Reuse the KV cache across turns #17

Do. Pin one slot per session (id_slot), keep cache_prompt, keep the prompt prefix byte-stable, and use slot save/restore (--slot-save-path) across daemon restarts. Tune --ctx-checkpoints for hybrid/SWA models (gpt-oss, Gemma 4, Qwen 3.5/3.6). Report tokens_cached as cache-read tokens.

Spec §5.2 #2.

Spec: SPEC_MUXCODE_HARNESS_AND_FEATURE_GAPS_2026_09_26.md