fix(anthropic): resolve issue with tool result caching

[elliottlawson]

Summary

• Fixes issue where tool_result_cache_type was being applied to each tool result message instead of only the last one
• Moves caching logic from message creation to final mapping stage
• Ensures only the last tool result across all messages gets cached

Changes

• Remove cache application in Text::handleToolCalls() and Stream::addMessagesToRequest()
• Add logic to MessageMap::map() to identify and cache only the last tool result message
• Add comprehensive test coverage for the caching behavior

Test plan

• Added new test file ToolResultCachingTest.php with 3 test cases

[kinsta]

Preview deployments for prism ⚡️

Status	Branch preview	Commit preview
❌ Failed to deploy	N/A	N/A

Commit: 5f53879d1f18be74cfe96e92390adb9aef8bd02b

Deployment ID: 485cd4cf-5586-4b8a-a1fb-eaae2154b561

Static site name: prism-97nz9

[sixlive]

@ChrisB-TL can you take a look at this?

[elliottlawson]

I've been running this patch for the past week with pretty good results. But it wouldn't hurt for someone to do a once-over.

[sixlive]

@ChrisB-TL can you take a look at this?

[ChrisB-TL]

Will this not result in missing the last cache on subsequent calls, as the breakpoint is no longer there? I.e. you'll miss the cache until the last call, which will undermine the point of it?

Anthropic does have a limit of four cache break points, so the current implementation will break eventually for chains with multiple tool calls.

[ChrisB-TL]

Will this not result in missing the last cache on subsequent calls, as the breakpoint is no longer there? I.e. you'll miss the cache until the last call, which will undermine the point of it?

Anthropic does have a limit of four cache break points, so the current implementation will break eventually for chains with multiple tool calls.

Actually, possibly not. From the docs:

If any of these previous positions match cached content from earlier requests, the system uses the longest matching prefix
This means you don’t need multiple breakpoints just to enable caching - one at the end is sufficient

Would you be able to test out with real requests to confirm, and add a test that uses recorded fixtures?