workers-ai-provider: tool_choice not forwarded to binding.run() inputs

[tomfuertes]

Package: workers-ai-provider v3.1.1

Bug

getArgs() correctly computes tool_choice via prepareToolsAndToolChoice(), but buildRunInputs() only forwards specific fields - silently dropping tool_choice:

// buildRunInputs returns:
{
  max_tokens: args.max_tokens,
  messages: finalMessages,
  temperature: args.temperature,
  tools: args.tools,       // included
  top_p: args.top_p,
  // tool_choice: args.tool_choice  <- MISSING
}

Impact

Models that require explicit tool_choice: "auto" to call tools (e.g. @cf/mistralai/mistral-small-3.1-24b-instruct) silently skip tool calls - the model acknowledges it should use tools in its text response but never calls them. @cf/zai-org/glm-4.7-flash was unaffected because it defaults to using tools when provided.

Workaround

Shim the binding to inject tool_choice: "auto" when tools are present:

const shimmedBinding = {
  run: (model, inputs, options) =>
    env.AI.run(model, inputs?.tools ? { ...inputs, tool_choice: "auto" } : inputs, options)
};
const provider = createWorkersAI({ binding: shimmedBinding });

Fix

One line in buildRunInputs in workersai-chat-language-model.ts:

tools: args.tools,
tool_choice: args.tool_choice,  // add this

[tomfuertes]

Edit: Might just be a mistral issue...

[threepointone]

yeah mistral's a bit weird, but maybe we could work around it in our lib