diff --git a/api/core/workflow/nodes/llm/node.py b/api/core/workflow/nodes/llm/node.py
index c413ad758b..28c4925456 100644
--- a/api/core/workflow/nodes/llm/node.py
+++ b/api/core/workflow/nodes/llm/node.py
@@ -573,7 +573,11 @@ class LLMNode(Node[LLMNodeData]):
                 "tool_calls": [self._serialize_tool_call(item) for item in generation_data.tool_calls],
                 "sequence": generation_data.sequence,
             }
-            files_to_output = generation_data.files
+            files_to_output = list(generation_data.files)
+            # Merge auto-collected/structured-output files from self._file_outputs
+            if self._file_outputs:
+                existing_ids = {f.id for f in files_to_output}
+                files_to_output.extend(f for f in self._file_outputs if f.id not in existing_ids)
         else:
             # Classical runtime: use pre-computed generation-specific text pair,
             # falling back to native model reasoning if no <think> tags were found.
@@ -2126,6 +2130,12 @@ class LLMNode(Node[LLMNodeData]):
 
             result = yield from self._process_tool_outputs(outputs)
 
+            # Auto-collect sandbox output/ files, deduplicate by id
+            collected_files = session.collect_output_files()
+            if collected_files:
+                existing_ids = {f.id for f in self._file_outputs}
+                self._file_outputs.extend(f for f in collected_files if f.id not in existing_ids)
+
         if result is None:
             raise LLMNodeError("SandboxSession exited unexpectedly")
 
diff --git a/web/app/components/workflow/constants.ts b/web/app/components/workflow/constants.ts
index d49eacc7ac..73e80dd0a2 100644
--- a/web/app/components/workflow/constants.ts
+++ b/web/app/components/workflow/constants.ts
@@ -182,6 +182,10 @@ export const LLM_OUTPUT_STRUCT: Var[] = [
     variable: 'usage',
     type: VarType.object,
   },
+  {
+    variable: 'files',
+    type: VarType.arrayFile,
+  },
 ]
 
 export const KNOWLEDGE_RETRIEVAL_OUTPUT_STRUCT: Var[] = [
diff --git a/web/app/components/workflow/nodes/llm/panel.tsx b/web/app/components/workflow/nodes/llm/panel.tsx
index 4d8b24676a..06043da9a9 100644
--- a/web/app/components/workflow/nodes/llm/panel.tsx
+++ b/web/app/components/workflow/nodes/llm/panel.tsx
@@ -461,6 +461,11 @@ const Panel: FC<NodePanelProps<LLMNodeType>> = ({
             type="object"
             description={t(`${i18nPrefix}.outputVars.usage`, { ns: 'workflow' })}
           />
+          <VarItem
+            name="files"
+            type="array[file]"
+            description={t(`${i18nPrefix}.outputVars.files`, { ns: 'workflow' })}
+          />
           {inputs.structured_output_enabled && (
             <>
               <Split className="mt-3" />
diff --git a/web/i18n/en-US/workflow.json b/web/i18n/en-US/workflow.json
index 78ee062559..bc4d22155a 100644
--- a/web/i18n/en-US/workflow.json
+++ b/web/i18n/en-US/workflow.json
@@ -828,6 +828,7 @@
   "nodes.llm.jsonSchema.warningTips.saveSchema": "Please finish editing the current field before saving the schema",
   "nodes.llm.model": "model",
   "nodes.llm.notSetContextInPromptTip": "To enable the context feature, please fill in the context variable in PROMPT.",
+  "nodes.llm.outputVars.files": "Files produced during execution",
   "nodes.llm.outputVars.generation": "Generation Information",
   "nodes.llm.outputVars.output": "Generate content",
   "nodes.llm.outputVars.reasoning_content": "Reasoning Content",
diff --git a/web/i18n/zh-Hans/workflow.json b/web/i18n/zh-Hans/workflow.json
index 29ad24c01e..373c6553dd 100644
--- a/web/i18n/zh-Hans/workflow.json
+++ b/web/i18n/zh-Hans/workflow.json
@@ -828,6 +828,7 @@
   "nodes.llm.jsonSchema.warningTips.saveSchema": "请先完成当前字段的编辑",
   "nodes.llm.model": "模型",
   "nodes.llm.notSetContextInPromptTip": "要启用上下文功能，请在提示中填写上下文变量。",
+  "nodes.llm.outputVars.files": "执行过程中产出的文件",
   "nodes.llm.outputVars.generation": "生成信息",
   "nodes.llm.outputVars.output": "生成内容",
   "nodes.llm.outputVars.reasoning_content": "推理内容",