The model often makes mistakes when using `rg`, with an error rate of about 10%. The issue is that `rg`'s handling of `-rn` is inconsistent with `grep`, and the model is more familiar with `grep`, so it frequently uses it incorrectly. In the AI era, new tools replacing old ones should seamlessly accept all inputs from the old tools, especially those that LLMs are familiar with.