cloudflare/ai workers-ai-provider@3.1.5

on GitHub

Patch Changes

• #451 2a62e23 Thanks @mchenco! - Fix reasoning content being concatenated into assistant message content in multi-turn conversations

  Previously, reasoning parts in assistant messages were concatenated into the content string when building message history. This caused models like kimi-k2.5 and deepseek-r1 to receive their own internal reasoning as if it were spoken text, corrupting the conversation history and resulting in empty text responses or leaked special tokens on subsequent turns.

  Reasoning parts are now sent as the reasoning field on the assistant message object, which is the field name vLLM expects on input for reasoning models (kimi-k2.5, glm-4.7-flash).