include retries, timeouts, limited concurrency, and save the task tree plus event log to a JSON state file so I can review it later.

Let the agents optionally use safe basic tools like running Python, reading and writing files, and searching project files. Add tests for the contracts, memory persistence, tools, CLI parsing, and orchestration.