OpenAI reasoning intigiert

math-tutor/backend/app/deterministic_services/llm_client.py

@@ -6,15 +6,13 @@ from typing import Any, Callable
 
 import ollama
 from mistralai.client import Mistral
-from openai import OpenAI
+from openai import BadRequestError, OpenAI
 
 from app import config
 from app.deterministic_services import llm_quota
 from app.deterministic_services.llm_tool_client import (
     append_tool_trace_log,
-    resolve_mcp_base_url,
     run_tool_chat_sync,
-    with_mcp_session,
 )
 
 
@@ -71,6 +69,110 @@ def _get_message_field(message: object, field: str) -> str:
         return getattr(message, field) or ""
     return ""
 
+# ---
+# OpenAI response extraction helpers
+# ---
+
+def _field(value: object, name: str, default=None):
+    if isinstance(value, dict):
+        return value.get(name, default)
+    return getattr(value, name, default)
+
+
+def _extract_block_text(blocks: list[object]) -> str:
+    return "".join(str(text) for block in blocks or [] if (text := _field(block, "text")))
+
+
+def _extract_response_thinking(output: list[object]) -> str:
+    parts: list[str] = []
+    for item in output:
+        if _field(item, "type") != "reasoning":
+            continue
+        parts.append(_extract_block_text(_field(item, "summary", []) or []))
+        parts.append(_extract_block_text(_field(item, "content", []) or []))
+    return "\n".join(part for part in parts if part)
+
+
+def _extract_response_openai(response) -> object:
+    output = _field(response, "output", []) or []
+    thinking = _extract_response_thinking(output)
+
+    if isinstance(response, dict):
+        output_text = response.get("output_text")
+        if output_text is not None:
+            message = {"role": "assistant", "content": output_text}
+            if thinking:
+                message["thinking"] = thinking
+            return message
+    else:
+        output_text = getattr(response, "output_text", None)
+        if output_text is not None:
+            message = {"role": "assistant", "content": output_text}
+            if thinking:
+                message["thinking"] = thinking
+            return message
+
+    parts: list[str] = []
+    for item in output:
+        content = item.get("content", []) if isinstance(item, dict) else getattr(item, "content", []) or []
+        for block in content or []:
+            block_type = block.get("type") if isinstance(block, dict) else getattr(block, "type", "")
+            text = block.get("text") if isinstance(block, dict) else getattr(block, "text", None)
+            if block_type in {"output_text", "text"} and text:
+                parts.append(str(text))
+    message = {"role": "assistant", "content": "".join(parts)}
+    if thinking:
+        message["thinking"] = thinking
+    return message
+
+
+def _openai_reasoning() -> dict[str, str]:
+    if config.get_llm_thinking_enabled():
+        return {"effort": "high", "summary": "detailed"}
+    return {"effort": "none"}
+
+
+def _message_content_text(content: object) -> str:
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = []
+        for item in content:
+            text = _field(item, "text")
+            if text:
+                parts.append(str(text))
+        return "\n".join(parts)
+    return str(content or "")
+
+
+def _messages_to_responses_input(messages: list[dict]) -> tuple[str | None, str]:
+    instructions = "\n\n".join(
+        _message_content_text(message.get("content"))
+        for message in messages
+        if message.get("role") == "system"
+    ).strip() or None
+
+    turns = []
+    for message in messages:
+        role = message.get("role")
+        if role == "system":
+            continue
+        content = _message_content_text(message.get("content")).strip()
+        if content:
+            turns.append(f"{role or 'user'}: {content}")
+    return instructions, "\n\n".join(turns)
+
+
+def _is_unsupported_temperature_error(exc: Exception) -> bool:
+    text = str(exc)
+    return (
+        isinstance(exc, BadRequestError)
+        and (getattr(exc, "param", None) == "temperature" or "temperature" in text)
+        and "Unsupported parameter" in text
+    )
+
+
+
 
