fix(model): support DeepSeek V4 Pro (#3079)

configs/dbgpt-proxy-deepseek.toml

@@ -23,9 +23,12 @@ persist_path = "pilot/data"
 # Model Configurations
 [models]
 [[models.llms]]
-name = "deepseek-reasoner"
+name = "deepseek-v4-pro"
+# name = "deepseek-reasoner"
 # name = "deepseek-chat"
 provider = "proxy/deepseek"
+# Disable V4-Pro thinking mode so ReAct responses stay parseable.
+thinking_enabled = false
 api_key = "your_deepseek_api_key"
 
 [[models.embeddings]]

docs/docs/getting-started/providers/deepseek.md

@@ -45,9 +45,11 @@ Edit `configs/dbgpt-proxy-deepseek.toml`:
 ```toml
 [models]
 [[models.llms]]
-name = "deepseek-reasoner"
+name = "deepseek-v4-pro"
 provider = "proxy/deepseek"
 api_key = "your-deepseek-api-key"
+# Disable V4-Pro thinking mode for strict ReAct output parsing.
+thinking_enabled = false
 
 [[models.embeddings]]
 name = "BAAI/bge-large-zh-v1.5"
@@ -60,9 +62,14 @@ provider = "hf"
 
 | Model | Config name | Notes |
 |---|---|---|
+| DeepSeek-V4-Pro | `deepseek-v4-pro` | 1M context, advanced reasoning, coding, and agent tasks |
 | DeepSeek-R1 | `deepseek-reasoner` | Strong reasoning, chain-of-thought |
 | DeepSeek-V3 | `deepseek-chat` | General purpose chat |
 
+For ReAct agents, keep `thinking_enabled = false` with `deepseek-v4-pro`. DeepSeek
+V4-Pro enables thinking mode by default, which can add reasoning blocks before the
+strict `Thought/Action/Action Input` response format.
+
 ## Start the server
 
 ```bash

docs/docs/getting-started/troubleshooting/llm.md

@@ -44,7 +44,7 @@ export OPENAI_API_KEY="sk-your-actual-key"
 | Provider | Example Name |
 |---|---|
 | OpenAI | `chatgpt_proxyllm`, `gpt-4o` |
-| DeepSeek | `deepseek-chat`, `deepseek-reasoner` |
+| DeepSeek | `deepseek-v4-pro`, `deepseek-chat`, `deepseek-reasoner` |
 | Ollama | `qwen2.5:latest` (must be pulled first) |
 | HuggingFace | `THUDM/glm-4-9b-chat-hf` |
 

docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/getting-started/providers/deepseek.md

@@ -45,9 +45,11 @@ uv sync --all-packages \
 ```toml
 [models]
 [[models.llms]]
-name = "deepseek-reasoner"
+name = "deepseek-v4-pro"
 provider = "proxy/deepseek"
 api_key = "your-deepseek-api-key"
+# 为严格的 ReAct 输出解析关闭 V4-Pro 思考模式。
+thinking_enabled = false
 
 [[models.embeddings]]
 name = "BAAI/bge-large-zh-v1.5"
@@ -60,9 +62,14 @@ provider = "hf"
 
 | 模型 | 配置名 | 说明 |
 |---|---|---|
+| DeepSeek-V4-Pro | `deepseek-v4-pro` | 1M 上下文，适合高级推理、代码与 Agent 任务 |
 | DeepSeek-R1 | `deepseek-reasoner` | 推理能力强，适合复杂思考任务 |
 | DeepSeek-V3 | `deepseek-chat` | 通用聊天与问答 |
 
+ReAct Agent 使用 `deepseek-v4-pro` 时建议保留 `thinking_enabled = false`。
+DeepSeek V4-Pro 默认开启思考模式，可能在严格的
+`Thought/Action/Action Input` 输出格式前产生推理块，导致解析失败。
+
 ## 启动服务
 
 ```bash

docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/getting-started/troubleshooting/llm.md

