Enable OpenAI JSON mode for guaranteed valid JSON output

Adds response_format: {type: 'json_object'} to the ChatOpenAI
constructor via model_kwargs. This is a model-level setting that
constrains token generation to valid JSON — no extra LLM call,
no post-processing, just guaranteed JSON from every response.

This is different from LangGraph's response_format parameter
(which adds a separate LLM call). This is OpenAI's native JSON
mode applied at the API level during the same call.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: abossard

backend/agent_builder/engine/react_runner.py

@@ -41,6 +41,7 @@ def build_llm(
         "api_key": api_key,
         "base_url": base_url or None,
         "temperature": temperature,
+        "model_kwargs": {"response_format": {"type": "json_object"}},
     }
     if max_tokens > 0:
         kwargs["max_tokens"] = max_tokens