Good catch on the render-churn issue — setContext unconditionally created a new context object on every layout effect, and Immer/Zustand notified subscribers even when nothing changed. The guard logic is sound: setContext bails out via an early return inside the Immer producer when no updated key actually differs, and setCallbacks preserves 'full replacement' semantics by checking key-count first before comparing values, so omitted keys are correctly treated as changes. The EMPTY_RESPONSE_ADS constant is a nice touch to avoid a fresh {} allocation feeding into the equality check on every render.

A few things worth polishing before this lands:

• The key-set comparison in setCallbacks only checks prevKeys.length !== nextKeys.length, not actual set equality. It happens to work because any differing key pair with equal counts still produces a value mismatch on the iterated key, but it's fragile — a small comment explaining why length-check + per-key diff is sufficient would help future readers, or just do a proper key-set diff.
• Consider whether this shallow-equality pattern belongs in a shared helper (e.g. zustand/shallow or a local shallowEqual util) rather than being hand-rolled twice in this file, since other stores in cli/src/state likely have the same churn problem.
• Tests are solid and cover the four cases described in the PR body — appreciated for a first-time contributor.

Overall this is a small, well-scoped, testable fix in cli/src/state and chat.tsx, not touching any restricted paths. Worth porting after a maintainer double-checks the key-set comparison edge case.