diff --git a/app/actions/batch.py b/app/actions/batch.py
index df66bbc..1d3781f 100644
--- a/app/actions/batch.py
+++ b/app/actions/batch.py
@@ -7,7 +7,7 @@ from copy import deepcopy
 
 from app.actions.registry import apply_action
 from app.actions.update import fit_existing_path
-from app.resources.material_fields import check_update_field
+from app.resources.material_fields import check_update_field, find_material
 from app.schemas.actions import Action, AddAction, UpdateAction
 from app.schemas.capabilities import DATA_BINDING_ERROR, DATA_BINDING_SUPPORTED
 from app.schemas.component_resource import ComponentCatalog
@@ -39,18 +39,8 @@ def merge_config(defaults: dict, overrides: dict) -> dict:
 
 
 def complete_add(action: AddAction, catalog: ComponentCatalog) -> AddAction:
-    """用物料默认配置补全新增组件，模型显式写出的字段优先，并按物料设置 sample_data"""
-    material = next(
-        (item for item in catalog.components if item.key == action.component_key), None
-    )
-    if material is None:
-        raise ValueError(
-            f"物料 {action.component_key} 不存在，可用物料用 list_components 查看"
-        )
-    if material.type != action.node.type:
-        raise ValueError(
-            f"物料 {material.key} 的类型是 {material.type}，node.type 必须与之一致"
-        )
+    """按 node.type 找物料补全新增组件，模型显式写出的字段优先，并按物料设置 sample_data"""
+    material = find_material(catalog, action.node.type)
     if not DATA_BINDING_SUPPORTED and action.node.model_extra.get("dataId"):
         raise ValueError(DATA_BINDING_ERROR)
     overrides = action.node.model_dump(mode="json", exclude_unset=True)
diff --git a/app/clients/mcp.py b/app/clients/mcp.py
index 6ddfb33..eb70446 100644
--- a/app/clients/mcp.py
+++ b/app/clients/mcp.py
@@ -8,6 +8,7 @@
 import asyncio
 import json
 import os
+import re
 import socket
 import subprocess
 import sys
@@ -29,18 +30,37 @@ CONFIG_PATH = PROJECT_ROOT / "mcp_servers.json"
 # 物料服务返回的目录版本须与主进程一致，否则模型查到的字段和校验依据不同
 MATERIAL_SERVER = "materials"
 LOCAL_HOSTS = {"localhost", "127.0.0.1"}
+# 配置中引用环境变量的写法
+ENV_REFERENCE = re.compile(r"\$\{(\w+)\}")
 
 
 class MCPServiceError(RuntimeError):
     """MCP 配置错误、服务不可用，或物料目录版本不一致"""
 
 
+def _expand_env(value):
+    """只展开字符串值里的 ${NAME}；变量未设置时报错，不把字面量当成地址或密钥"""
+    if isinstance(value, dict):
+        return {key: _expand_env(item) for key, item in value.items()}
+    if isinstance(value, list):
+        return [_expand_env(item) for item in value]
+    if not isinstance(value, str):
+        return value
+
+    def lookup(match: re.Match) -> str:
+        if (found := os.environ.get(match[1])) is None:
+            raise MCPServiceError(f"MCP 服务配置引用的环境变量 {match[1]} 未设置")
+        return found
+
+    return ENV_REFERENCE.sub(lookup, value)
+
+
 def load_mcp_config(path: Path = CONFIG_PATH) -> dict[str, dict]:
     """读取服务列表并展开环境变量；文件缺失或格式不对时报错"""
     if not path.exists():
         raise MCPServiceError(f"未找到 MCP 服务配置: {path}")
     try:
-        config = json.loads(os.path.expandvars(path.read_text(encoding="utf-8")))
+        config = _expand_env(json.loads(path.read_text(encoding="utf-8")))
     except json.JSONDecodeError as error:
         raise MCPServiceError(f"MCP 服务配置不是合法 JSON: {error}") from error
     if not isinstance(config, dict) or not all(
diff --git a/app/main.py b/app/main.py
index 7338cd3..99e5124 100644
--- a/app/main.py
+++ b/app/main.py
@@ -17,6 +17,7 @@ from app.core.logging import setup_logging
 @asynccontextmanager
 async def lifespan(application: FastAPI):
     setup_logging()
+    from app.agent.screen_agent import get_screen_agent
     from app.clients.decision import close_decision_engine
     from app.clients.mcp import start_mcp_servers, stop_mcp_servers
     from app.clients.user_center import close_user_center
@@ -35,6 +36,11 @@ async def lifespan(application: FastAPI):
         await warm_material_cache(persistence.db)
         # 拉起 mcp_servers.json 中由应用管理的服务；单个服务失败只记录，用到它的请求执行时报错
         mcp_processes = await start_mcp_servers()
+        # 提前拉取 MCP 工具并组装 agent，问题在启动时暴露；失败只记录，首个请求会重试
+        try:
+            await get_screen_agent()
+        except Exception as error:
+            logger.opt(exception=error).error("screen agent 预先组装失败")
         application.state.agent_service = AgentService(
             build_graph(persistence.checkpointer), persistence.store
         )
diff --git a/app/observability/tracing.py b/app/observability/tracing.py
index 8eba04b..113e34b 100644
--- a/app/observability/tracing.py
+++ b/app/observability/tracing.py
@@ -55,26 +55,6 @@ def flush_tracing() -> None:
         logger.opt(exception=exc).warning("Langfuse flush 失败")
 