 # ---
 # Quota
@@ -152,11 +254,21 @@ def _chat_openai_compatible(
     settings = settings or _require_openai_chat_settings()
     timeout = settings.timeout or 60.0
     client = OpenAI(api_key=settings.api_key, base_url=settings.base_url, timeout=timeout)
-    kwargs = {"messages": messages, "model": settings.model}
+    instructions, response_input = _messages_to_responses_input(messages)
+    kwargs = {"input": response_input, "model": settings.model}
+    if instructions:
+        kwargs["instructions"] = instructions
     if settings.temperature is not None:
         kwargs["temperature"] = settings.temperature
-    response = client.chat.completions.create(**kwargs)
-    message = response.choices[0].message if response.choices else {}
+    kwargs["reasoning"] = _openai_reasoning()
+    try:
+        response = client.responses.create(**kwargs)
+    except BadRequestError as exc:
+        if "temperature" not in kwargs or not _is_unsupported_temperature_error(exc):
+            raise
+        kwargs.pop("temperature")
+        response = client.responses.create(**kwargs)
+    message = _extract_response_openai(response)
     return {"raw": response, "message": message}
 
 

math-tutor/backend/app/deterministic_services/llm_tool_client.py

@@ -6,7 +6,7 @@ import httpx
 import ollama
 from mcp.client.session import ClientSession
 from mcp.client.streamable_http import streamable_http_client
-from openai import AsyncOpenAI
+from openai import AsyncOpenAI, BadRequestError
 
 from app import config
 from app.deterministic_services import tool_log_context
@@ -14,6 +14,15 @@ from app.deterministic_services import tool_log_context
 AUTH_HEADER = "X-MCP-Shared-Secret"
 
 
+def _is_unsupported_temperature_error(exc: Exception) -> bool:
+    text = str(exc)
+    return (
+        isinstance(exc, BadRequestError)
+        and (getattr(exc, "param", None) == "temperature" or "temperature" in text)
+        and "Unsupported parameter" in text
+    )
+
+
 def _extract_message(response) -> object:
     if isinstance(response, dict) and "message" in response:
         return response["message"]
@@ -30,6 +39,38 @@ def _extract_tool_calls(message: object) -> list[Any]:
     return []
 
 
+def _extract_response_tool_calls(response: object) -> list[Any]:
+    output = response.get("output", []) if isinstance(response, dict) else getattr(response, "output", []) or []
+    return [
+        item for item in output
+        if (item.get("type") if isinstance(item, dict) else getattr(item, "type", "")) == "function_call"
+    ]
+
+
+def _extract_response_output(response: object) -> list[Any]:
+    return response.get("output", []) if isinstance(response, dict) else getattr(response, "output", []) or []
+
+
+def _field(value: object, name: str, default=None):
+    if isinstance(value, dict):
+        return value.get(name, default)
+    return getattr(value, name, default)
+
+
+def _extract_block_text(blocks: list[object]) -> str:
+    return "".join(str(text) for block in blocks or [] if (text := _field(block, "text")))
+
+
+def _extract_response_thinking(response: object) -> str:
+    parts: list[str] = []
+    for item in _extract_response_output(response):
+        if _field(item, "type") != "reasoning":
+            continue
+        parts.append(_extract_block_text(_field(item, "summary", []) or []))
+        parts.append(_extract_block_text(_field(item, "content", []) or []))
+    return "\n".join(part for part in parts if part)
+
+
 def _normalize_message(message: object) -> dict[str, Any]:
     if isinstance(message, dict):
         return dict(message)
@@ -44,6 +85,10 @@ def _normalize_message(message: object) -> dict[str, Any]:
 
 
 def _extract_tool_call_name(tool_call: object) -> str:
+    if hasattr(tool_call, "name"):
+        return str(getattr(tool_call, "name") or "")
+    if isinstance(tool_call, dict) and "name" in tool_call:
+        return str(tool_call.get("name") or "")
     function = getattr(tool_call, "function", None)
     if function is not None and hasattr(function, "name"):
         return getattr(function, "name")
@@ -54,11 +99,16 @@ def _extract_tool_call_name(tool_call: object) -> str:
 
 
 def _extract_tool_call_arguments(tool_call: object) -> dict[str, Any]:
-    function = getattr(tool_call, "function", None)
     raw_arguments: Any = None
-    if function is not None and hasattr(function, "arguments"):
+    if hasattr(tool_call, "arguments"):
+        raw_arguments = getattr(tool_call, "arguments")
+    elif isinstance(tool_call, dict) and "arguments" in tool_call:
+        raw_arguments = tool_call.get("arguments")
+
+    function = getattr(tool_call, "function", None)
+    if raw_arguments is None and function is not None and hasattr(function, "arguments"):
         raw_arguments = getattr(function, "arguments")
-    elif isinstance(tool_call, dict):
+    elif raw_arguments is None and isinstance(tool_call, dict):
         function_dict = tool_call.get("function") or {}
         raw_arguments = function_dict.get("arguments")
 
