battlewen/jailbreakskill ? reverse-engineered prompt

Build me a Python tool that runs a two stage red teaming workflow for testing an LLM with reusable prompt rewrite skills.

I want a command line app that can read seed prompts from benchmark files, run a first stage where it tries a library of rewrite skills and keeps track of which ones work best, then run a second stage that looks at the failed cases, groups them by category, and creates or improves new skills from those failures. It should save checkpoints and outputs in an output folder so stage two can be rerun later without starting over.

Please include support for separate model endpoints for the planner, the skill rewrites, the target model, and the guard or judge model, all configured from an env file and a YAML config. Also add a way to load and run example evolved skills, plus a simple structure for adding new skills later. If anything is unclear, look up the current docs online and make sensible defaults.