Tool Suggestions
Sphinx: The Git Hook That Quizzes You on Your AI's Code Before You Push
Every AI code review tool reviews the code. Sphinx reviews the developer. It is a Git pre-push hook that, when you try to push to a protected branch, stops the push and asks you scenario questions about the diff you are shipping: what the change does, where its edge cases are, and what happens when something fails. Answer correctly and the push proceeds. Answer wrong and Sphinx explains the gap and gives you another scenario. Its repository states the premise in one line — "You can outsource your thinking, but not your understanding" — and its Show HN launch on October 18 drew 225 points and 76 comments, with the underlying anxiety confirmed a day later by an Ask HN thread on shipping AI-written code you do not understand that passed 590 points and 700 comments.
How it works
The demo exchange shows the design. Your webhook handler now saves the event ID before calling charge(). Stripe times out and sends the same event again. What happens on the retry? Work through it and the answer is uncomfortable: the saved ID makes the retry look already-processed, so the customer is never billed. That is the class of question Sphinx asks — not "what does this function do" but "what breaks, and who pays." Mechanically, a separate reviewer model reads your diff through read-only Git tools, deliberately without your chat history, and generates one to three questions scaled to the size of the change (the project's own analysis describes adaptive budgets of up to five). It runs on minimal tokens, from your existing coding agent or any OpenAI-compatible endpoint, including OpenRouter. It works with Claude Code, Codex, Cursor, OpenCode, Pi or plain Git. Your answers never reach the coding agent, and Sphinx does not open .env files or keys.
Installation is an npm package — sphinx-understand, requiring Node 22.19 or newer, on macOS, Linux or WSL — and the license is MIT. The project is days old: first commit October 14, a solo maintainer, and roughly 200 GitHub stars within its first week. There is an escape hatch, and it is the most honest feature in the design: sphinx yolo lets a push through without the quiz — and logs every skip. The project's README is equally plain about what it is: "practice, not a security gate."
Why developers are paying attention
The bottleneck in software delivery has moved. Agents can now open multiple files, run commands and produce complete pull requests; what has not scaled is a human's verified understanding of what was produced. A diff can be idiomatic, pass its tests, and still encode a wrong assumption about retries, authorization or failure handling — the failure surfaces during an incident, which is the most expensive classroom. Sphinx's architectural choice is placement: the question arrives at the delivery stage, when the patch is concrete and about to become part of the system, rather than as an optional reflection inside the coding session. It also inverts the usual economics of AI review tooling — instead of paying for more machine output added to a queue nobody reads, it spends a few of your tokens to test the one component that will be paged at 3 a.m.
The honest problems
Friction is the obvious one, and the HN thread named it precisely: one commenter built a similar tool, got annoyed at being questioned by it, and stopped using it. A quiz on every production push will be loved in principle and resented in practice, and a gate that can be bypassed with a logged yolo is a team norm, not an enforcement layer — it will catch the conscientious and be routed around by everyone else. The sharpest suggestion in the discussion was to move the gate earlier in the social pipeline: quiz before a human is asked to review, not before production, so the check prevents wasted reviewer attention rather than sitting at the last step. And a quiz is only as good as its questions; a reviewer model that misreads a diff will teach wrong lessons with great confidence. Sphinx mitigates none of these fully. What it does is make unexamined understanding harder to disguise as review — which, this week, several hundred developers agreed is a real problem worth a small daily annoyance.
The takeaway
Sphinx is free, MIT-licensed, cheap to run and honest about its limits — a rare combination. Teams shipping agent-written code to production branches they protect are exactly its audience, and the trial cost is one npm install. Treat it as what it says it is: practice that builds the understanding your agents cannot give you, with a log that shows how often you skipped class.