@@ -70,6 +120,10 @@ def _extract_tool_call_arguments(tool_call: object) -> dict[str, Any]:
 
 
 def _extract_tool_call_id(tool_call: object) -> str:
+    if hasattr(tool_call, "call_id"):
+        return str(getattr(tool_call, "call_id") or "")
+    if isinstance(tool_call, dict) and "call_id" in tool_call:
+        return str(tool_call.get("call_id") or "")
     if hasattr(tool_call, "id"):
         return str(getattr(tool_call, "id") or "")
     if isinstance(tool_call, dict):
@@ -92,27 +146,46 @@ def _extract_structured_tool_output(tool_result: Any) -> Any:
 
 
 def append_tool_trace_log(result: dict, provider: str) -> None:
-    tool_trace = result.get("tool_trace")
-    if not tool_trace:
+    tool_trace = result.get("tool_trace") or []
+    raw = result.get("raw")
+    output = _extract_response_output(raw) if provider in {"openai", "gwdg"} else []
+    reasoning = _extract_response_thinking(raw) if output else ""
+
+    if not tool_trace and not output:
         return
 
     tool_log = tool_log_context.get_active_tool_log()
     if tool_log is None:
         return
 
-    tool_log.append(
-        {
-            "name": "llm_tool_trace",
-            "arguments": {
-                "provider": provider,
-                "tool_call_count": len(tool_trace),
-            },
-            "response": {
-                "tool_trace": tool_trace,
-                "response_ids": result.get("response_ids", []),
-            },
-        }
-    )
+    if output:
+        message = result.get("message")
+        tool_log.append(
+            {
+                "name": "llm_response_metadata",
+                "arguments": {"provider": provider},
+                "response": {
+                    "output_item_types": [_field(item, "type", "") for item in output],
+                    "reasoning_summary_chars": len(reasoning),
+                    "has_message_thinking": isinstance(message, dict) and bool(message.get("thinking")),
+                },
+            }
+        )
+
+    if tool_trace:
+        tool_log.append(
+            {
+                "name": "llm_tool_trace",
+                "arguments": {
+                    "provider": provider,
+                    "tool_call_count": len(tool_trace),
+                },
+                "response": {
+                    "tool_trace": tool_trace,
+                    "response_ids": result.get("response_ids", []),
+                },
+            }
+        )
 
 
 def resolve_mcp_base_url(mcp_settings: dict[str, Any]) -> str:
@@ -171,18 +244,49 @@ def _tool_to_chat_tool(tool: Any) -> dict[str, Any]:
     }
 
 
-def _append_openai_tool_message(
-    messages: list[dict[str, Any]],
+def _tool_to_response_tool(tool: Any) -> dict[str, Any]:
+    return {
+        "type": "function",
+        "name": _tool_attr(tool, "name", "name"),
+        "description": _tool_attr(tool, "description", "description") or "",
+        "parameters": _tool_attr(tool, "input_schema", "inputSchema"),
+        "strict": False,
+    }
+
+
+def _extract_response_message(response: object) -> dict[str, str]:
+    output_text = response.get("output_text") if isinstance(response, dict) else getattr(response, "output_text", None)
+    message = {"role": "assistant", "content": output_text or ""}
+    thinking = _extract_response_thinking(response)
+    if thinking:
+        message["thinking"] = thinking
+    return message
+
+
+def _openai_reasoning() -> dict[str, str]:
+    if config.get_llm_thinking_enabled():
+        return {"effort": "high", "summary": "detailed"}
+    return {"effort": "none"}
+
+
+def _response_input_item(item: Any) -> Any:
+    if isinstance(item, dict):
+        return item
+    if hasattr(item, "model_dump"):
+        return item.model_dump(exclude_none=True)
+    return item
+
+
+def _append_response_tool_output(
+    items: list[dict[str, Any]],
     tool_call_id: str,
-    tool_name: str,
     structured_output: Any,
 ) -> None:
-    messages.append(
+    items.append(
         {
-            "role": "tool",
-            "tool_call_id": tool_call_id,
-            "name": tool_name,
-            "content": json.dumps(structured_output),
+            "type": "function_call_output",
+            "call_id": tool_call_id,
+            "output": json.dumps(structured_output),
         }
     )
 