 
-async def ainvoke_screen(
-    graph: Any,
-    payload: Any,
-    *,
-    thread_id: str,
-    session_id: str | None = None,
-    user_id: int | None = None,
-) -> Any:
-    """每轮调用记录一个根 span，返回、等待用户或异常时结束。
-
-    session_id 由调用方在首次与恢复时一致传入（通常为 conversation_id）。
-    未传时始终使用 thread_id，不从仅首次请求才有的 payload 推断。
-    user_id 为用户中心的用户 ID，接口调用时传入，本地脚本不传。
-    """
-    with _screen_trace(payload, thread_id, session_id, user_id) as (config, root):
-        result = await graph.ainvoke(payload, config=config)
-        _finish_trace(root, result)
-        return result
-
-
 async def astream_screen(graph, payload, *, thread_id, session_id=None, user_id=None):
     """同一轮流保持同一个追踪上下文，关闭流时结束追踪"""
     from contextlib import aclosing
diff --git a/app/processor/graph.py b/app/processor/graph.py
index e5b77ec..18372d9 100644
--- a/app/processor/graph.py
+++ b/app/processor/graph.py
@@ -81,7 +81,3 @@ def build_graph(checkpointer=None):
     if checkpointer is None:
         checkpointer = InMemorySaver(serde=build_checkpoint_serde())
     return workflow.compile(checkpointer=checkpointer)
-
-
-# 模块加载时创建共享图实例，复用该实例的调用也共用其检查点存储
-query_app = build_graph()
diff --git a/app/processor/nodes/screen_agent_node.py b/app/processor/nodes/screen_agent_node.py
index 01c971f..d150479 100644
--- a/app/processor/nodes/screen_agent_node.py
+++ b/app/processor/nodes/screen_agent_node.py
@@ -62,8 +62,6 @@ class ScreenAgentNode(BaseNode):
                 ]
             },
             context=ScreenAgentContext(screen=context.screen, catalog=catalog),
-            # 调用次数由 ModelCallLimitMiddleware 限制，这里只防止异常循环
-            config={"recursion_limit": 4 * self.config.max_agent_steps + 10},
         )
         if result.get("structured_response") is None:
             last = result["messages"][-1]
diff --git a/app/resources/material_fields.py b/app/resources/material_fields.py
index 85af7ab..3056d88 100644
--- a/app/resources/material_fields.py
+++ b/app/resources/material_fields.py
@@ -105,7 +105,7 @@ def list_components(catalog: ComponentCatalog) -> dict:
 
 
 def get_materials(catalog: ComponentCatalog, component_types: list[str]) -> dict:
-    """指定类型的物料默认配置，同一类型的多个预设全部返回"""
+    """指定类型的物料默认配置"""
     known = {item.type for item in catalog.components}
     if unknown := [item for item in component_types if item not in known]:
         raise ValueError(
diff --git a/app/schemas/actions/add.py b/app/schemas/actions/add.py
index 6aef149..a4b06b5 100644
--- a/app/schemas/actions/add.py
+++ b/app/schemas/actions/add.py
@@ -5,7 +5,6 @@ from typing import Literal
 from pydantic import Field
 
 from app.schemas.actions.base import ActionModel
-from app.schemas.base import NonBlankString
 from app.schemas.screen import ScreenNode
 
 
@@ -13,13 +12,10 @@ class AddAction(ActionModel):
     """按物料新增组件"""
 
