`supports_tool_choice` and other logic for Claude models with thinking enabled is incorrect

[awg66]

Here are the docs on extended thinking: https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-extended-thinking.html

According to this, all Claude models with thinking enabled (3.7, 4 Opus, 4 Sonnet) should support a tool choice of "auto". And this is confirmed by manual testing. Yet the ChatBedrockConverse model does not support that tool choice for these models with thinking enabled

Additionally, much of the logic in tool binding and structured output that specifically checks if the model is 3.7 with thinking enabled hasn't been updated to also check if the model is a 4.0 Opus or Sonnet model with thinking enabled.

[michaelnchin]

The Bedrock docs page says any is valid:

Tool choice limitation: Tool use with thinking only supports tool_choice: any. It does not support providing a specific tool, auto, or any other values.

I guess this might be a typo since the Anthropic docs state otherwise:

When using extended thinking with tool use, tool_choice: {"type": "any"} and tool_choice: {"type": "tool", "name": "..."} are not supported and will result in an error. Only tool_choice: {"type": "auto"} (the default) and tool_choice: {"type": "none"} are compatible with extended thinking.

[michaelnchin]

Checked with supports_tool_choice_values set manually - auto works on Claude 3.7+ while any throws an error.

Thanks for the catch Alex Gatto (@awg66) !