feat: extract tool output from openai-agents sdk (#507)

Signed-off-by: CormickKneey <[email protected]>

Author: CormickKneey

areal/experimental/openai/client.py

@@ -15,6 +15,7 @@
 from openai.types.chat import (
     ChatCompletion,
     ChatCompletionMessage,
+    ChatCompletionToolMessageParam,
     ChatCompletionToolParam,
 )
 from openai.types.chat.chat_completion import Choice
@@ -106,7 +107,7 @@ async def create(
         if extra_body is None:
             extra_body = {}
         # Convert messages to prompt format
-        tools = tools if tools is not NOT_GIVEN else None
+        tools = tools if not is_omitted(tools) else None
         if self.chat_template_type == "hf":
             prompt_token_ids = self.tokenizer.apply_chat_template(
                 messages_list,
@@ -128,23 +129,23 @@ async def create(
                 f"Unsupported chat_template_type {self.chat_template_type}"
             )
 
-        temp = 1.0 if temperature is NOT_GIVEN else (temperature or 0.0)
+        temp = 1.0 if is_omitted(temperature) else (temperature or 0.0)
         max_new_tokens = 512
-        if max_tokens is not NOT_GIVEN and max_tokens is not None:
+        if not is_omitted(max_tokens):
             max_new_tokens = max_tokens - len(prompt_token_ids)
             if max_new_tokens <= 0:
                 raise RuntimeError(
                     "max_tokens must be greater than the number of prompt tokens"
                 )
-        if max_completion_tokens is not NOT_GIVEN and max_completion_tokens is not None:
+        if not is_omitted(max_completion_tokens):
             max_new_tokens = min(max_new_tokens, max_completion_tokens)
 
-        top_p_val = 1.0 if top_p is NOT_GIVEN else (top_p or 1.0)
-        stop_tokens = None if stop is NOT_GIVEN else stop
+        top_p_val = 1.0 if is_omitted(top_p) else (top_p or 1.0)
+        stop_tokens = None if is_omitted(stop) else stop
         if stop_tokens is not None and not isinstance(stop_tokens, list):
             stop_tokens = [stop_tokens]
 
-        if frequency_penalty is NOT_GIVEN or frequency_penalty is None:
+        if is_omitted(frequency_penalty):
             frequency_penalty = 0.0
 
         # Create generation config
@@ -165,7 +166,7 @@ async def create(
             input_ids=prompt_token_ids,
             gconfig=gconfig,
             rid=str(uuid.uuid4()),
-            metadata=metadata if metadata is not NOT_GIVEN else {},
+            metadata=metadata if not is_omitted(metadata) else {},
             tokenizer=self.tokenizer,
         )
 
@@ -215,7 +216,7 @@ async def create(
             ),
         )
 
-        if store is NOT_GIVEN or store:
+        if is_omitted(store) or store:
             # Cache the completion with its input messages
             self._cache[completion_id] = InteractionWithTokenLogpReward(
                 completion=deepcopy(chat_completion),
@@ -269,13 +270,45 @@ async def create(
 
         # Build a simple messages list compatible with tokenizer chat template
         messages_list: list[dict] = []
-        if instructions is not NOT_GIVEN and instructions is not None:
+        if not is_omitted(instructions):
             messages_list = [
                 {"role": "system", "content": instructions},
             ]
-        if input is NOT_GIVEN or input is None:
+        if is_omitted(input):
             raise ValueError("input is required for Responses.create")
 
+        def _convert_tool_output_format(
+            item: dict,
+        ) -> ChatCompletionToolMessageParam | dict:
+            """Convert custom tool output format to standard chat template format.
+
+            Converts openai.types.responses.response_input_item_param.FunctionCallOutput
+            to openai.types.chat.ChatCompletionToolMessageParam.
+
+            Args:
+                item: Input dict, could be FunctionCallOutput from openai-agents SDK
+                      with format: {'call_id': str, 'output': str, 'type': 'function_call_output'}
+
+            Returns:
+                ChatCompletionToolMessageParam (TypedDict) with format:
+                {'role': 'tool', 'content': str, 'tool_call_id': str}
+                or the original dict if conversion is not needed.
+            """
+            if (
+                isinstance(item, dict)
+                and "output" in item
+                and item.get("type") == "function_call_output"
+            ):
+                converted = {
+                    "role": "tool",
+                    "content": item["output"],
+                }
+                # Add tool_call_id if present
+                if "call_id" in item:
+                    converted["tool_call_id"] = item["call_id"]
+                return converted
+            return item
+
         def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
             messages_list = []
             if "content" in item:
@@ -286,13 +319,17 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
                 elif isinstance(item["content"], Iterable):
                     for content in item["content"]:
                         if isinstance(content, dict):
-                            messages_list.append(deepcopy(content))
+                            # Convert tool output format if needed
+                            converted = _convert_tool_output_format(content)
+                            messages_list.append(deepcopy(converted))
                         else:
                             raise ValueError("Unsupported content format")
                 else:
                     raise ValueError("Unsupported input item format")
             else:
-                messages_list.append(deepcopy(item))
+                # Convert tool output format if needed
+                converted = _convert_tool_output_format(item)
+                messages_list.append(deepcopy(converted))
             return messages_list
 
