This seems cool! I’ve had success giving Claude Code the ability to reflect on runtime performance by making it write rust perf benchmarks and then reflect on the results; this seems like a nice generalization