wip first pass at new OpenAI Responses API support

Author: cpsievert

chatlas/__init__.py

@@ -27,7 +27,9 @@
 from ._provider_huggingface import ChatHuggingFace
 from ._provider_mistral import ChatMistral
 from ._provider_ollama import ChatOllama
-from ._provider_openai import ChatAzureOpenAI, ChatOpenAI
+from ._provider_openai import ChatOpenAI
+from ._provider_openai_azure import ChatAzureOpenAI
+from ._provider_openai_responses import ChatOpenAIResponses
 from ._provider_openrouter import ChatOpenRouter
 from ._provider_perplexity import ChatPerplexity
 from ._provider_portkey import ChatPortkey
@@ -59,6 +61,7 @@
     "ChatMistral",
     "ChatOllama",
     "ChatOpenAI",
+    "ChatOpenAIResponses",
     "ChatOpenRouter",
     "ChatAzureOpenAI",
     "ChatPerplexity",

chatlas/_auto.py

@@ -17,7 +17,8 @@
 from ._provider_huggingface import ChatHuggingFace
 from ._provider_mistral import ChatMistral
 from ._provider_ollama import ChatOllama
-from ._provider_openai import ChatAzureOpenAI, ChatOpenAI
+from ._provider_openai import ChatOpenAI
+from ._provider_openai_azure import ChatAzureOpenAI
 from ._provider_openrouter import ChatOpenRouter
 from ._provider_perplexity import ChatPerplexity
 from ._provider_portkey import ChatPortkey

chatlas/_chat.py

@@ -2539,12 +2539,11 @@ def emit(text: str | Content):
                 result,
                 has_data_model=data_model is not None,
             )
-
             if echo == "all":
                 emit_other_contents(turn, emit)
 
         else:
-            response = self.provider.chat_perform(
+            result = self.provider.chat_perform(
                 stream=False,
                 turns=[*self._turns, user_turn],
                 tools=self._tools,
@@ -2553,7 +2552,7 @@ def emit(text: str | Content):
             )
 
             turn = self.provider.value_turn(
-                response, has_data_model=data_model is not None
+                result, has_data_model=data_model is not None
             )
             if turn.text:
                 emit(turn.text)
@@ -2610,7 +2609,7 @@ def emit(text: str | Content):
                 emit_other_contents(turn, emit)
 
         else:
-            response = await self.provider.chat_perform_async(
+            result = await self.provider.chat_perform_async(
                 stream=False,
                 turns=[*self._turns, user_turn],
                 tools=self._tools,
@@ -2619,7 +2618,7 @@ def emit(text: str | Content):
             )
 
             turn = self.provider.value_turn(
-                response, has_data_model=data_model is not None
+                result, has_data_model=data_model is not None
             )
             if turn.text:
                 emit(turn.text)

chatlas/_content.py

@@ -125,6 +125,7 @@ def from_tool(cls, tool: "Tool") -> "ToolInfo":
     "tool_result_resource",
     "json",
     "pdf",
+    "thinking",
 ]
 """
 A discriminated union of all content types.
@@ -682,6 +683,40 @@ def __repr__(self, indent: int = 0):
         return " " * indent + f"<ContentPDF size={len(self.data)}>"
 
 
+class ContentThinking(Content):
+    """
+    Thinking/reasoning content
+
+    This content type represents reasoning traces from models that support
+    extended thinking (like OpenAI's o-series models). The thinking content
+    is not meant to be sent back to the model but is useful for debugging
+    and understanding the model's reasoning process.
+
+    Parameters
+    ----------
+    thinking
+        The thinking/reasoning text from the model.
+    extra
+        Additional metadata associated with the thinking content (e.g.,
+        encrypted content, status information).
+    """
+
+    thinking: str
+    extra: Optional[dict[str, Any]] = None
+
+    content_type: ContentTypeEnum = "thinking"
+
+    def __str__(self):
+        return f"<thinking>\n{self.thinking}\n</thinking>\n"
+
+    def _repr_markdown_(self):
+        return self.__str__()
+
+    def __repr__(self, indent: int = 0):
+        preview = self.thinking[:50] + "..." if len(self.thinking) > 50 else self.thinking
+        return " " * indent + f"<ContentThinking thinking='{preview}'>"
+
+
 ContentUnion = Union[
     ContentText,
     ContentImageRemote,
@@ -692,6 +727,7 @@ def __repr__(self, indent: int = 0):
     ContentToolResultResource,
     ContentJson,
     ContentPDF,
+    ContentThinking,
 ]
 
 
@@ -724,6 +760,8 @@ def create_content(data: dict[str, Any]) -> ContentUnion:
         return ContentJson.model_validate(data)
     elif ct == "pdf":
         return ContentPDF.model_validate(data)
+    elif ct == "thinking":
+        return ContentThinking.model_validate(data)
     else:
         raise ValueError(f"Unknown content type: {ct}")
 

chatlas/_provider_anthropic.py

@@ -242,28 +242,6 @@ def list_models(self):
 
         return res
 
-    @overload
-    def chat_perform(
-        self,
-        *,
-        stream: Literal[False],
-        turns: list[Turn],
-        tools: dict[str, Tool],
-        data_model: Optional[type[BaseModel]] = None,
-        kwargs: Optional["SubmitInputArgs"] = None,
-    ): ...
-
-    @overload
-    def chat_perform(
-        self,
-        *,
-        stream: Literal[True],
-        turns: list[Turn],
-        tools: dict[str, Tool],
-        data_model: Optional[type[BaseModel]] = None,
-        kwargs: Optional["SubmitInputArgs"] = None,
-    ): ...
-
     def chat_perform(
         self,
         *,
@@ -276,28 +254,6 @@ def chat_perform(
         kwargs = self._chat_perform_args(stream, turns, tools, data_model, kwargs)
         return self._client.messages.create(**kwargs)  # type: ignore
 
-    @overload
-    async def chat_perform_async(
-        self,
-        *,
-        stream: Literal[False],
-        turns: list[Turn],
-        tools: dict[str, Tool],
-        data_model: Optional[type[BaseModel]] = None,
-        kwargs: Optional["SubmitInputArgs"] = None,
-    ): ...
-
-    @overload
-    async def chat_perform_async(
-        self,
-        *,
-        stream: Literal[True],
-        turns: list[Turn],
-        tools: dict[str, Tool],
-        data_model: Optional[type[BaseModel]] = None,
-        kwargs: Optional["SubmitInputArgs"] = None,
-    ): ...
-
     async def chat_perform_async(
         self,
         *,

chatlas/_provider_github.py

@@ -7,7 +7,8 @@
 
 from ._chat import Chat
 from ._logging import log_model_default
-from ._provider_openai import ModelInfo, OpenAIProvider
+from ._provider import ModelInfo
+from ._provider_openai import OpenAIProvider
 from ._utils import MISSING, MISSING_TYPE, is_testing
 
 if TYPE_CHECKING:

chatlas/_provider_ollama.py

@@ -7,7 +7,8 @@
 import orjson
 
 from ._chat import Chat
-from ._provider_openai import ModelInfo, OpenAIProvider
+from ._provider import ModelInfo
+from ._provider_openai import OpenAIProvider
 from ._utils import MISSING_TYPE, is_testing
 
 if TYPE_CHECKING: