feat(deepseek): update reasoner model for V3.2 tool call support

daniel7an · daniel7an · commit 41facbca82e0 · 2026-02-09T19:16:11.000Z
- Remove 'tools' from reasoner unsupported params since DeepSeek V3.2 reasoner now supports tool calls with thinking mode - Fix bug where copy.deepcopy overwrote the filtered config for reasoner models, making param filtering ineffective - Add reasoning_content injection/extraction for tool call continuations (required by DeepSeek API for multi-turn tool calls in thinking mode) - Fix typo in constant name (REASONSER -> REASONER) - Update documentation URL to point to current thinking_mode guide Closes #3811
diff --git a/camel/models/deepseek_model.py b/camel/models/deepseek_model.py
@@ -49,14 +49,13 @@
 
 logger = get_logger(__name__)
 
-REASONSER_UNSUPPORTED_PARAMS = [
+REASONER_UNSUPPORTED_PARAMS = [
     "temperature",
     "top_p",
     "presence_penalty",
     "frequency_penalty",
     "logprobs",
     "top_logprobs",
-    "tools",
 ]
 
 
@@ -124,32 +123,97 @@ def __init__(
             max_retries=max_retries,
             **kwargs,
         )
+        # Store the last reasoning_content from model response.
+        # For DeepSeek reasoner models with tool calls, the
+        # reasoning_content must be passed back in subsequent requests.
+        self._last_reasoning_content: Optional[str] = None
+
+    def _inject_reasoning_content(
+        self,
+        messages: List[OpenAIMessage],
+    ) -> List[OpenAIMessage]:
+        r"""Inject the last reasoning_content into assistant messages.
+
+        For DeepSeek reasoner models with tool call support, the
+        reasoning_content from the model response needs to be passed back
+        in subsequent requests for proper context management.
+
+        Args:
+            messages: The original messages list.
+
+        Returns:
+            Messages with reasoning_content added to the last assistant
+            message that has tool_calls.
+        """
+        if not self._last_reasoning_content:
+            return messages
+
+        # Find the last assistant message with tool_calls and inject
+        # reasoning_content
+        processed: List[OpenAIMessage] = []
+        reasoning_injected = False
+
+        for msg in reversed(messages):
+            if (
+                not reasoning_injected
+                and isinstance(msg, dict)
+                and msg.get("role") == "assistant"
+                and msg.get("tool_calls")
+                and "reasoning_content" not in msg
+            ):
+                new_msg = dict(msg)
+                new_msg["reasoning_content"] = self._last_reasoning_content
+                processed.append(new_msg)  # type: ignore[arg-type]
+                reasoning_injected = True
+            else:
+                processed.append(msg)
+
+        if reasoning_injected:
+            self._last_reasoning_content = None
+
+        return list(reversed(processed))
+
+    def _extract_reasoning_content(
+        self, response: ChatCompletion
+    ) -> Optional[str]:
+        r"""Extract reasoning_content from the model response.
+
+        Args:
+            response: The model response.
+
+        Returns:
+            The reasoning_content if available, None otherwise.
+        """
+        if response.choices:
+            return getattr(
+                response.choices[0].message, "reasoning_content", None
+            )
+        return None
 
     def _prepare_request(
         self,
         messages: List[OpenAIMessage],
         response_format: Optional[Type[BaseModel]] = None,
         tools: Optional[List[Dict[str, Any]]] = None,
     ) -> Dict[str, Any]:
-        request_config = self.model_config_dict.copy()
+        import copy
+
+        request_config = copy.deepcopy(self.model_config_dict)
 
         if self.model_type in [
             ModelType.DEEPSEEK_REASONER,
         ]:
             logger.warning(
-                "Warning: You are using an DeepSeek Reasoner model, "
+                "Warning: You are using a DeepSeek Reasoner model, "
                 "which has certain limitations, reference: "
-                "`https://api-docs.deepseek.com/guides/reasoning_model"
+                "`https://api-docs.deepseek.com/guides/thinking_mode"
                 "#api-parameters`.",
             )
             request_config = {
                 key: value
                 for key, value in request_config.items()
-                if key not in REASONSER_UNSUPPORTED_PARAMS
+                if key not in REASONER_UNSUPPORTED_PARAMS
             }
-        import copy
-
-        request_config = copy.deepcopy(self.model_config_dict)
         # Remove strict from each tool's function parameters since DeepSeek
         # does not support them
         if tools:
@@ -183,6 +247,10 @@ def _run(
         """
         self._log_and_trace()
 
+        # Inject reasoning_content for reasoner tool call continuations
+        if self.model_type in [ModelType.DEEPSEEK_REASONER]:
+            messages = self._inject_reasoning_content(messages)
+
         request_config = self._prepare_request(
             messages, response_format, tools
         )
@@ -193,6 +261,12 @@ def _run(
             **request_config,
         )
 
