cloudflare/ai @cloudflare/tanstack-ai@0.1.8

on GitHub

Patch Changes

• #505 f43f6f0 Thanks @threepointone! - Add passthrough for reasoning_effort and chat_template_kwargs in createWorkersAiChat. Pass them per-call through modelOptions:

  const adapter = createWorkersAiChat("@cf/zai-org/glm-4.7-flash", {
  	binding: env.AI,
  });
  
  chat({
  	adapter,
  	messages,
  	modelOptions: {
  		reasoning_effort: "low",
  		chat_template_kwargs: { enable_thinking: false },
  	},
  });
  

  Previously these fields were silently dropped, which could cause reasoning models (GLM-4.7-flash, Kimi K2.5/K2.6, GPT-OSS) to burn the entire output token budget on chain-of-thought with no visible content. They now reach binding.run(model, inputs) at the inputs level as required by Workers AI.

  A new WorkersAiTextModelOptions type is exported from @cloudflare/tanstack-ai and @cloudflare/tanstack-ai/adapters/workers-ai.

  Closes #503.