@@ -44,7 +44,7 @@ export OPENAI_API_KEY="sk-your-actual-key"
 | Provider | 示例名称 |
 |---|---|
 | OpenAI | `chatgpt_proxyllm`, `gpt-4o` |
-| DeepSeek | `deepseek-chat`, `deepseek-reasoner` |
+| DeepSeek | `deepseek-v4-pro`, `deepseek-chat`, `deepseek-reasoner` |
 | Ollama | `qwen2.5:latest`（需先拉取） |
 | HuggingFace | `THUDM/glm-4-9b-chat-hf` |
 

packages/dbgpt-core/src/dbgpt/agent/util/react_parser.py

@@ -93,9 +93,7 @@ class ReActOutputParser:
 
     def _find_prefix_matches(self, text: str, escaped_prefix: str) -> List[re.Match]:
         """Find line-start ReAct prefix matches outside markdown code fences."""
-        pattern = re.compile(
-            self._prefix_line_pattern(escaped_prefix), re.MULTILINE
-        )
+        pattern = re.compile(self._prefix_line_pattern(escaped_prefix), re.MULTILINE)
         fence_spans = self._markdown_fence_spans(text)
         return [
             match
@@ -196,9 +194,7 @@ class ReActOutputParser:
         text = self._normalize_react_text(text).strip()
 
         # Find all line-start instances of the thought prefix outside code fences.
-        thought_matches = self._find_prefix_matches(
-            text, self.thought_prefix_escaped
-        )
+        thought_matches = self._find_prefix_matches(text, self.thought_prefix_escaped)
 
         if not thought_matches:
             return []
@@ -272,12 +268,8 @@ class ReActOutputParser:
             self.action_reason_prefix_escaped
         )
         action_line = self._prefix_line_pattern(self.action_prefix_escaped)
-        action_input_line = self._prefix_line_pattern(
-            self.action_input_prefix_escaped
-        )
-        observation_line = self._prefix_line_pattern(
-            self.observation_prefix_escaped
-        )
+        action_input_line = self._prefix_line_pattern(self.action_input_prefix_escaped)
+        observation_line = self._prefix_line_pattern(self.observation_prefix_escaped)
 
         thought_match = re.search(
             rf"{thought_line}(.*?)(?={phase_line}|{action_intention_line}|"

packages/dbgpt-core/src/dbgpt/agent/util/tests/test_react_parser.py

@@ -55,10 +55,7 @@ Action Input: {"sql": "select distinct major_type from assets"}"""
         assert len(steps) == 1
         assert steps[0].thought == "I need to inspect the database schema."
         assert steps[0].action_intention == "Query the available asset categories."
-        assert (
-            steps[0].action_reason
-            == "The user asked about assets by major type."
-        )
+        assert steps[0].action_reason == "The user asked about assets by major type."
         assert steps[0].action == "query_db"
         assert steps[0].action_input == {
             "sql": "select distinct major_type from assets"
@@ -486,9 +483,7 @@ Action Input: {"result": "网络设备资产共 30 条"}"""
         assert len(all_steps) == 2
         assert len(current_steps) == 1
         assert current_steps[0].action == "query_db"
-        assert current_steps[0].action_input == {
-            "sql": "select count(*) from assets"
-        }
+        assert current_steps[0].action_input == {"sql": "select count(*) from assets"}
 
     def test_react_markers_inside_action_input_code_fence_are_not_steps(self):
         """Do not split Action Input code fences that mention ReAct labels."""

packages/dbgpt-core/src/dbgpt/model/proxy/llms/deepseek.py

@@ -1,8 +1,18 @@
 import os
 from dataclasses import dataclass, field
-from typing import TYPE_CHECKING, Any, Dict, Optional, Type, Union, cast
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    AsyncIterator,
+    Dict,
+    List,
+    Optional,
+    Type,
+    Union,
+    cast,
+)
 
-from dbgpt.core import ModelMetadata
+from dbgpt.core import ModelMetadata, ModelOutput
 from dbgpt.core.awel.flow import (
     TAGS_ORDER_HIGH,
     ResourceCategory,
@@ -57,6 +67,16 @@ class DeepSeekDeployModelParameters(OpenAICompatibleDeployModelParameters):
         },
     )
 
