ai.CacheDelimiter

Signature

A position in a prompt: everything before it is the prefix a provider may cache. ${cache()} in an LLM function's prompt: puts one where it is written. A client that has a cache marker on its wire sends one there; every other client drops the delimiter and sends the prompt as if it were not there.

${cache()} leaves args null, and the client that sends the request fills in its own default_cache_args(): the marker its wire takes, or nothing when its model would reject one. So one ${cache()} is right on every client, including each member of a fallback.

${cache(args)} sets args, which rides along untouched and is sent exactly as written. It is the provider's own cache object, so it is written for one provider's wire:

${cache({ "type": "default", "ttl": "1h" })}   // aws.BedrockClient: a `cachePoint` block
${cache({ "type": "ephemeral", "ttl": "1h" })} // anthropic.Client: `cache_control` on the block before
${cache({ "mode": "explicit" })}               // openai.ChatClient, openai.ResponsesClient:
                                               // `prompt_cache_breakpoint` on the part before

${cache(null)} is ${cache()}. To mark a position on some calls only, put the ${cache()} inside an ${if (...)}.

Outside an LLM function's prompt: there is no bare cache; interpolate the value itself, ${ai.CacheDelimiter { args: null }}.

Source:<builtin>/ai/spec.bamlbytes 3016–3066