🤖 fix: implement proper Anthropic cache strategy for system prompts and tools

[ibetitsmike]

Summary

This PR fixes the Anthropic cache strategy to properly cache system prompts, tool definitions, and conversation messages according to the Vercel AI SDK documentation.

Problem

The current implementation only cached conversation messages but missed two critical components:

• System prompts (which rarely change and should always be cached)
• Tool definitions (which are static per model and should always be cached)

Solution

1. Enhanced Cache Strategy Module

• Added createCachedSystemMessage() to convert system prompts to cached messages
• Added applyCacheControlToTools() to cache tool definitions
• Added supportsAnthropicCache() helper for cleaner code

2. StreamManager Integration

• System messages are now converted to cached messages at the start of the message array
• Tool definitions are cached when passed to streamText()
• Messages continue to be cached up to the second-to-last message

3. Comprehensive Testing

• Added unit tests for all cache strategy functions (13 tests, all passing)
• Added integration tests to verify end-to-end functionality

Impact

This change significantly reduces token usage and costs when using Anthropic models by:

• System prompts: Cached across all conversations (rarely change)
• Tool definitions: Cached for all tool calls (static schemas)
• Message history: Cached up to current message (reused context)

Implementation Details

Uses the proper providerOptions structure as documented:

providerOptions: {
  anthropic: {
    cacheControl: {
      type: "ephemeral"
    }
  }
}

Testing

• ✅ Unit tests: bun test src/common/utils/ai/cacheStrategy.test.ts
• ✅ Type checking: make typecheck
• ✅ Build: make build

Generated with mux