Fix OpenAI Responses API tool calls with reasoning (pydantic#2869)

Author: DouweM

pydantic_ai_slim/pydantic_ai/models/openai.py

@@ -878,7 +878,9 @@ def _process_response(self, response: responses.Response) -> ModelResponse:
                     if isinstance(content, responses.ResponseOutputText):  # pragma: no branch
                         items.append(TextPart(content.text))
             elif isinstance(item, responses.ResponseFunctionToolCall):
-                items.append(ToolCallPart(item.name, item.arguments, tool_call_id=item.call_id))
+                items.append(
+                    ToolCallPart(item.name, item.arguments, tool_call_id=_combine_tool_call_ids(item.call_id, item.id))
+                )
 
         finish_reason: FinishReason | None = None
         provider_details: dict[str, Any] | None = None
@@ -1084,27 +1086,29 @@ async def _map_messages(  # noqa: C901
                     elif isinstance(part, UserPromptPart):
                         openai_messages.append(await self._map_user_prompt(part))
                     elif isinstance(part, ToolReturnPart):
-                        openai_messages.append(
-                            FunctionCallOutput(
-                                type='function_call_output',
-                                call_id=_guard_tool_call_id(t=part),
-                                output=part.model_response_str(),
-                            )
+                        call_id = _guard_tool_call_id(t=part)
+                        call_id, _ = _split_combined_tool_call_id(call_id)
+                        item = FunctionCallOutput(
+                            type='function_call_output',
+                            call_id=call_id,
+                            output=part.model_response_str(),
                         )
+                        openai_messages.append(item)
                     elif isinstance(part, RetryPromptPart):
                         # TODO(Marcelo): How do we test this conditional branch?
                         if part.tool_name is None:  # pragma: no cover
                             openai_messages.append(
                                 Message(role='user', content=[{'type': 'input_text', 'text': part.model_response()}])
                             )
                         else:
-                            openai_messages.append(
-                                FunctionCallOutput(
-                                    type='function_call_output',
-                                    call_id=_guard_tool_call_id(t=part),
-                                    output=part.model_response(),
-                                )
+                            call_id = _guard_tool_call_id(t=part)
+                            call_id, _ = _split_combined_tool_call_id(call_id)
+                            item = FunctionCallOutput(
+                                type='function_call_output',
+                                call_id=call_id,
+                                output=part.model_response(),
                             )
+                            openai_messages.append(item)
                     else:
                         assert_never(part)
             elif isinstance(message, ModelResponse):
@@ -1141,12 +1145,18 @@ async def _map_messages(  # noqa: C901
 
     @staticmethod
     def _map_tool_call(t: ToolCallPart) -> responses.ResponseFunctionToolCallParam:
-        return responses.ResponseFunctionToolCallParam(
-            arguments=t.args_as_json_str(),
-            call_id=_guard_tool_call_id(t=t),
+        call_id = _guard_tool_call_id(t=t)
+        call_id, id = _split_combined_tool_call_id(call_id)
+
+        param = responses.ResponseFunctionToolCallParam(
             name=t.tool_name,
+            arguments=t.args_as_json_str(),
+            call_id=call_id,
             type='function_call',
         )
+        if id:  # pragma: no branch
+            param['id'] = id
+        return param
 
     def _map_json_schema(self, o: OutputObjectDefinition) -> responses.ResponseFormatTextJSONSchemaConfigParam:
         response_format_param: responses.ResponseFormatTextJSONSchemaConfigParam = {
@@ -1365,7 +1375,7 @@ async def _get_event_iterator(self) -> AsyncIterator[ModelResponseStreamEvent]:
                         vendor_part_id=chunk.item.id,
                         tool_name=chunk.item.name,
                         args=chunk.item.arguments,
-                        tool_call_id=chunk.item.call_id,
+                        tool_call_id=_combine_tool_call_ids(chunk.item.call_id, chunk.item.id),
                     )
                 elif isinstance(chunk.item, responses.ResponseReasoningItem):
                     pass
@@ -1506,3 +1516,17 @@ def _map_usage(response: chat.ChatCompletion | ChatCompletionChunk | responses.R
             u.input_audio_tokens = response_usage.prompt_tokens_details.audio_tokens or 0
             u.cache_read_tokens = response_usage.prompt_tokens_details.cached_tokens or 0
         return u
+
+
+def _combine_tool_call_ids(call_id: str, id: str | None) -> str:
+    # When reasoning, the Responses API requires the `ResponseFunctionToolCall` to be returned with both the `call_id` and `id` fields.
+    # Our `ToolCallPart` has only the `call_id` field, so we combine the two fields into a single string.
+    return f'{call_id}|{id}' if id else call_id
+
+
+def _split_combined_tool_call_id(combined_id: str) -> tuple[str, str | None]:
+    if '|' in combined_id:
+        call_id, id = combined_id.split('|', 1)
+        return call_id, id
+    else:
+        return combined_id, None  # pragma: no cover