context-cache

Shows how full the context window is and how the prompt cache is doing.

Context ┃ Window 17% ┃ Cache hit 99% ┃ Cache TTL 4m33s

Cells

| Cell | Meaning | |---|---| | Window | Context fill. Green below 70%, orange from 70%, red from 85%. | | Cache hit | Share of the last prompt served from cache. Red below 20%. | | Cache TTL | Countdown to cache expiry, from the TTL the API reports for the entry it just wrote (5 minutes or 1 hour). expired once it ran out. |

Cache hit and Cache TTL only show once the context holds at least 20k tokens, below that a cold cache costs little. Until then, and before the first response, they read -.

Toasts

• Context crosses 70%, 85% or 95%: consider /compact.
• Cache miss: a response re-wrote 20k+ tokens at full price with a hit rate under 20%.
• Cache expires in 60 seconds: send a prompt to keep it.
• Cache expired: the next prompt re-reads the context uncached.

Behavior

• Row in the bottom-right grid, refreshed every second.
• Only the main conversation counts; a subagent has its own cache entry.