@@ -217,36 +321,42 @@ async def _run_openai_compatible_tool_loop(
     client = openai_client or AsyncOpenAI(**client_kwargs)
 
     async def _handler(session, tools) -> dict:
-        chat_tools = [_tool_to_chat_tool(tool) for tool in tools]
-        request_messages = [dict(message) for message in messages]
+        response_tools = [_tool_to_response_tool(tool) for tool in tools]
+        request_input = [dict(message) for message in messages]
         response_ids: list[str] = []
         tool_trace: list[dict[str, Any]] = []
 
         while True:
             request_kwargs: dict[str, Any] = {
                 "model": settings.model,
-                "messages": request_messages,
-                "tools": chat_tools,
+                "input": request_input,
+                "tools": response_tools,
             }
             if settings.temperature is not None:
                 request_kwargs["temperature"] = settings.temperature
-
-            response = await client.chat.completions.create(**request_kwargs)
-            response_id = getattr(response, "id", "")
+            request_kwargs["reasoning"] = _openai_reasoning()
+
+            try:
+                response = await client.responses.create(**request_kwargs)
+            except BadRequestError as exc:
+                if "temperature" not in request_kwargs or not _is_unsupported_temperature_error(exc):
+                    raise
+                request_kwargs.pop("temperature")
+                response = await client.responses.create(**request_kwargs)
+            response_id = response.get("id", "") if isinstance(response, dict) else getattr(response, "id", "")
             if response_id:
                 response_ids.append(response_id)
-            message = response.choices[0].message if response.choices else {}
-            tool_calls = _extract_tool_calls(message)
+            tool_calls = _extract_response_tool_calls(response)
 
             if not tool_calls:
                 return {
                     "raw": response,
-                    "message": message,
+                    "message": _extract_response_message(response),
                     "tool_trace": tool_trace,
                     "response_ids": response_ids,
                 }
 
-            request_messages.append(_normalize_message(message))
+            request_input.extend(_response_input_item(item) for item in _extract_response_output(response))
             for tool_call in tool_calls:
                 tool_name = _extract_tool_call_name(tool_call)
                 arguments = _extract_tool_call_arguments(tool_call)
@@ -259,10 +369,9 @@ async def _run_openai_compatible_tool_loop(
                         "output": structured_output,
                     }
                 )
-                _append_openai_tool_message(
-                    request_messages,
+                _append_response_tool_output(
+                    request_input,
                     _extract_tool_call_id(tool_call),
-                    tool_name,
                     structured_output,
                 )
 

math-tutor/backend/test/test_llm_provider.py

@@ -25,6 +25,7 @@ ollama_stub.AsyncClient = object
 openai_stub = types.ModuleType("openai")
 openai_stub.OpenAI = object
 openai_stub.AsyncOpenAI = object
+openai_stub.BadRequestError = RuntimeError
 
 mistralai_stub = types.ModuleType("mistralai")
 mistralai_client_stub = types.ModuleType("mistralai.client")
@@ -90,50 +91,84 @@ from app.deterministic_services.orchestrators.orchestrator_base import ChatState
 MESSAGES = [{"role": "user", "content": "Hallo"}]
 
 
-def _make_openai_response(message, response_id: str = "resp-1", total_tokens: int = 7):
+def _make_responses_response(
+    output_text: str = "",
+    *,
+    output: list[object] | None = None,
+    response_id: str = "resp-1",
+    total_tokens: int = 7,
+):
     usage = types.SimpleNamespace(total_tokens=total_tokens)
-    choice = types.SimpleNamespace(message=message)
-    return types.SimpleNamespace(id=response_id, choices=[choice], usage=usage)
+    return types.SimpleNamespace(
+        id=response_id,
+        output_text=output_text,
+        output=output or [],
+        usage=usage,
+    )
 
