Support for anthropic models

[dimavrem22]

• added anthropic models and anthropic llm client (with rate limits exponential decay)
• all of the new specialized agents (network spy, trace hound, and docs digger) can use openai or anthropic models
• added unit tests to anthropic llm client and for the vendors

[dimavrem22]

@claude review pretty pretty please

[dimavrem22]

@claude check again!

[dimavrem22]

@claude

[rayruizhiliao]

Can we update README, where OpenAI API key is currently mentioned as a prerequisite? And also code snippet

# Set your OpenAI API key
export OPENAI_API_KEY="sk-..."

[dimavrem22]

for now discovery and guide agent still require openai models because of vectorstores. Only specialized agents can be run with anthropic currently. so i think we should keep the docs as is for now

Dima: for now discovery and guide agent still require openai models because of vectorstores. Only specialized agents can be run with anthropic currently. so i think we should keep the docs as is for now

[claude]

Code Review

I found an issue during the code review:

Issue: Missing "none" tool_choice handling

Files affected:

• bluebox/llms/openai_client.py (lines 104-106)
• bluebox/llms/anthropic_client.py (lines 149-156)

Description:

The abstract base class AbstractLLMVendorClient documents "none" as a valid tool_choice option:

https://github.com/VectorlyApp/bluebox-sdk/blob/d5c9b19a2695c2e658323afee1ed94737e569846/bluebox/llms/abstract_llm_vendor_client.py#L128-L134

However, neither the OpenAIClient nor the AnthropicClient handle this value correctly. When tool_choice="none" is passed:

OpenAI Client: Incorrectly converts it to {"type": "function", "name": "none"} instead of passing "none" directly (which is the correct format for OpenAI's API).

https://github.com/VectorlyApp/bluebox-sdk/blob/d5c9b19a2695c2e658323afee1ed94737e569846/bluebox/llms/openai_client.py#L103-L107

Anthropic Client: Incorrectly converts it to {"type": "tool", "name": "none"}. However, Anthropic's Messages API doesn't have a "none" equivalent - to disable tools, you simply don't include them in the request.

https://github.com/VectorlyApp/bluebox-sdk/blob/d5c9b19a2695c2e658323afee1ed94737e569846/bluebox/llms/anthropic_client.py#L148-L157

Recommendation:

1. OpenAI Client: Add "none" to the list of pass-through values:

   if tool_choice is not None and tool_choice not in ["auto", "required", "none"]:
       return {"type": "function", "name": tool_choice}
   return tool_choice

2. Anthropic Client: Either:

   • Add a check to handle "none" (perhaps by returning None to skip tool_choice entirely)
   • Remove "none" from the abstract base class documentation if it's not universally supported

Impact: Users attempting to use tool_choice="none" will encounter API errors since both clients treat it as a tool name.