Keep planner and executor in separate sessions or you pay full price for every turn after.

DeepSeek-Reasonix is a single static CGO-free Go binary built entirely around prefix caching, and its central design choice is worth stealing: never inject planning turns into the executor's conversation, because a mid-conversation planning turn breaks the exact-token prefix match and forces a full-price cache miss on everything after. That's transferable to any cached-prefix provider. Built-in tools cover read/write/edit/multi_edit/glob/grep/ls/bash/web_fetch/todo_write behind a permission gate with sandboxed bash.