fix: preserve deepseek reasoning content in tool loops

tests/test_langchain_deepseek_compat.py

@@ -0,0 +1,146 @@
+import importlib.util
+import sys
+import unittest
+from pathlib import Path
+from types import ModuleType
+
+from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
+
+
+def _stub_module(name: str, **attrs):
+    module = sys.modules.get(name)
+    if module is None:
+        module = ModuleType(name)
+        sys.modules[name] = module
+    for key, value in attrs.items():
+        setattr(module, key, value)
+    return module
+
+
+class _DummyLogger:
+    def __getattr__(self, _name):
+        return lambda *args, **kwargs: None
+
+
+def _build_tool_call(name: str = "search", arguments: str = "{}"):
+    return [
+        {
+            "id": "call_1",
+            "type": "tool_call",
+            "name": name,
+            "args": {},
+        }
+    ]
+
+
+class _FakeChatDeepSeek:
+    def __init__(self, model_name: str, model_kwargs: dict | None = None):
+        self.model_name = model_name
+        self.model_kwargs = model_kwargs or {}
+
+    def _get_request_payload(self, input_, *, stop=None, **kwargs):
+        messages = []
+        for message in input_:
+            payload_message = {
+                "role": message.type,
+                "content": message.content,
+            }
+            if message.type == "human":
+                payload_message["role"] = "user"
+            elif message.type == "ai":
+                payload_message["role"] = "assistant"
+                tool_calls = getattr(message, "tool_calls", None)
+                if tool_calls:
+                    payload_message["tool_calls"] = tool_calls
+            elif message.type == "tool":
+                payload_message["role"] = "tool"
+                payload_message["tool_call_id"] = message.tool_call_id
+            messages.append(payload_message)
+        return {"messages": messages}
+
+
+_ORIGINAL_GET_REQUEST_PAYLOAD = _FakeChatDeepSeek._get_request_payload
+
+
+sys.modules.pop("app.helper.llm", None)
+_stub_module(
+    "app.core.config",
+    settings=ModuleType("settings"),
+)
+sys.modules["app.core.config"].settings.LLM_PROVIDER = "deepseek"
+sys.modules["app.core.config"].settings.LLM_MODEL = "deepseek-v4-pro"
+sys.modules["app.core.config"].settings.LLM_API_KEY = "sk-test"
+sys.modules["app.core.config"].settings.LLM_BASE_URL = "https://api.deepseek.com"
+sys.modules["app.core.config"].settings.LLM_THINKING_LEVEL = None
+sys.modules["app.core.config"].settings.LLM_DISABLE_THINKING = False
+sys.modules["app.core.config"].settings.LLM_REASONING_EFFORT = None
+sys.modules["app.core.config"].settings.LLM_TEMPERATURE = 0.1
+sys.modules["app.core.config"].settings.LLM_MAX_CONTEXT_TOKENS = 64
+sys.modules["app.core.config"].settings.PROXY_HOST = None
+_stub_module("app.log", logger=_DummyLogger())
+_stub_module("langchain_deepseek", ChatDeepSeek=_FakeChatDeepSeek)
+
+module_path = Path(__file__).resolve().parents[1] / "app" / "helper" / "llm.py"
+spec = importlib.util.spec_from_file_location("test_llm_module_for_deepseek_compat", module_path)
+llm_module = importlib.util.module_from_spec(spec)
+assert spec and spec.loader
+spec.loader.exec_module(llm_module)
+
+
+class DeepSeekCompatPatchTest(unittest.TestCase):
+    def setUp(self):
+        _FakeChatDeepSeek._get_request_payload = _ORIGINAL_GET_REQUEST_PAYLOAD
+        if hasattr(_FakeChatDeepSeek, "_moviepilot_reasoning_content_patched"):
+            delattr(_FakeChatDeepSeek, "_moviepilot_reasoning_content_patched")
+        llm_module._patch_deepseek_reasoning_content_support()
+
+    def test_injects_reasoning_content_for_assistant_tool_calls(self):
+        llm = _FakeChatDeepSeek("deepseek-v4-pro")
+        messages = [
+            HumanMessage(content="天气如何？"),
+            AIMessage(
+                content="",
+                tool_calls=_build_tool_call(),
+                additional_kwargs={"reasoning_content": "先调用天气工具"},
+            ),
+            ToolMessage(content="晴天", tool_call_id="call_1"),
+        ]
+
+        payload = llm._get_request_payload(messages)
+
+        self.assertEqual(
+            payload["messages"][1]["reasoning_content"],
+            "先调用天气工具",
+        )
+
+    def test_falls_back_to_empty_reasoning_content_when_missing(self):
+        llm = _FakeChatDeepSeek("deepseek-v4-flash")
+        messages = [
+            HumanMessage(content="天气如何？"),
+            AIMessage(content="", tool_calls=_build_tool_call()),
+            ToolMessage(content="晴天", tool_call_id="call_1"),
+        ]
+
+        payload = llm._get_request_payload(messages)
+
+        self.assertIn("reasoning_content", payload["messages"][1])
+        self.assertEqual(payload["messages"][1]["reasoning_content"], "")
+
+    def test_skips_injection_when_thinking_is_disabled(self):
+        llm = _FakeChatDeepSeek(
+            "deepseek-v4-pro",
+            model_kwargs={"extra_body": {"thinking": {"type": "disabled"}}},
+        )
+        messages = [
+            HumanMessage(content="天气如何？"),
+            AIMessage(
+                content="",
+                tool_calls=_build_tool_call(),
+                additional_kwargs={"reasoning_content": "先调用天气工具"},
+            ),
+            ToolMessage(content="晴天", tool_call_id="call_1"),
+        ]
+
+        payload = llm._get_request_payload(messages)
+
+        self.assertNotIn("reasoning_content", payload["messages"][1])

