所有节点得先"报到"才能被造出来。register_nodes（api/core/workflow/node_factory.py:113）负责这次报到：

# api/core/workflow/node_factory.py:113
def register_nodes() -> None:
    """Import production node modules so they self-register with ``Node``."""
    _import_node_package("graphon.nodes")          # ① 导入引擎内置的通用节点
    _import_node_package("core.workflow.nodes")    # ② 导入 Dify 本地的专有节点

def get_node_type_classes_mapping() -> Mapping[NodeType, Mapping[str, type[Node]]]:
    register_nodes()                                # ③ 先确保都注册了
    return Node.get_node_type_classes_mapping()     # ④ 返回"类型→版本→类"的登记表快照

登记表是 {类型: {版本: 类}} 两层结构。resolve_workflow_node_class（api/core/workflow/node_factory.py:131）负责查表：

# api/core/workflow/node_factory.py:131
def resolve_workflow_node_class(*, node_type, node_version) -> type[Node]:
    node_mapping = get_node_type_classes_mapping().get(node_type)   # ① 先按类型取"版本表"
    if not node_mapping:
        raise ValueError(f"No class mapping found for node type: {node_type}")

    latest_node_class = node_mapping.get(LATEST_VERSION)            # ② 取"latest"兜底
    matched_node_class = node_mapping.get(node_version)             # ③ 取精确匹配的版本
    node_class = matched_node_class or latest_node_class           # ④ ★精确优先，否则退 latest
    if not node_class:
        raise ValueError(f"No latest version class found for node type: {node_type}")
    return node_class

Day08 说拓扑不选起点。选起点是 get_default_root_node_id（api/core/workflow/node_factory.py:150）的活：

# api/core/workflow/node_factory.py:150
def get_default_root_node_id(graph_config) -> str:
    nodes = graph_config.get("nodes")
    for node in nodes:
        if node.get("type") == "custom-note":       # ① 跳过"便签"这种非执行节点
            continue
        node_id = node.get("id"); data = node.get("data")
        node_type = data.get("type")
        if is_start_node_type(node_type):            # ② ★找到第一个"开始类"节点
            return node_id
    raise ValueError("Unable to determine default root node ID from workflow graph")

# api/core/workflow/node_factory.py:145
def is_start_node_type(node_type) -> bool:
    return node_type in _START_NODE_TYPES            # 开始/数据源/各种触发器

核心方法 DifyNodeFactory.create_node（api/core/workflow/node_factory.py:377）。它先解析类，再按节点类型选一份"依赖配方"：

# api/core/workflow/node_factory.py:377
def create_node(self, node_config) -> Node:
    adapted_node_config = adapt_node_config_for_graph(node_config)
    typed_node_config = NodeConfigDictAdapter.validate_python(adapted_node_config)  # ① 校验配置
    node_id = typed_node_config["id"]; node_data = typed_node_config["data"]
    node_class = self._resolve_node_class(node_type=node_data.type, node_version=str(node_data.version))
    resolved_node_data = self._validate_resolved_node_data(node_class, node_data)  # ② 用具体类再校验一遍
    node_type = node_data.type

    # ③ ★按节点类型选"依赖配方"——每种节点要的料不同
    node_init_kwargs_factories = {
        BuiltinNodeTypes.CODE:  lambda: {"code_executor": self._code_executor, "code_limits": ...},
        BuiltinNodeTypes.HTTP_REQUEST: lambda: {"http_client": ..., "file_manager": ...},
        BuiltinNodeTypes.LLM:   lambda: self._build_llm_compatible_node_init_kwargs(node_class=node_class, ...),
        BuiltinNodeTypes.TOOL:  lambda: {"tool_file_manager": ..., "runtime": self._tool_runtime},
        BuiltinNodeTypes.AGENT: lambda: self._build_agent_node_init_kwargs(node_class=node_class),
        # ... QUESTION_CLASSIFIER / PARAMETER_EXTRACTOR / DOCUMENT_EXTRACTOR / HUMAN_INPUT ...
    }
    node_init_kwargs = node_init_kwargs_factories.get(node_type, lambda: {})()     # ④ 取配方；没有就空
    constructor_node_data = resolved_node_data.model_dump(mode="python", by_alias=True)
    return node_class(                                                             # ⑤ ★造节点，注入依赖
        node_id=node_id, data=constructor_node_data,
        graph_init_params=self.graph_init_params,
        graph_runtime_state=self.graph_runtime_state,
        **node_init_kwargs,
    )

注意配方表里，LLM、QUESTION_CLASSIFIER（问题分类）、PARAMETER_EXTRACTOR（参数提取）三种节点都调同一个 _build_llm_compatible_node_init_kwargs（api/core/workflow/node_factory.py:534），只是开关不同：

# api/core/workflow/node_factory.py:418（LLM 节点的配方）
BuiltinNodeTypes.LLM: lambda: self._build_llm_compatible_node_init_kwargs(
    node_class=node_class, node_data=resolved_node_data,
    wrap_model_instance=True,              # ① 要把 ModelInstance 包成节点能用的适配器(Day10!)
    include_http_client=True,
    include_llm_file_saver=True,
    include_prompt_message_serializer=True, # ② 要 prompt 消息序列化器
    include_retriever_attachment_loader=True,
    include_jinja2_template_renderer=True,
),
# api/core/workflow/node_factory.py:442（参数提取节点：同一套，但关掉大部分开关）
BuiltinNodeTypes.PARAMETER_EXTRACTOR: lambda: self._build_llm_compatible_node_init_kwargs(
    node_class=node_class, node_data=resolved_node_data,
    wrap_model_instance=True,
    include_http_client=False, include_llm_file_saver=False,
    include_prompt_message_serializer=True,
    include_retriever_attachment_loader=False, include_jinja2_template_renderer=False,
),