fix: pass provider extra_kwargs from config to ChatOpenAI

[swinney]

Summary

• Bug: BaseReActAgent._build_provider_config() hardcoded extra_kwargs = {}, silently discarding any extra_kwargs defined in the provider YAML config
• Fix: Initialize from cfg.get("extra_kwargs", {}) so user-defined kwargs flow through to the LangChain model constructor
• Impact: Any extra_kwargs in provider config (e.g. extra_body, model_kwargs, stream_options) were being ignored

Root Cause

In src/archi/pipelines/agents/base_react.py:882, _build_provider_config() builds the provider config dict that gets passed to get_model() → get_provider() → ChatOpenAI(**kwargs). The extra_kwargs field is spread into the ChatOpenAI constructor via **self.config.extra_kwargs in each provider's get_chat_model() method.

However, _build_provider_config() was initializing extra = {} instead of reading from the YAML config, so the pipeline between config file → provider → LangChain was broken at this point:

YAML config → _build_provider_config() → ProviderConfig(extra_kwargs={})  ← lost here
                                                              ↓
                                        ChatOpenAI(**{})  ← nothing passed

Example

This config should disable Qwen3's thinking mode when served via vLLM:

providers:
  openai:
    base_url: http://localhost:8000/v1
    extra_kwargs:
      extra_body:
        chat_template_kwargs:
          enable_thinking: false

Before this fix, extra_body never reached ChatOpenAI and thinking output leaked into the response content. After this fix, vLLM correctly receives chat_template_kwargs in the request body.

Test plan

• Deploy with a provider config containing extra_kwargs and verify the kwargs reach the LLM client
• Verify existing providers without extra_kwargs still work (empty dict fallback)
• Verify local provider mode override still works alongside user-defined extra_kwargs

🤖 Generated with Claude Code