Agent Cache Meter
Prompt-Cache Meter

Challenge
Every request sends the whole conversation again. The prompt cache makes that cheap, since cached tokens cost a tenth of normal input, but the cache lives only an hour after the last request (five minutes on an API key).
Coming back after it expires costs one full rebuild. In a real 487k-token session, that first request cost $3.63, the price of 30 normal requests. A large context also costs on every request, even with a warm cache. None of this was visible while working, so there was no way to tell when to continue and when to compact.
Results
• One-line meter: sits above the composer and redraws every second, with read, write and uncached tokens, the hit rate and an expiry countdown.
• Four clear states: warm, expires soon, cold and reset by /compact, each with its own badge and next step.
• Continue or compact: on a cold cache, the meter compares both paths and recommends /compact only when it pays for itself.
• /cache panel: per-turn reads, writes, hit rate and cost, what the cache saved this session, plus /cache settings.
• Three parts: an omp extension, a Claude Code status line on npm (claude-cache-meter) and a Tern plugin with tab countdowns, expiry toasts and a canvas listing every agent.














