This looks so very cool, and so then all the more thought provoking that the tests themselves are closed-source, unlike the rest of the codebase. In this evolving world of rapidly improving llm coding agent productivity, the idea that the tests are more important than the implementation starts to ring true.

I was thinking about sqlite's test landscape as described here, in relation to simonw's recent writing about porting/recreating the justHTML engine from python to js via codex, nearly "automatically" with just a prompt and light steering.