The tool system breakdown is useful. I've been running Claude Code daily for a few months and some things in there explain behaviors I noticed but couldn't figure out. Like why sometimes it re-reads files it already read 30 seconds ago. Makes more sense now seeing how the context window management works internally, it's being more conservative about what it assumes is still accurate. The permission gate flow is the part I care about most. When I'm away from my machine and the agent hits a permission prompt, that's where my workflow falls apart. Knowing how it structures those requests internally is helpful for anyone building tooling around it. One thing the guide doesn't cover much is how sessions are meant to be resumed. The checkpoint system is interesting but in practice I find session resume pretty fragile once you go past a certain context length.