v0.7.68

Messages surface for Claude Code (Harbor / Akhara AI asks): every usage object (message_start, message_delta, non-stream) carries Anthropic's four legs including cache_read_input_tokens / cache_creation_input_tokens from the provider's cache report; the pre-dispatch estimate on message_start counts system, every turn and tool definitions the way count_tokens does; the cache_control disclosure reads not_forwarded(provider_caches_automatically; cache reads reported in usage.cache_read_input_tokens) (#920). A bare thinking: {type: "enabled"} resolves to the lane's default reasoning effort instead of a fixed medium, budget_tokens >= max_tokens is refused like Anthropic does, a tiny max_tokens on a default-reasoning lane no longer returns an empty reply, and an image on a text-only route is refused as image_input on messages instead of blaming thinking (#921). Carries exp-gateway-native 0.3.52. Evidence: the adversarial production probe of 2026-09-12 and the platform ledger for org Akhara AI.