Get prompt caching working for decomposing thinker

Author: jahooma

.agents/thinker/decomposing-thinker.ts

@@ -44,7 +44,7 @@ const definition: SecretAgentDefinition = {
 
     if (prompts.length > 1) {
       // Prompt cache with a dummy thinker first!
-      const prompt = `This request was cancelled. Please respond with an empty message only. Do not write any text at all.`
+      const prompt = `We are prompt caching the above files. This is a warm-up request so that subsequent (parallel!) requests can be cached. Please respond with just the phrase: 'The context is now cached!'. It's important that you don't include any other text or reasoning in your response, since that would increase latency and cost for this request, which is just meant to warm up the cache before firing off the real requests in parallel.`
       yield {
         toolName: 'spawn_agents',
         input: {

.agents/thinker/thinker-with-files-input.ts

@@ -1,6 +1,7 @@
 import { publisher } from '../constants'
 
 import type { SecretAgentDefinition } from '../types/secret-agent-definition'
+import type { ToolMessage } from '../types/util-types'
 
 const definition: SecretAgentDefinition = {
   id: 'thinker-with-files-input',
@@ -63,9 +64,16 @@ You must be extremely concise and to the point.
       )
       const promptMessages = messageHistory.slice(0, lastAssistantMessageIndex)
       const readFilesMessages = messageHistory.slice(lastAssistantMessageIndex)
+      const readFilesToolResult = readFilesMessages[
+        readFilesMessages.length - 1
+      ] as ToolMessage
+      // For getting prompt caching to work, we need to remove the unique tool call id from the tool result.
+      delete (readFilesToolResult.content as any).toolCallId
+
       yield {
         toolName: 'set_messages',
         input: { messages: [...readFilesMessages, ...promptMessages] },
+        includeToolCall: false,
       }
     }