 
-def _make_openai_tool_call(
+def _make_response_tool_call(
     name: str,
     arguments: str | dict,
     tool_call_id: str = "call-1",
 ):
-    function = types.SimpleNamespace(name=name, arguments=arguments)
-    return types.SimpleNamespace(id=tool_call_id, function=function)
+    return types.SimpleNamespace(
+        type="function_call",
+        call_id=tool_call_id,
+        name=name,
+        arguments=arguments,
+    )
+
 
+def _make_reasoning_item(text: str):
+    return types.SimpleNamespace(
+        type="reasoning",
+        summary=[types.SimpleNamespace(type="summary_text", text=text)],
+    )
 
-class _FakeOpenAICompletions:
+
+class _FakeOpenAIResponses:
     def __init__(self, responses: list[object]) -> None:
         self._responses = list(responses)
         self.calls: list[dict] = []
 
     async def create(self, **kwargs):
         self.calls.append(kwargs)
-        return self._responses.pop(0)
+        response = self._responses.pop(0)
+        if isinstance(response, Exception):
+            raise response
+        return response
 
 
 class _FakeOpenAIClient:
     def __init__(self, responses: list[object]) -> None:
-        completions = _FakeOpenAICompletions(responses)
-        self.chat = types.SimpleNamespace(completions=completions)
+        self.responses = _FakeOpenAIResponses(responses)
 
 
 class _FakeSyncOpenAIClient:
     def __init__(self, response: object) -> None:
         self.calls: list[dict] = []
-        self.chat = types.SimpleNamespace(
-            completions=types.SimpleNamespace(create=self._create)
-        )
+        self.responses = types.SimpleNamespace(create=self._create)
         self._response = response
 
     def _create(self, **kwargs):
         self.calls.append(kwargs)
+        if isinstance(self._response, list):
+            response = self._response.pop(0)
+            if isinstance(response, Exception):
+                raise response
+            return response
+        if isinstance(self._response, Exception):
+            raise self._response
         return self._response
 
 
+class _FakeBadRequestError(RuntimeError):
+    def __init__(self, message: str, param: str | None = None) -> None:
+        super().__init__(message)
+        self.param = param
+
+
 class _FakeMistralClient:
     calls: list[dict] = []
 
@@ -489,14 +524,83 @@ class LLMClientProviderTest(unittest.TestCase):
             timeout=30.0,
             temperature=None,
         )
-        fake_client = _FakeSyncOpenAIClient(_make_openai_response({"content": "done"}))
+        fake_client = _FakeSyncOpenAIClient(_make_responses_response("done"))
 
         with patch.dict(os.environ, {"LLM_THINKING_ENABLED": "true"}), patch.object(
             llm_client, "OpenAI", return_value=fake_client
         ):
             llm_client._chat_openai_compatible(MESSAGES, settings)
 
-        self.assertEqual(fake_client.calls[0]["reasoning_effort"], "medium")
+        self.assertEqual(fake_client.calls[0]["reasoning"], {"effort": "high", "summary": "detailed"})
+
+    def test_openai_compatible_chat_sends_system_as_instructions(self) -> None:
+        """Checks that plain Responses input uses text so reasoning summaries are returned."""
+        settings = config.OpenAIChatSettings(
+            base_url="https://example.com/v1",
+            api_key="key",
+            model="model",
+            timeout=30.0,
+            temperature=None,
+        )
+        fake_client = _FakeSyncOpenAIClient(_make_responses_response("done"))
+
+        with patch.object(llm_client, "OpenAI", return_value=fake_client):
+            llm_client._chat_openai_compatible(
+                [
+                    {"role": "system", "content": "Be precise."},
+                    {"role": "user", "content": "Hallo"},
+                    {"role": "assistant", "content": "Hi"},
+                ],
+                settings,
+            )
+
+        self.assertEqual(fake_client.calls[0]["instructions"], "Be precise.")
+        self.assertEqual(fake_client.calls[0]["input"], "user: Hallo\n\nassistant: Hi")
+
+    def test_openai_compatible_chat_formats_reasoning_as_thinking(self) -> None:
+        """Checks that Responses reasoning summaries use the existing thinking format."""
+        settings = config.OpenAIChatSettings(
+            base_url="https://example.com/v1",
+            api_key="key",
+            model="model",
+            timeout=30.0,
+            temperature=None,
+        )
+        fake_client = _FakeSyncOpenAIClient(
+            _make_responses_response("answer", output=[_make_reasoning_item("work")])
+        )
+
+        with patch.dict(os.environ, {"LLM_THINKING_ENABLED": "true"}), patch.object(
+            llm_client, "OpenAI", return_value=fake_client
+        ):
+            result = llm_client._chat_openai_compatible(MESSAGES, settings)
+
+        self.assertEqual(llm_client.get_message_content(result), "work</think>\n\nanswer")
+
+    def test_openai_compatible_chat_retries_without_unsupported_temperature(self) -> None:
+        """Checks that Responses models which reject temperature still work."""
+        settings = config.OpenAIChatSettings(
+            base_url="https://example.com/v1",
+            api_key="key",
+            model="model",
+            timeout=30.0,
+            temperature=0.2,
+        )
+        fake_client = _FakeSyncOpenAIClient(
+            [
+                _FakeBadRequestError("Unsupported parameter: 'temperature'", "temperature"),
+                _make_responses_response("done"),
+            ]
+        )
+
+        with patch.object(llm_client, "BadRequestError", _FakeBadRequestError), patch.object(
+            llm_client, "OpenAI", return_value=fake_client
+        ):
+            result = llm_client._chat_openai_compatible(MESSAGES, settings)
+
+        self.assertEqual(result["message"]["content"], "done")
+        self.assertEqual(fake_client.calls[0]["temperature"], 0.2)
+        self.assertNotIn("temperature", fake_client.calls[1])
 
