fix: patch o1 support (#1978)

letta/llm_api/openai.py

@@ -7,7 +7,7 @@ from openai import OpenAI
 from letta.constants import LETTA_MODEL_ENDPOINT
 from letta.helpers.datetime_helpers import timestamp_to_datetime
 from letta.llm_api.helpers import add_inner_thoughts_to_functions, convert_to_structured_output, make_post_request
-from letta.llm_api.openai_client import supports_parallel_tool_calling, supports_temperature_param
+from letta.llm_api.openai_client import accepts_developer_role, supports_parallel_tool_calling, supports_temperature_param
 from letta.local_llm.constants import INNER_THOUGHTS_KWARG, INNER_THOUGHTS_KWARG_DESCRIPTION, INNER_THOUGHTS_KWARG_DESCRIPTION_GO_FIRST
 from letta.local_llm.utils import num_tokens_from_functions, num_tokens_from_messages
 from letta.log import get_logger
@@ -114,8 +114,16 @@ def build_openai_chat_completions_request(
             put_inner_thoughts_first=put_inner_thoughts_first,
         )
 
+    use_developer_message = accepts_developer_role(llm_config.model)
+
     openai_message_list = [
-        cast_message_to_subtype(m.to_openai_dict(put_inner_thoughts_in_kwargs=llm_config.put_inner_thoughts_in_kwargs)) for m in messages
+        cast_message_to_subtype(
+            m.to_openai_dict(
+                put_inner_thoughts_in_kwargs=llm_config.put_inner_thoughts_in_kwargs,
+                use_developer_message=use_developer_message,
+            )
+        )
+        for m in messages
     ]
 
     if llm_config.model:

letta/llm_api/openai_client.py

@@ -40,7 +40,19 @@ def is_openai_reasoning_model(model: str) -> bool:
     """Utility function to check if the model is a 'reasoner'"""
 
     # NOTE: needs to be updated with new model releases
-    return model.startswith("o1") or model.startswith("o3")
+    is_reasoning = model.startswith("o1") or model.startswith("o3")
+    return is_reasoning
+
+
+def accepts_developer_role(model: str) -> bool:
+    """Checks if the model accepts the 'developer' role. Note that not all reasoning models accept this role.
+
+    See: https://community.openai.com/t/developer-role-not-accepted-for-o1-o1-mini-o3-mini/1110750/7
+    """
+    if is_openai_reasoning_model(model):
+        return True
+    else:
+        return False
 
 
 def supports_temperature_param(model: str) -> bool:
@@ -102,7 +114,7 @@ class OpenAIClient(LLMClientBase):
                 put_inner_thoughts_first=True,
             )
 
-        use_developer_message = is_openai_reasoning_model(llm_config.model)
+        use_developer_message = accepts_developer_role(llm_config.model)
 
         openai_message_list = [
             cast_message_to_subtype(

letta/schemas/providers.py

@@ -201,7 +201,9 @@ class OpenAIProvider(Provider):
             # for openai, filter models
             if self.base_url == "https://api.openai.com/v1":
                 allowed_types = ["gpt-4", "o1", "o3"]
-                disallowed_types = ["transcribe", "search", "realtime", "tts", "audio", "computer"]
+                # NOTE: o1-mini and o1-preview do not support tool calling
+                # NOTE: o1-pro is only available in Responses API
+                disallowed_types = ["transcribe", "search", "realtime", "tts", "audio", "computer", "o1-mini", "o1-preview", "o1-pro"]
                 skip = True
                 for model_type in allowed_types:
                     if model_name.startswith(model_type):