The project's v6.6 status report calls the Python-script delivery a confirmed blocker on modern macOS — Gatekeeper and CORS — and concludes that the current release needs terminal skills. Take that at face value: if you are comfortable running a local Python service, the interview-then-validate loop is a genuinely different idea from another autocomplete. If you were hoping to double-click something, wait.
A quality gate for AI-assisted coding that runs entirely on your own machine. It is a standard MCP server for Cursor, Windsurf, Claude Desktop and other clients, and it is also usable on its own as a web page at `http://localhost:8001`.
- A Socratic interview stage that asks the questions people skip — pessimistic or optimistic locking for concurrent orders, roll back or retry on payment failure
- A state machine that actually blocks: `LOCKED` to `INTERVIEWING` to `GREEN` to `CODING`, and a `/generate` request before `GREEN` is refused with HTTP 403
- 17 validation layers over generated code: syntax and AST parsing, file-tree and dependency checks including circular imports, then security auditing for `eval()`, `exec()` and SQL injection, Pydantic schema alignment, and package-whitelist checks against invented dependencies
- A keyword tripwire — a request containing `DROP TABLE` or `DELETE FROM` triggers a CRITICAL STOP that asks whether this is production
- Delivery as a project ZIP: source, architecture diagrams, installation guide, cost estimate and a validation report
- Bilingual operation in Traditional Chinese and English, with the interview and reports switching language dynamically
Python 3 with `pip install -r requirements.txt`, and port 8001 free. Start it with `./Start` on macOS and Linux or `Start.bat` on Windows; the package is `blueprint-little-mouse` and the entry point is `bluemouse`. No API key is needed — the validation runs locally either way, and an Anthropic or OpenAI key only improves the AI assistance. Licensed AGPLv3, so commercial use carries obligations.