Hi HN, I built a CLI tool that acts as a middleware between me and LLMs. It forces a "Socratic" planning phase before allowing code generation. The problem: I found "Vibe Coding" (prompting until it works) was making my architecture sloppy. The solution: A 17-layer validation script and a prompt chain that asks modifying questions (e.g. "How do you handle JWT revocation?") before writing code. Stack: Python, Local MCP Server. License: AGPLv3. Repo: https://github.com/peijun1700/bluemouse