Add workflow support and enhance runtime configuration in models and services

- Introduce RuntimeModelResource and RuntimeKnowledgeBase classes to manage workflow resources.
- Update AssistantConfig to include workflow_model_resources and workflow_knowledge_bases for better integration.
- Refactor validation and processing logic in routes and services to accommodate workflow types.
- Implement dynamic variable support for workflow assistants and enhance graph normalization.
- Add ToolExecutor for reusable tool execution across different assistant types.
- Update various services to ensure compatibility with new workflow features and improve error handling.

backend/services/runtime_variables.py

@@ -175,9 +175,9 @@ def prepare_dynamic_config(
     trusted_values: dict[str, Any] | None = None,
 ) -> AssistantConfig:
     """Validate and merge one call's values without mutating stored config."""
-    if cfg.type != "prompt":
+    if cfg.type not in {"prompt", "workflow"}:
         if client_values:
-            raise DynamicVariableError("动态变量目前仅支持 prompt 助手")
+            raise DynamicVariableError("动态变量仅支持 prompt 和 workflow 助手")
         return cfg
 
     supplied = _validate_public(client_values)
@@ -214,9 +214,13 @@ def prepare_dynamic_config(
     prepared = cfg.model_copy(deep=True)
     prepared.dynamic_variables = merged
     prepared.conversation_id = conversation_id
-    # Validate prompt and greeting before media resources are allocated.
+    # Validate top-level prompt/greeting before media resources are allocated.
+    # Workflow node templates are rendered lazily as their nodes become active.
     store = DynamicVariableStore.from_config(prepared)
-    store.render(prepared.prompt)
+    if prepared.type == "prompt":
+        store.render(prepared.prompt)
+    elif prepared.type == "workflow":
+        store.render_data(prepared.graph)
     store.render(prepared.greeting)
     return prepared