The pitch is simple and the mechanism is elegant: Sphinx sits between your AI coding agent and your remote branch, reads the diff, and asks you a question about what the code actually does under failure conditions. Get it right and the push goes through. Get it wrong and you learn something before your teammates do. The Stripe webhook example in the README is the best advertisement. The tool reads a diff where event IDs are saved before charge() is called, then asks what happens on a retry when the charge fails. The user gets it wrong — twice — and in doing so discovers a real idempotency bug they would have shipped. This is not a code review tool. It is a comprehension test, and the distinction matters. What Sphinx is actually solving is the growing gap between code velocity and code understanding in agent-assisted workflows. When Codex or Claude Code writes 5,000 lines across 11 files, the human in the loop often rubber-stamps the diff with a cursory glance. Sphinx forces a Socratic pause — not to slow you down for slowness's sake, but to confirm you could explain the change to someone who wasn't in your chat session. The design choices are disciplined. Your answers never reach the coding agent. Environment files and keys are never opened. There's a sphinx yolo escape hatch for when you genuinely need to ship fast, but every skip is logged — creating an audit trail of the moments you chose speed over understanding. The tool works with Codex, Claude Code, Cursor, OpenCode, Pi, and plain Git. Installation is one npm command plus sphinx install. The limitations are inherent to the form. Sphinx can only ask about what it sees in the diff, which means architectural-level misunderstandings — the kind where the whole approach is wrong — will sail past it. It also depends entirely on the quality of the reviewing model's questions. A weak model asks surface-level questions; a strong one probes edge cases. The tool is only as good as its interrogator. The Sun Tzu quote is doing more work than it probably should. "Outsource your thinking, not your understanding" is the entire thesis, and it's a good one. But Sphinx is a guard rail, not a solution. It cannot make you understand code you fundamentally don't have the context to evaluate. It can catch the case where you know enough to answer but weren't paying attention — which, in the age of agent-assisted development, is the most common failure mode. This is a well-scoped developer tool with a clear opinion about what's going wrong in AI-assisted workflows. It doesn't try to do too much. Whether it becomes essential depends on whether the comprehension gap it targets keeps widening — and right now, every indicator says it will.
