tr4m0ryp/fable-5-jailbreak ? reverse-engineered prompt

Build me a Python toolkit for security research that can test how robust an AI safety classifier is when the same request gets rewritten in different ways. I want it to work both as a simple command line tool and as an MCP tool that Claude Code can auto load. The main flow should take a user query, split it into smaller pieces, optionally obfuscate the text with unicode style substitutions, wrap it in different harmless sounding framings like study guide or documentation, send those variants through a model client, then merge the results into a clear report.

Please include commands to preview the decomposition, run the full evaluation, check config status, and try just the obfuscation step. I also want a benchmark runner that can do a dry run without an API key, a quick smoke test, and save results as JSON or HTML. Add sensible defaults for things like obfuscation strength, max pieces, padding, and whether helper and multi agent modes are on. Keep it clearly positioned for authorized defense evaluation only, and look up current Claude Code MCP docs if needed.