+    thinking_enabled: Optional[bool] = field(
+        default=None,
+        metadata={
+            "help": _(
+                "Whether to enable DeepSeek thinking mode. If None, thinking is "
+                "disabled for deepseek-v4-pro to keep ReAct output parseable."
+            ),
+        },
+    )
+
 
 async def deepseek_generate_stream(
     model: ProxyModel, tokenizer, params, device, context_len=2048
@@ -88,6 +108,7 @@ class DeepseekLLMClient(OpenAILLMClient):
         context_length: Optional[int] = None,
         openai_client: Optional["ClientType"] = None,
         openai_kwargs: Optional[Dict[str, Any]] = None,
+        thinking_enabled: Optional[bool] = None,
         **kwargs,
     ):
         api_base = (
@@ -96,7 +117,9 @@ class DeepseekLLMClient(OpenAILLMClient):
         api_key = api_key or os.getenv("DEEPSEEK_API_KEY")
         model = model or _DEFAULT_MODEL
         if not context_length:
-            if "deepseek-chat" in model:
+            if "deepseek-v4" in model:
+                context_length = 1024 * 1024
+            elif "deepseek-chat" in model:
                 context_length = 1024 * 32
             elif "deepseek-coder" in model:
                 context_length = 1024 * 16
@@ -109,6 +132,7 @@ class DeepseekLLMClient(OpenAILLMClient):
                 "Deepseek API key is required, please set 'DEEPSEEK_API_KEY' in "
                 "environment variable or pass it to the client."
             )
+        self._thinking_enabled = thinking_enabled
         super().__init__(
             api_key=api_key,
             api_base=api_base,
@@ -124,6 +148,69 @@ class DeepseekLLMClient(OpenAILLMClient):
             **kwargs,
         )
 
+    @classmethod
+    def new_client(
+        cls,
+        model_params: DeepSeekDeployModelParameters,
+        default_executor=None,
+    ) -> "DeepseekLLMClient":
+        """Create a new client with the model parameters."""
+        return cls(
+            api_key=model_params.api_key,
+            api_base=model_params.api_base,
+            api_type=model_params.api_type,
+            api_version=model_params.api_version,
+            model=model_params.real_provider_model_name,
+            proxy=model_params.http_proxy,
+            model_alias=model_params.real_provider_model_name,
+            context_length=model_params.context_length,
+            thinking_enabled=model_params.thinking_enabled,
+        )
+
+    def _build_request(self, request, stream: Optional[bool] = False) -> Dict[str, Any]:
+        payload = super()._build_request(request, stream)
+        model = payload.get("model") or self.default_model
+        extra_body = payload.get("extra_body") or {}
+        if "thinking" not in extra_body and self._should_set_thinking(model):
+            thinking_enabled = self._thinking_enabled is True
+            extra_body["thinking"] = {
+                "type": "enabled" if thinking_enabled else "disabled"
+            }
+            payload["extra_body"] = extra_body
+        return payload
+
+    def _should_set_thinking(self, model: str) -> bool:
+        return self._thinking_enabled is not None or model == "deepseek-v4-pro"
+
+    def _is_thinking_disabled(self, payload: Dict[str, Any]) -> bool:
+        thinking = (payload.get("extra_body") or {}).get("thinking") or {}
+        return thinking.get("type") == "disabled"
+
+    def _drop_thinking_if_disabled(
+        self, output: ModelOutput, payload: Dict[str, Any]
+    ) -> ModelOutput:
+        if not self._is_thinking_disabled(payload) or not output.has_thinking:
+            return output
+        text = output.text if output.has_text else ""
+        return ModelOutput.build(
+            text=text,
+            usage=output.usage,
+            finish_reason=output.finish_reason,
+            metrics=output.metrics,
+        )
+
+    async def generate_v1(
+        self, messages: List[Dict[str, Any]], payload: Dict[str, Any]
+    ) -> ModelOutput:
+        output = await super().generate_v1(messages, payload)
+        return self._drop_thinking_if_disabled(output, payload)
+
+    async def generate_stream_v1(
+        self, messages: List[Dict[str, Any]], payload: Dict[str, Any]
+    ) -> AsyncIterator[ModelOutput]:
+        async for output in super().generate_stream_v1(messages, payload):
+            yield self._drop_thinking_if_disabled(output, payload)
+
     def check_sdk_version(self, version: str) -> None:
         if not version >= "1.0":
             raise ValueError(
@@ -154,6 +241,14 @@ class DeepseekLLMClient(OpenAILLMClient):
 register_proxy_model_adapter(
     DeepseekLLMClient,
     supported_models=[
+        ModelMetadata(
+            model="deepseek-v4-pro",
+            context_length=1024 * 1024,
+            max_output_length=384 * 1024,
+            description="DeepSeek-V4-Pro by DeepSeek",
+            link="https://api-docs.deepseek.com/news/news260424",
+            function_calling=True,
+        ),
         ModelMetadata(
             model="deepseek-chat",
             context_length=64 * 1024,

