Prompt caching with Anthropic requires the use of cache_control breakpoints. 
The cache_control breakpoint can only be inserted into the text part of a multipart message.

typescript
// Build the request using BAML's modular API
const req = await b.request.AgentLoop(
  bamlMessages,
  commandValues,
  conversationContext,
  currentDateTime,
  pdfAttachments,
  imageAttachments,
  { clientRegistry },
);

// Extract body and inject cache_control
const body = req.body.json();

// Find system message and split at my static doc marker
const systemMsgIndex = body.messages.findIndex(m => m.role === 'system');
if (systemMsgIndex !== -1) {
  const textContent = /* extract text from system msg */;
  const staticEndMarker = '</static_documentation>';
  const splitIndex = textContent.indexOf(staticEndMarker);

  if (splitIndex !== -1) {
    const staticContent = textContent.substring(0, splitIndex + staticEndMarker.length);
    const dynamicContent = textContent.substring(splitIndex + staticEndMarker.length);

    // Restructure with cache_control on static portion
    body.messages[systemMsgIndex] = {
      role: 'system',
      content: [
        { type: 'text', text: staticContent, cache_control: { type: 'ephemeral' } },
        { type: 'text', text: dynamicContent },
      ],
    };
  }
}

// Add anthropic-beta header required for caching
const headers = {
  ...req.headers,
  'anthropic-beta': 'prompt-caching-2024-07-31',
};

// Manual fetch
const response = await fetch(req.url, {
  method: req.method,
  headers,
  body: JSON.stringify(body),
});

// Parse with BAML
const decision = await b.parse.AgentLoop(responseContent);

b.AgentLoop()

traceAsync

setTags

cache_control

typescript
await b.AgentLoop(args, {
  modifyRequest: (body) => {
    // inject cache_control here
    return body;
  }
});