From 3ff28f3559c3598cd5efb6f3469735e14ce7f3e0 Mon Sep 17 00:00:00 2001
From: Yijia-Xiao <yijia-xiao@outlook.com>
Date: Sun, 22 Mar 2026 20:34:03 +0000
Subject: [PATCH] fix: use OpenAI Responses API for native models

Enable use_responses_api for native OpenAI provider, which supports
reasoning_effort with function tools across all model families.
Removes the UnifiedChatOpenAI subclass workaround.

Closes #403
---
 tradingagents/llm_clients/openai_client.py | 69 +++++++++++-----------
 1 file changed, 36 insertions(+), 33 deletions(-)

diff --git a/tradingagents/llm_clients/openai_client.py b/tradingagents/llm_clients/openai_client.py
index 4605c1f9..c314d077 100644
--- a/tradingagents/llm_clients/openai_client.py
+++ b/tradingagents/llm_clients/openai_client.py
@@ -6,28 +6,28 @@ from langchain_openai import ChatOpenAI
 from .base_client import BaseLLMClient
 from .validators import validate_model
 
+# Kwargs forwarded from user config to ChatOpenAI
+_PASSTHROUGH_KWARGS = (
+    "timeout", "max_retries", "reasoning_effort",
+    "api_key", "callbacks", "http_client", "http_async_client",
+)
 
-class UnifiedChatOpenAI(ChatOpenAI):
-    """ChatOpenAI subclass that strips temperature/top_p for GPT-5 family models.
-
-    GPT-5 family models use reasoning natively. temperature/top_p are only
-    accepted when reasoning.effort is 'none'; with any other effort level
-    (or for older GPT-5/GPT-5-mini/GPT-5-nano which always reason) the API
-    rejects these params. Langchain defaults temperature=0.7, so we must
-    strip it to avoid errors.
-
-    Non-GPT-5 models (GPT-4.1, xAI, Ollama, etc.) are unaffected.
-    """
-
-    def __init__(self, **kwargs):
-        if "gpt-5" in kwargs.get("model", "").lower():
-            kwargs.pop("temperature", None)
-            kwargs.pop("top_p", None)
-        super().__init__(**kwargs)
+# Provider base URLs and API key env vars
+_PROVIDER_CONFIG = {
+    "xai": ("https://api.x.ai/v1", "XAI_API_KEY"),
+    "openrouter": ("https://openrouter.ai/api/v1", "OPENROUTER_API_KEY"),
+    "ollama": ("http://localhost:11434/v1", None),
+}
 
 
 class OpenAIClient(BaseLLMClient):
-    """Client for OpenAI, Ollama, OpenRouter, and xAI providers."""
+    """Client for OpenAI, Ollama, OpenRouter, and xAI providers.
+
+    For native OpenAI models, uses the Responses API (/v1/responses) which
+    supports reasoning_effort with function tools across all model families
+    (GPT-4.1, GPT-5). Third-party compatible providers (xAI, OpenRouter,
+    Ollama) use standard Chat Completions.
+    """
 
     def __init__(
         self,
@@ -43,27 +43,30 @@ class OpenAIClient(BaseLLMClient):
         """Return configured ChatOpenAI instance."""
         llm_kwargs = {"model": self.model}
 
-        if self.provider == "xai":
-            llm_kwargs["base_url"] = "https://api.x.ai/v1"
-            api_key = os.environ.get("XAI_API_KEY")
-            if api_key:
-                llm_kwargs["api_key"] = api_key
-        elif self.provider == "openrouter":
-            llm_kwargs["base_url"] = "https://openrouter.ai/api/v1"
-            api_key = os.environ.get("OPENROUTER_API_KEY")
-            if api_key:
-                llm_kwargs["api_key"] = api_key
-        elif self.provider == "ollama":
-            llm_kwargs["base_url"] = "http://localhost:11434/v1"
-            llm_kwargs["api_key"] = "ollama"  # Ollama doesn't require auth
+        # Provider-specific base URL and auth
+        if self.provider in _PROVIDER_CONFIG:
+            base_url, api_key_env = _PROVIDER_CONFIG[self.provider]
+            llm_kwargs["base_url"] = base_url
+            if api_key_env:
+                api_key = os.environ.get(api_key_env)
+                if api_key:
+                    llm_kwargs["api_key"] = api_key
+            else:
+                llm_kwargs["api_key"] = "ollama"
         elif self.base_url:
             llm_kwargs["base_url"] = self.base_url
 
-        for key in ("timeout", "max_retries", "reasoning_effort", "api_key", "callbacks", "http_client", "http_async_client"):
+        # Forward user-provided kwargs
+        for key in _PASSTHROUGH_KWARGS:
             if key in self.kwargs:
                 llm_kwargs[key] = self.kwargs[key]
 
-        return UnifiedChatOpenAI(**llm_kwargs)
+        # Native OpenAI: use Responses API for consistent behavior across
+        # all model families. Third-party providers use Chat Completions.
+        if self.provider == "openai":
+            llm_kwargs["use_responses_api"] = True
+
+        return ChatOpenAI(**llm_kwargs)
 
     def validate_model(self) -> bool:
         """Validate model for the provider."""