Sphinx blocks your git push until you can explain what the agent wrote: the first understanding gate shipped as a Git hook
A Show HN on Sunday took 235 points and 77 comments for a tool that does one thing: when you push to a protected branch, Git waits while a reviewer model asks you to predict what your change does when things go wrong. Answer right and the push continues. Answer wrong and it explains what you missed and gives you a new scenario. Sphinx is by parthh01, the repo was created on October 14 and had 98 stars by Sunday night, MIT licensed, installed with npm install -g sphinx-understand and then sphinx install. The tagline on the README is a joke attributed to Sun Tzu: you can outsource your thinking, but not your understanding.
The design choices are the point. The reviewer model never sees the conversation that produced the code; it reads the committed diff through read-only Git tools, so it is quizzing you on the artifact rather than on what the agent told you it did. Questions are about behavior, failure modes and invariants, not file names. The README example is a background job queue: the worker commits the operation but crashes before acknowledging the job, so after redelivery, does the operation happen twice? The check is open-book, you can read the code or ask your agent to walk you through it while you answer. Reviews run on your existing Claude Code or Codex plan, OpenRouter, or any OpenAI-compatible endpoint, and there is no hosted service. It works whether you or the coding agent ran the push, which means an agent on autopilot gets stopped at the same gate. And for the honest case, sphinx yolo lets the push through and logs the skip, so the record of how often you bypassed your own understanding is kept.
The caveats are in the README too, which is rare for a day-one project. It sets the global core.hooksPath, so it applies to every repo on the machine until you run sphinx init per repo. A matching push blocks for up to 25 minutes, and your coding agent's shell may time out first. The changed code, your answers and a context file go to whichever reviewer provider you chose; .env files, keys and common credential paths are excluded by default, but a secret pasted into ordinary source is not caught. No native Windows.
Why this one matters more than its star count. The past month produced a stack of essays saying the same thing: Simon Willison's don't be a meat proxy, the Brain Sandwich post arguing your brain goes before and after the AI, Pensieve, you still have to think, and this week Karri Saarinen of Linear writing that each person operating a small software factory produces a product that is the average of all those outputs. Those are arguments. Sphinx is the first enforcement. It sits in the same slot as the Claude Code team's review loop and Postgres delaying a release because AI reviewers found more than humans could fix, except it gates the human instead of the code. Anthropic's own employment data says reading Claude's output is now most of a work week; a hook that makes you prove you read it is the cheapest possible version of that job. Whether people keep it installed past the first 25-minute block is the real test, and the yolo log will tell them.
Links: https://github.com/parthh01/sphinx , https://sphinx-six-theta.vercel.app/ , https://news.ycombinator.com/item?id=50137950 , https://karrisaarinen.com/posts/someone-behind-the-software/ , https://terriblesoftware.org/2026/10/02/the-brain-sandwich/
The design choices are the point. The reviewer model never sees the conversation that produced the code; it reads the committed diff through read-only Git tools, so it is quizzing you on the artifact rather than on what the agent told you it did. Questions are about behavior, failure modes and invariants, not file names. The README example is a background job queue: the worker commits the operation but crashes before acknowledging the job, so after redelivery, does the operation happen twice? The check is open-book, you can read the code or ask your agent to walk you through it while you answer. Reviews run on your existing Claude Code or Codex plan, OpenRouter, or any OpenAI-compatible endpoint, and there is no hosted service. It works whether you or the coding agent ran the push, which means an agent on autopilot gets stopped at the same gate. And for the honest case, sphinx yolo lets the push through and logs the skip, so the record of how often you bypassed your own understanding is kept.
The caveats are in the README too, which is rare for a day-one project. It sets the global core.hooksPath, so it applies to every repo on the machine until you run sphinx init per repo. A matching push blocks for up to 25 minutes, and your coding agent's shell may time out first. The changed code, your answers and a context file go to whichever reviewer provider you chose; .env files, keys and common credential paths are excluded by default, but a secret pasted into ordinary source is not caught. No native Windows.
Why this one matters more than its star count. The past month produced a stack of essays saying the same thing: Simon Willison's don't be a meat proxy, the Brain Sandwich post arguing your brain goes before and after the AI, Pensieve, you still have to think, and this week Karri Saarinen of Linear writing that each person operating a small software factory produces a product that is the average of all those outputs. Those are arguments. Sphinx is the first enforcement. It sits in the same slot as the Claude Code team's review loop and Postgres delaying a release because AI reviewers found more than humans could fix, except it gates the human instead of the code. Anthropic's own employment data says reading Claude's output is now most of a work week; a hook that makes you prove you read it is the cheapest possible version of that job. Whether people keep it installed past the first 25-minute block is the real test, and the yolo log will tell them.
Links: https://github.com/parthh01/sphinx , https://sphinx-six-theta.vercel.app/ , https://news.ycombinator.com/item?id=50137950 , https://karrisaarinen.com/posts/someone-behind-the-software/ , https://terriblesoftware.org/2026/10/02/the-brain-sandwich/
β Previous
DeepSeek sells its search engine to agents: $0.75 per thousand calls, results come with a summary written for a model, not a person
Next β
Anthropic measured its own supervision bottleneck: agent hours outnumber employee hours six to one, and 80 percent of those hours are the agent waiting for a human
Comments