feat: add chatgpt oauth client for codex routing (#8774)

* base

* refresh

* use default model fallback

* patch

* streaming

* generate

letta/adapters/simple_llm_stream_adapter.py

@@ -2,6 +2,9 @@ import json
 from typing import AsyncGenerator, List
 
 from letta.adapters.letta_llm_stream_adapter import LettaLLMStreamAdapter
+from letta.log import get_logger
+
+logger = get_logger(__name__)
 from letta.helpers.datetime_helpers import get_utc_timestamp_ns
 from letta.interfaces.anthropic_parallel_tool_call_streaming_interface import SimpleAnthropicStreamingInterface
 from letta.interfaces.gemini_streaming_interface import SimpleGeminiStreamingInterface
@@ -75,12 +78,22 @@ class SimpleLLMStreamAdapter(LettaLLMStreamAdapter):
                 run_id=self.run_id,
                 step_id=step_id,
             )
-        elif self.llm_config.model_endpoint_type in [ProviderType.openai, ProviderType.deepseek, ProviderType.zai]:
+        elif self.llm_config.model_endpoint_type in [
+            ProviderType.openai,
+            ProviderType.deepseek,
+            ProviderType.zai,
+            ProviderType.chatgpt_oauth,
+        ]:
             # Decide interface based on payload shape
             use_responses = "input" in request_data and "messages" not in request_data
             # No support for Responses API proxy
             is_proxy = self.llm_config.provider_name == "lmstudio_openai"
 
+            # ChatGPT OAuth always uses Responses API format
+            if self.llm_config.model_endpoint_type == ProviderType.chatgpt_oauth:
+                use_responses = True
+                is_proxy = False
+
             if use_responses and not is_proxy:
                 self.interface = SimpleOpenAIResponsesStreamingInterface(
                     is_openai_proxy=False,
@@ -109,9 +122,6 @@ class SimpleLLMStreamAdapter(LettaLLMStreamAdapter):
         else:
             raise ValueError(f"Streaming not supported for provider {self.llm_config.model_endpoint_type}")
 
-        # Extract optional parameters
-        # ttft_span = kwargs.get('ttft_span', None)
-
         # Start the streaming request (map provider errors to common LLMError types)
         try:
             # Gemini uses async generator pattern (no await) to maintain connection lifecycle