packages/dbgpt-core/src/dbgpt/model/proxy/tests/test_deepseek.py

@@ -0,0 +1,94 @@
+"""Tests for DeepSeek proxy LLM client."""
+
+import os
+from unittest.mock import patch
+
+from dbgpt.core import ModelMessage, ModelOutput, ModelRequest
+from dbgpt.model.proxy.llms.deepseek import DeepseekLLMClient
+
+
+class TestDeepSeekModelRegistration:
+    """DeepSeek provider/model adapter resolution."""
+
+    def test_supported_models_contains_v4_pro_metadata(self):
+        from dbgpt.model.adapter.base import get_model_adapter
+
+        adapter = get_model_adapter("proxy/deepseek", "deepseek-v4-pro")
+        assert adapter is not None
+
+        metadata = {metadata.model: metadata for metadata in adapter.supported_models()}
+        v4_pro = metadata["deepseek-v4-pro"]
+
+        assert v4_pro.context_length == 1024 * 1024
+        assert v4_pro.max_output_length == 384 * 1024
+        assert v4_pro.function_calling is True
+
+
+class TestDeepSeekContextLength:
+    """DeepSeek model context length defaults."""
+
+    @patch.dict(os.environ, {"DEEPSEEK_API_KEY": "test-key"})
+    def test_v4_pro_uses_one_million_context(self):
+        client = DeepseekLLMClient(
+            model="deepseek-v4-pro",
+            openai_client=_FakeOpenAIClient(),
+        )
+
+        assert client.context_length == 1024 * 1024
+        assert client._context_length == 1024 * 1024
+
+
+class TestDeepSeekThinkingMode:
+    """DeepSeek V4 thinking mode defaults for parse-sensitive agents."""
+
+    def _request(self):
+        return ModelRequest(
+            model="deepseek-v4-pro",
+            messages=[ModelMessage(role="user", content="hi")],
+        )
+
+    @patch.dict(os.environ, {"DEEPSEEK_API_KEY": "test-key"})
+    def test_v4_pro_disables_thinking_by_default(self):
+        client = DeepseekLLMClient(
+            model="deepseek-v4-pro",
+            openai_client=_FakeOpenAIClient(),
+        )
+
+        payload = client._build_request(self._request())
+
+        assert payload["extra_body"]["thinking"] == {"type": "disabled"}
+
+    @patch.dict(os.environ, {"DEEPSEEK_API_KEY": "test-key"})
+    def test_v4_pro_can_enable_thinking_explicitly(self):
+        client = DeepseekLLMClient(
+            model="deepseek-v4-pro",
+            thinking_enabled=True,
+            openai_client=_FakeOpenAIClient(),
+        )
+
+        payload = client._build_request(self._request())
+
+        assert payload["extra_body"]["thinking"] == {"type": "enabled"}
+
+    @patch.dict(os.environ, {"DEEPSEEK_API_KEY": "test-key"})
+    def test_disabled_thinking_drops_reasoning_content(self):
+        client = DeepseekLLMClient(
+            model="deepseek-v4-pro",
+            openai_client=_FakeOpenAIClient(),
+        )
+        output = ModelOutput.build(
+            text="Thought: continue\nAction: terminate\nAction Input: {}",
+            thinking="private reasoning",
+        )
+        payload = client._build_request(self._request())
+
+        sanitized = client._drop_thinking_if_disabled(output, payload)
+
+        assert sanitized.has_thinking is False
+        assert (
+            sanitized.text == "Thought: continue\nAction: terminate\nAction Input: {}"
+        )
+
+
+class _FakeOpenAIClient:
+    default_headers = {}