Have you ever wondered why sometimes Claude Code completes tasks with the current set of allowed tools, and other times it wants to execute a bash command that requires permission?
Why will it sometimes buzz along with the built in coding/editing tools but occasionally try to pipe an obscure npx command into python via tr? Is it a problem? What can we do about it?
Read more in this post by David Morgan-Brown:
https://morgan-brown.com/writing/agentic-reflective-practice