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
Show HN: BlueMouse – open-source, local Socratic firewall for AI coding | Heykuki News