     def test_mistral_chat_sets_reasoning_effort_from_env(self) -> None:
         """Checks that Mistral chat maps thinking mode to reasoning effort."""
@@ -663,8 +767,7 @@ class LLMClientToolHelpersTest(unittest.TestCase):
             timeout=30.0,
             temperature=0.2,
         )
-        final_message = types.SimpleNamespace(role="assistant", content="done", tool_calls=None)
-        fake_client = _FakeOpenAIClient([_make_openai_response(final_message)])
+        fake_client = _FakeOpenAIClient([_make_responses_response("done")])
 
         async def run_test():
             async def fake_with_mcp_session(mcp_settings, handler, http_client=None):
@@ -677,7 +780,7 @@ class LLMClientToolHelpersTest(unittest.TestCase):
                     openai_client=fake_client,
                 )
 
-            self.assertEqual(result["message"].content, "done")
+            self.assertEqual(result["message"]["content"], "done")
             self.assertEqual(result["response_ids"], ["resp-1"])
             self.assertEqual(result["tool_trace"], [])
 
@@ -693,17 +796,11 @@ class LLMClientToolHelpersTest(unittest.TestCase):
             timeout=30.0,
             temperature=0.2,
         )
-        tool_call = _make_openai_tool_call("solve", "{\"x\": 2}")
-        tool_message = types.SimpleNamespace(
-            role="assistant",
-            content="",
-            tool_calls=[tool_call],
-        )
-        final_message = types.SimpleNamespace(role="assistant", content="x=2", tool_calls=None)
+        tool_call = _make_response_tool_call("solve", "{\"x\": 2}")
         fake_client = _FakeOpenAIClient(
             [
-                _make_openai_response(tool_message, response_id="resp-1"),
-                _make_openai_response(final_message, response_id="resp-2"),
+                _make_responses_response(output=[tool_call], response_id="resp-1"),
+                _make_responses_response("x=2", response_id="resp-2"),
             ]
         )
         fake_tool = types.SimpleNamespace(
@@ -724,16 +821,81 @@ class LLMClientToolHelpersTest(unittest.TestCase):
                     openai_client=fake_client,
                 )
 