     type: Literal["add"] = "add"
-    component_key: NonBlankString = Field(
-        description="物料目录编号，即 list_components 返回的 key"
-    )
     node: ScreenNode = Field(
         description="新组件。id、type、layout 必填，其余只写需要覆盖物料默认配置的字段；"
-        "后端按物料默认配置补全：对象逐层覆盖，数组整体替换，null 明确覆盖。"
-        "id 不能与现有组件重复，同一批里先 delete 的 ID 可以再用；type 必须与物料一致"
+        "后端按 type 找到物料，用默认配置补全：对象逐层覆盖，数组整体替换，null 明确覆盖。"
+        "id 不能与现有组件重复，同一批里先 delete 的 ID 可以再用；type 用 list_components 返回的 type"
     )
     sample_data: bool = Field(
         default=False,
diff --git a/app/utils/__init__.py b/app/utils/__init__.py
deleted file mode 100644
index e69de29..0000000
diff --git a/main.py b/main.py
deleted file mode 100644
index eb389a0..0000000
--- a/main.py
+++ /dev/null
@@ -1,16 +0,0 @@
-# 这是一个示例 Python 脚本。
-
-# 按 Shift+F10 执行或将其替换为您的代码。
-# 按 双击 Shift 在所有地方搜索类、文件、工具窗口、操作和设置。
-
-
-def print_hi(name):
-    # 在下面的代码行中使用断点来调试脚本。
-    print(f'Hi, {name}')  # 按 Ctrl+F8 切换断点。
-
-
-# 按装订区域中的绿色按钮以运行脚本。
-if __name__ == '__main__':
-    print_hi('PyCharm')
-
-# 访问 https://www.jetbrains.com/help/pycharm/ 获取 PyCharm 帮助
diff --git a/mcp_servers/material_server.py b/mcp_servers/material_server.py
index ecceee7..04c7189 100644
--- a/mcp_servers/material_server.py
+++ b/mcp_servers/material_server.py
@@ -46,10 +46,7 @@ def _require_catalog() -> ComponentCatalog:
 
 @mcp.tool()
 async def list_components() -> dict[str, Any]:
-    """列出可以新增的组件：key、type、名称、用途和默认宽高
-
-    新建大屏、新增组件或更换类型时，先从这里选类型，再用 get_component_materials 查看默认配置
-    """
+    """列出可以新增的组件：key、type、名称、用途和默认宽高"""
     return material_fields.list_components(_require_catalog())
 
 
@@ -60,10 +57,7 @@ async def get_component_materials(
         Field(description="要使用的物料类型，可一次传多个", min_length=1),
     ],
 ) -> dict[str, Any]:
-    """获取物料的完整默认配置（schema）和使用说明
-
-    新增组件前先查看所选类型的默认配置，只覆盖需要修改的字段，不从零编写 props
-    """
+    """获取物料的完整默认配置（schema）和使用说明"""
     return material_fields.get_materials(_require_catalog(), component_types)
 
 
diff --git a/skills/chart-replace/SKILL.md b/skills/chart-replace/SKILL.md
index 135e449..b825728 100644
--- a/skills/chart-replace/SKILL.md
+++ b/skills/chart-replace/SKILL.md
@@ -11,7 +11,7 @@ description: 把已有组件换成另一种组件类型，例如"把折线图换
 2. 用 list_components 选出用户要的新类型，用 get_component_materials 查看它的默认配置
 3. 在同一批动作里先 delete 原组件，再 add 新组件：
     - node.id 沿用原组件 ID，layout 沿用原组件的位置和尺寸（用户另有要求除外）
-    - component_key 选新类型对应的物料
+    - node.type 填新类型
 4. 用 apply_actions 提交
 
 ## 数据迁移
diff --git a/skills/screen-design/SKILL.md b/skills/screen-design/SKILL.md
index 96d0de0..c78e55d 100644
--- a/skills/screen-design/SKILL.md
+++ b/skills/screen-design/SKILL.md
@@ -8,7 +8,7 @@ description: 新建大屏，或在大屏上新增组件、整体重新布局、
 ## 选组件
 
 1. 用 list_components 查看可用组件，按用户要展示的内容选类型；不编造目录中没有的组件类型
-2. 用 get_component_materials 查看所选类型的默认配置（可一次查多个类型），新组件以默认配置为基础，只写需要改的字段，例如标题
+2. 用 get_component_materials 查看所选类型的默认配置
 3. 新建大屏时，根据主题安排一个标题文本和若干图表；图表标题按要展示的内容写，不沿用物料的示例标题
 
 ## 布局
@@ -22,7 +22,3 @@ description: 新建大屏，或在大屏上新增组件、整体重新布局、
 
 - 重新布局：用 move 和 resize 调整已有组件，不删除、不重建
 - 统一风格：先用 get_nodes 读取组件配置，用 search_fields 找到颜色等字段，再逐个修改；画布背景用 update_canvas
-
-## 提交
-
-本次的所有新增和调整放在同一批，用 apply_actions 提交
diff --git a/skills/screen-edit/SKILL.md b/skills/screen-edit/SKILL.md
index f6e873d..10b28a6 100644
--- a/skills/screen-edit/SKILL.md
+++ b/skills/screen-edit/SKILL.md
@@ -23,7 +23,3 @@ description: 修改大屏上已有的组件或画布：改标题、文字、颜
 
 - 相对变化（"放大一倍""往右移 100"）依据大屏摘要中的当前位置和尺寸换算
 - "放到正中间"等按画布尺寸计算目标位置
-
-## 提交
-
-本次的所有修改放在同一批，用 apply_actions 提交