CCBench: How do agents perform on codebases that aren't part of training data? | Dark Hacker News