tests/test_llm_helper_testcall.py

@@ -144,6 +144,7 @@ class LlmHelperTestCallTest(unittest.TestCase):
 
     def test_get_llm_uses_deepseek_thinking_level_controls(self):
         calls = []
+        patch_calls = []
 
         class _FakeChatDeepSeek:
             def __init__(self, **kwargs):
@@ -154,6 +155,10 @@ class LlmHelperTestCallTest(unittest.TestCase):
         with patch.dict(
             sys.modules,
             {"langchain_deepseek": SimpleNamespace(ChatDeepSeek=_FakeChatDeepSeek)},
+        ), patch.object(
+            llm_module,
+            "_patch_deepseek_reasoning_content_support",
+            side_effect=lambda: patch_calls.append(True),
         ):
             llm_module.LLMHelper.get_llm(
                 provider="deepseek",
@@ -168,11 +173,13 @@ class LlmHelperTestCallTest(unittest.TestCase):
             calls[0].get("extra_body"),
             {"thinking": {"type": "enabled"}},
         )
+        self.assertEqual(patch_calls, [True])
         self.assertEqual(calls[0].get("reasoning_effort"), "max")
         self.assertEqual(calls[0].get("api_base"), "https://api.deepseek.com")
 
     def test_get_llm_disables_deepseek_thinking_via_thinking_level(self):
         calls = []
+        patch_calls = []
 
         class _FakeChatDeepSeek:
             def __init__(self, **kwargs):
@@ -183,6 +190,10 @@ class LlmHelperTestCallTest(unittest.TestCase):
         with patch.dict(
             sys.modules,
             {"langchain_deepseek": SimpleNamespace(ChatDeepSeek=_FakeChatDeepSeek)},
+        ), patch.object(
+            llm_module,
+            "_patch_deepseek_reasoning_content_support",
+            side_effect=lambda: patch_calls.append(True),
         ):
             llm_module.LLMHelper.get_llm(
                 provider="deepseek",
@@ -197,6 +208,7 @@ class LlmHelperTestCallTest(unittest.TestCase):
             calls[0].get("extra_body"),
             {"thinking": {"type": "disabled"}},
         )
+        self.assertEqual(patch_calls, [True])
         self.assertIsNone(calls[0].get("reasoning_effort"))
         self.assertEqual(calls[0].get("api_base"), "https://proxy.example.com")