If compression has some black-tech—for example, Codex’s private cloud compression model—within a single conversation, it keeps compressing continuously without opening new threads. The agent can continuously obtain the entire repository’s content, memory, etc. You don’t need to keep telling it to scan the repository every time, which is just useless chatter, and it also saves tokens and is more convenient to operate. This may also be why Codex previously could cancel the context display. It wasn’t mature back in version 5.5; but now with 5.6, I feel it’s excellent.