🐛 Patch Changes
- 65bca08: Fix token-cost calculation for providers that report cache-read prompt tokens. Manifest now uses models.dev cache-read/cache-write prices when available, so cached DeepSeek input tokens are billed at the cache-hit rate instead of the full input-token rate.
- 08f9c6f: Fix
/v1/responsesstreaming to emit the full OpenAI Responses API event lifecycle when bridging a Chat Completions upstream. The converter now opens a message item and content part (response.output_item.added/response.content_part.added) before the text deltas and closes them (response.output_text.done/response.content_part.done/response.output_item.done) with a populatedresponse.completed, instead of emitting bareoutput_text.delta+response.completed{output:[]}. Strict Responses API clients (Pi, OpenClaw-style) previously dropped the deltas and rendered empty assistant messages.