-            self.assertEqual(result["message"].content, "x=2")
+            self.assertEqual(result["message"]["content"], "x=2")
             self.assertEqual(result["response_ids"], ["resp-1", "resp-2"])
             self.assertEqual(
                 result["tool_trace"],
                 [{"name": "solve", "arguments": {"x": 2}, "output": {"result": 2}}],
             )
             self.assertEqual(session.calls, [("solve", {"x": 2})])
-            second_call_messages = fake_client.chat.completions.calls[1]["messages"]
-            self.assertEqual(second_call_messages[-1]["role"], "tool")
-            self.assertEqual(second_call_messages[-1]["tool_call_id"], "call-1")
+            second_call_input = fake_client.responses.calls[1]["input"]
+            self.assertEqual(second_call_input[-1]["type"], "function_call_output")
+            self.assertEqual(second_call_input[-1]["call_id"], "call-1")
+
+        asyncio_result = __import__("asyncio").run(run_test())
+        self.assertIsNone(asyncio_result)
+
+    def test_openai_tool_loop_formats_reasoning_as_thinking(self) -> None:
+        """Checks that tool-mode Responses reasoning summaries are preserved."""
+        settings = config.OpenAIChatSettings(
+            base_url="https://example.com/v1",
+            api_key="key",
+            model="model",
+            timeout=30.0,
+            temperature=0.2,
+        )
+        fake_client = _FakeOpenAIClient(
+            [_make_responses_response("done", output=[_make_reasoning_item("checked")])]
+        )
+
+        async def run_test():
+            async def fake_with_mcp_session(mcp_settings, handler, http_client=None):
+                return await handler(_FakeSession([]), [])
+
+            with patch.object(llm_tool_client, "with_mcp_session", side_effect=fake_with_mcp_session):
+                result = await llm_tool_client._run_openai_compatible_tool_loop(
+                    MESSAGES,
+                    settings,
+                    openai_client=fake_client,
+                )
+
+            self.assertEqual(llm_client.get_message_content(result), "checked</think>\n\ndone")
+
+        asyncio_result = __import__("asyncio").run(run_test())
+        self.assertIsNone(asyncio_result)
+
+    def test_openai_tool_loop_retries_without_unsupported_temperature(self) -> None:
+        """Checks that the tool path also tolerates models rejecting temperature."""
+        settings = config.OpenAIChatSettings(
+            base_url="https://example.com/v1",
+            api_key="key",
+            model="model",
+            timeout=30.0,
+            temperature=0.2,
+        )
+        fake_client = _FakeOpenAIClient(
+            [
+                _FakeBadRequestError("Unsupported parameter: 'temperature'", "temperature"),
+                _make_responses_response("done"),
+            ]
+        )
+
+        async def run_test():
+            async def fake_with_mcp_session(mcp_settings, handler, http_client=None):
+                return await handler(_FakeSession([]), [])
+
+            with patch.object(llm_tool_client, "BadRequestError", _FakeBadRequestError), patch.object(
+                llm_tool_client, "with_mcp_session", side_effect=fake_with_mcp_session
+            ):
+                result = await llm_tool_client._run_openai_compatible_tool_loop(
+                    MESSAGES,
+                    settings,
+                    openai_client=fake_client,
+                )
+
+            self.assertEqual(result["message"]["content"], "done")
+            self.assertEqual(fake_client.responses.calls[0]["temperature"], 0.2)
+            self.assertNotIn("temperature", fake_client.responses.calls[1])
 
         asyncio_result = __import__("asyncio").run(run_test())
         self.assertIsNone(asyncio_result)
@@ -748,15 +910,13 @@ class LLMClientToolHelpersTest(unittest.TestCase):
             temperature=0.2,
         )
         tool_calls = [
-            _make_openai_tool_call("first", {"a": 1}, tool_call_id="call-1"),
-            _make_openai_tool_call("second", "{\"b\": 2}", tool_call_id="call-2"),
+            _make_response_tool_call("first", {"a": 1}, tool_call_id="call-1"),
+            _make_response_tool_call("second", "{\"b\": 2}", tool_call_id="call-2"),
         ]
-        tool_message = types.SimpleNamespace(role="assistant", content="", tool_calls=tool_calls)
-        final_message = types.SimpleNamespace(role="assistant", content="done", tool_calls=None)
         fake_client = _FakeOpenAIClient(
             [
-                _make_openai_response(tool_message, response_id="resp-1"),
-                _make_openai_response(final_message, response_id="resp-2"),
+                _make_responses_response(output=tool_calls, response_id="resp-1"),
+                _make_responses_response("done", response_id="resp-2"),
             ]
         )
         fake_tools = [
@@ -781,7 +941,7 @@ class LLMClientToolHelpersTest(unittest.TestCase):
                     openai_client=fake_client,
                 )
 
-            self.assertEqual(result["message"].content, "done")
+            self.assertEqual(result["message"]["content"], "done")
             self.assertEqual(
                 result["tool_trace"],
                 [