+        # Extract and store reasoning_content for next request
+        if isinstance(response, ChatCompletion):
+            self._last_reasoning_content = self._extract_reasoning_content(
+                response
+            )
+
         return response
 
     @observe()
@@ -215,6 +289,10 @@ async def _arun(
         """
         self._log_and_trace()
 
+        # Inject reasoning_content for reasoner tool call continuations
+        if self.model_type in [ModelType.DEEPSEEK_REASONER]:
+            messages = self._inject_reasoning_content(messages)
+
         request_config = self._prepare_request(
             messages, response_format, tools
         )
@@ -224,4 +302,10 @@ async def _arun(
             **request_config,
         )
 
+        # Extract and store reasoning_content for next request
+        if isinstance(response, ChatCompletion):
+            self._last_reasoning_content = self._extract_reasoning_content(
+                response
+            )
+
         return response
diff --git a/test/models/test_deepseek_model.py b/test/models/test_deepseek_model.py
@@ -54,3 +54,99 @@ def test_deepseek_model_create(model_type: ModelType):
         model_config_dict=DeepSeekConfig(temperature=1.3).as_dict(),
     )
     assert model.model_type == model_type
+
+
+@pytest.mark.model_backend
+def test_deepseek_reasoner_prepare_request_filters_params():
+    r"""Test that reasoner model filters unsupported parameters."""
+    model_config_dict = DeepSeekConfig(temperature=0.5, top_p=0.9).as_dict()
+    model = DeepSeekModel(ModelType.DEEPSEEK_REASONER, model_config_dict)
+    messages = [{"role": "user", "content": "Hello"}]
+    request_config = model._prepare_request(messages)
+    assert "temperature" not in request_config
+    assert "top_p" not in request_config
+
+
+@pytest.mark.model_backend
+def test_deepseek_reasoner_allows_tools():
+    r"""Test that reasoner model supports tool calls (V3.2+)."""
+    model_config_dict = DeepSeekConfig().as_dict()
+    model = DeepSeekModel(ModelType.DEEPSEEK_REASONER, model_config_dict)
+    messages = [{"role": "user", "content": "Hello"}]
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "description": "Get the weather",
+                "parameters": {"type": "object", "properties": {}},
+                "strict": True,
+            },
+        }
+    ]
+    request_config = model._prepare_request(messages, tools=tools)
+    assert "tools" in request_config
+    # Verify strict is removed from function parameters
+    assert "strict" not in request_config["tools"][0]["function"]
+
+
+@pytest.mark.model_backend
+def test_deepseek_chat_prepare_request_keeps_params():
+    r"""Test that chat model keeps all parameters."""
+    model_config_dict = DeepSeekConfig(temperature=0.5, top_p=0.9).as_dict()
+    model = DeepSeekModel(ModelType.DEEPSEEK_CHAT, model_config_dict)
+    messages = [{"role": "user", "content": "Hello"}]
+    request_config = model._prepare_request(messages)
+    assert request_config.get("temperature") == 0.5
+    assert request_config.get("top_p") == 0.9
+
+
+@pytest.mark.model_backend
+def test_deepseek_reasoning_content_injection():
+    r"""Test reasoning_content injection for tool call continuations."""
+    model_config_dict = DeepSeekConfig().as_dict()
+    model = DeepSeekModel(ModelType.DEEPSEEK_REASONER, model_config_dict)
+
+    # Simulate stored reasoning_content from a previous response
+    model._last_reasoning_content = "Let me think about this..."
+
+    messages = [
+        {"role": "user", "content": "What's the weather?"},
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {
+                        "name": "get_weather",
+                        "arguments": '{"city": "Beijing"}',
+                    },
+                }
+            ],
+        },
+        {
+            "role": "tool",
+            "content": "Sunny, 25°C",
+            "tool_call_id": "call_1",
+        },
+    ]
+
+    processed = model._inject_reasoning_content(messages)
+
+    # reasoning_content should be injected into the assistant message
+    assert processed[1]["reasoning_content"] == "Let me think about this..."
+    # Should be cleared after injection
+    assert model._last_reasoning_content is None
+
+
+@pytest.mark.model_backend
+def test_deepseek_no_injection_without_reasoning_content():
+    r"""Test that no injection happens when there's no reasoning_content."""
+    model_config_dict = DeepSeekConfig().as_dict()
+    model = DeepSeekModel(ModelType.DEEPSEEK_REASONER, model_config_dict)
+
+    messages = [{"role": "user", "content": "Hello"}]
+    processed = model._inject_reasoning_content(messages)
+    assert processed == messages
