feat: O1 support in goose for new models

src/exchange/providers/openai.py

@@ -14,6 +14,8 @@
     tools_to_openai_spec,
 )
 from exchange.tool import Tool
+from exchange.content import ToolUse, ToolResult, Text
+
 
 OPENAI_HOST = "https://api.openai.com/"
 
@@ -64,7 +66,7 @@ def complete(
         messages: List[Message],
         tools: Tuple[Tool],
         **kwargs: Dict[str, Any],
-    ) -> Tuple[Message, Usage]:
+    ) -> Tuple[Message, Usage]:        
         payload = dict(
             messages=[
                 {"role": "system", "content": system},
@@ -77,16 +79,77 @@ def complete(
         payload = {k: v for k, v in payload.items() if v}
         response = self._send_request(payload)
 
+
+
         # Check for context_length_exceeded error for single, long input message
         if "error" in response.json() and len(messages) == 1:
             openai_single_message_context_length_exceeded(response.json()["error"])
 
         data = raise_for_status(response).json()
-
         message = openai_response_to_message(data)
+
+        # optionally use the reasoning model to get a better answer if tool usage isn't required.
+        resoning_model = self.get_reasoning_model()
+        if resoning_model and len(self.tool_use(message)) == 0: # if a tool is needed we let things continue on without extra reasoning.
+            # will limit its invocation to non trivial things for now
+            filtered_messages = self.messages_filtered(messages)
+            latest_message =  filtered_messages[-1]
+            latest_completion = data["choices"][0]["message"]["content"]
+            if len(latest_message['content']) > 50 and len(latest_completion) > 100:                
+                print("---> using deep reasoning")
+                payload = dict(
+                    messages=[                        
+                        *filtered_messages,
+                        {"role": "user", "content": "TASK: please check the answer that follows, if it is ok then return it, otherwise rewrite it and return it:" + latest_completion},
+                    ],
+                    model=resoning_model,
+                    **kwargs,
+                )
+                payload = {k: v for k, v in payload.items() if v}
+                response = self._send_request(payload)
+
+                # Check for context_length_exceeded error for single, long input message
+                if "error" in response.json() and len(messages) == 1:
+                    openai_single_message_context_length_exceeded(response.json()["error"])
+
+                data = raise_for_status(response).json()
+                message = openai_response_to_message(data)
+
         usage = self.get_usage(data)
         return message, usage
 
+    def tool_use(self, message):
+        """ checks if the returned message is asking for tool usage or not """
+        return [content for content in message.content if isinstance(content, ToolUse) or isinstance(content, ToolResult)]
+
     @retry_httpx_request()
     def _send_request(self, payload: Any) -> httpx.Response:  # noqa: ANN401
         return self.client.post("v1/chat/completions", json=payload)
+
+
+    def messages_filtered(self, messages: List[Message]) -> List[Dict[str, Any]]:
+        """This is for models that don't handle tool call output or images directly"""
+        messages_spec = []
+        for message in messages:
+            converted = {"role": "user"}
+            output = []
+            for content in message.content:
+                if isinstance(content, Text):
+                    converted["content"] = content.text
+                elif isinstance(content, ToolResult):
+                    output.append(
+                        {
+                            "role": "user",
+                            "content": content.output,
+                        }
+                    )
+
+            if "content" in converted or "tool_calls" in converted:
+                output = [converted] + output
+            messages_spec.extend(output)
+        return messages_spec
+
+    def get_reasoning_model(self):
+        """ set this to o1-mini or o1-preview depending if you are doing code or not """
+        return os.environ.get("OPENAI_REASONING", None)
+