v1.20.5

[v1.20.5] - 2026-06-16

Fixed

• Split cached prompt tokens from fresh input for OpenAI-compatible providers, so per-turn input usage no longer multi-counts the re-read cache and cost applies the cache-read rate (@yanmxa in #234)