The _parseOpenAIUsageFromSSE method was not capturing cache-related
tokens (cache_read_input_tokens, cache_creation_input_tokens) from
OpenAI format responses, while the Anthropic endpoint correctly
captured them.
This fix adds extraction of:
- cached_tokens from input_tokens_details
- cache_creation_input_tokens from both input_tokens_details and
top-level usage object
This ensures proper cache statistics tracking and cost calculation
for OpenAI models (like GPT-5/Codex) when using the Droid provider.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>