Context Caching Disk Kv Cache

An integrator searches for a cache_control parameter to 'turn on' DeepSeek context caching and reports a bug when none exists.

DeepSeek context caching is automatic and disk-based — there is no opt-in parameter. Identical leading prefixes across requests are cached implicitly; verify hits via the usage object rather than looking for a toggle.

Cost telemetry sums prompt_tokens and ignores prompt_cache_hit_tokens / prompt_cache_miss_tokens, so cached requests are billed at the full input rate.

Read usage.prompt_cache_hit_tokens and usage.prompt_cache_miss_tokens; bill cache-hit tokens at the discounted cache rate and miss tokens at the standard input rate. Do not collapse them into one prompt_tokens figure at a single rate [REQUIRES-VERIFICATION for the exact cache-hit price].

A reusable system prompt and a per-request user question are concatenated, but the per-request question is placed before the stable system text in messages[].

Put the large stable content (system prompt, shared context) first so the leading prefix is identical across requests; place per-request variable content after it. Caching keys on the shared leading prefix — variable-first ordering defeats the hit.