applyPromptCache

function applyPromptCache(
   model,
   sessionId,
   messages): object;

Single wiring point for provider prompt caching, called by the driver streamText sites. Gated by the conservative detectors so non-matching providers are untouched:

• Anthropic → applies cache_control breakpoints (caches the system+tools prefix + recent history; ~75% input-cost + TTFT off multi-turn turns).
• OpenAI Responses → sets promptCacheKey (= sessionId, pins same-session turns to one cache slot) + truncation: 'auto' overflow safety net. Returns the (possibly transformed) messages + an optional providerOptions bag ready to spread into streamText. Default-on: both are no-ops on providers that ignore the fields.

Parameters

Parameter	Type
model	unknown
sessionId	string
messages	ModelMessage[]

Returns

object

messages

messages: ModelMessage[];

providerOptions?

optional providerOptions?: Record<string, Record<string, JSONValue>>;