tpte

Usage and cost footer appended to every reply — tokens, cache hit rate, cost and context occupancy, plus a per-tool time breakdown showing which tool actually consumed the wall clock. Waiting time (a blocking user prompt) is reported separately and excluded from the active-time denominator. Pure computation, never calls a model; any failure leaves the reply untouched. Disclosure — the usage footer is appended to the stored assistant reply, so it is replayed to the model on later turns and adds tokens to every turn.