What are cached tokens in /usage and /cost?

In an agent session, every step sends the conversation so far, and most of it is the same as the previous step. Line keeps that history append-only: earlier messages are sent exactly as before, so the repeated part can be served from a cache. Those are cached tokens.

Line reports them in /usage, /cost, the print-mode usage line and JSON output (cached_tokens, cache_hit). Old tool output is cleared only during compaction, in one batch, to keep the cache effective.