Cache write tokens now count on OpenAI and LiteLLM

Agno's OpenAIChat fills MessageMetrics.cache_write_tokens from usage.prompt_tokens_details.cache_write_tokens, and OpenAIResponses from usage.input_tokens_details.cache_write_tokens. Both already read cached_tokens from the same object into cache_read_tokens, so a run now shows both sides of the prompt cache.

The LiteLLM model reads prompt_tokens_details.cache_write_tokens and falls back to cache_creation_tokens, the mirror LiteLLM also fills. That covers Claude and any other cache-writing provider behind LiteLLM, whether the usage arrives as a dict or as LiteLLM's Usage object. A missing or null field still counts as zero, and the value adds up into run and session metrics like every other token count.

We passed each parser a usage payload built with the provider SDK's own types: 3,000 written tokens for the two OpenAI APIs, and LiteLLM's Usage with cache_creation_tokens=4000.

# Before (real output)
OpenAIChat       cache_read_tokens=12000  cache_write_tokens=0
OpenAIResponses  cache_read_tokens=12000  cache_write_tokens=0
LiteLLM          cache_read_tokens=8000   cache_write_tokens=0
 
# Now (real output)
OpenAIChat       cache_read_tokens=12000  cache_write_tokens=3000
OpenAIResponses  cache_read_tokens=12000  cache_write_tokens=3000
LiteLLM          cache_read_tokens=8000   cache_write_tokens=4000

Learn more about cache and reasoning token metrics and LiteLLM in the documentation.