         if isinstance(input, str):
@@ -309,7 +346,7 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
             )
 
         # Apply chat template
-        tools = list(tools) if tools is not NOT_GIVEN else None
+        tools = list(tools) if not is_omitted(tools) else None
         if self.chat_template_type == "hf":
             prompt_token_ids = self.tokenizer.apply_chat_template(
                 messages_list,
@@ -332,10 +369,10 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
             )
 
         # Map sampling params
-        temp = 1.0 if temperature is NOT_GIVEN else (temperature or 0.0)
-        top_p_val = 1.0 if top_p is NOT_GIVEN else (top_p or 1.0)
+        temp = 1.0 if is_omitted(temperature) else (temperature or 0.0)
+        top_p_val = 1.0 if is_omitted(top_p) else (top_p or 1.0)
         max_new_tokens = 512
-        if max_output_tokens is not NOT_GIVEN and max_output_tokens is not None:
+        if not is_omitted(max_output_tokens):
             max_new_tokens = max_output_tokens
 
         stop = kwargs.get("stop", None)
@@ -359,7 +396,7 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
             input_ids=prompt_token_ids,
             gconfig=gconfig,
             rid=str(uuid.uuid4()),
-            metadata=metadata if metadata is not NOT_GIVEN else {},
+            metadata=metadata if not is_omitted(metadata) else {},
             tokenizer=self.tokenizer,
         )
 
@@ -369,7 +406,7 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
 
         # Parse tool calls.
         tool_calls = None
-        if tool_choice != "none" and tools:
+        if not is_omitted(tool_choice) and tool_choice != "none" and tools:
             tool_calls, output_text, engine_resp.stop_reason = process_tool_calls(
                 output_text,
                 tools,
@@ -420,14 +457,14 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
             created_at=current_time,
             error=None,
             incomplete_details=None,
-            instructions=None if instructions is NOT_GIVEN else instructions,
-            metadata=None if metadata is NOT_GIVEN else metadata,
+            instructions=None if is_omitted(instructions) else instructions,
+            metadata=None if is_omitted(metadata) else metadata,
             model="None",
             object="response",
             output=resp_output,
             parallel_tool_calls=False,
             temperature=temp,
-            tool_choice=tool_choice if tool_choice is not NOT_GIVEN else "none",
+            tool_choice=tool_choice if not is_omitted(tool_choice) else "none",
             tools=tools,
             top_p=top_p_val,
             background=None,
@@ -453,7 +490,7 @@ def _build_messages_list(item: ResponseInputItemParam) -> list[dict]:
             response=deepcopy(response),
             model_response=engine_resp,  # Should not deepcopy because of tokenizer
             input_data=(
-                deepcopy(input) if input is not NOT_GIVEN else ""
+                deepcopy(input) if not is_omitted(input) else ""
             ),  # Store a copy of the input data
             chat_template_type=self.chat_template_type,
         )
@@ -751,3 +788,23 @@ def export_responses(
             "export_responses is deprecated. Please use export_interactions instead."
         )
         return self.export_interactions(style)
+
+
+def is_omitted(value) -> bool:
+    """Check if a value is NOT_GIVEN or Omit type or None."""
+    if value is NOT_GIVEN or value is None:
+        return True
+    # Use isinstance for type safety and robustness
+    # Check for common omitted types from OpenAI SDK
+    try:
+        from openai import Omit
+
+        if isinstance(value, Omit):
+            return True
+    except ImportError:
+        pass
+
+    # Fallback for other omit types
+    if hasattr(value, "__class__"):
+        return value.__class__.__name__ in ("NotGiven", "Omit")
+    return False