v1.0.3

What's Changed

• Fix system prompt caching for string format by @roblesjm in #4 #5

Changes

Bug Fixes

System Prompt Caching for String Format (#4, #5)

Fixed issue where system prompts using the string format were not being cached, even when exceeding the 1024-token minimum.

Root Cause:
The Anthropic API supports two formats for system prompts:

1. String format: "system": "text here"
2. Blocks format: "system": [{"type": "text", "text": "..."}]

Cache control can only be applied to the blocks format. Previously, autocache only handled prompts already in blocks format, so system strings were never cached regardless of size.

Solution:

• Convert system strings to SystemBlocks when cacheable (>= min tokens)
• Add cache_control to converted blocks
• Serialize SystemBlocks as "system" array in JSON output

Technical Details:

• Add MarshalJSON to output SystemBlocks as "system" array
• Add UnmarshalJSON to accept both string and array formats
• Ensures compatibility with all client formats

Testing

• All 81 tests pass
• End-to-end test confirms caching works with real API
• First request: cache_creation_input_tokens = 1201
• Second request: cache_read_input_tokens = 1201

Files Changed

• internal/cache/injector.go: Convert system strings to blocks
• internal/types/models.go: Add Marshal/UnmarshalJSON methods
• internal/cache/cache_injector_test.go: Add conversion tests
• internal/types/models_test.go: Add JSON handling tests
• internal/server/handler_test.go: Update mock server

Full Changelog: v1.0.2...v1.0.3