mirror of
https://github.com/labring/FastGPT.git
synced 2026-05-08 01:08:43 +08:00
7506a147e6
* batch node (#6732) * batch node * docs: add local code quality standards and style guides for automated review * refactor: remove enforced minimum for parallel concurrency, simplify edge handling in task runtime context, and fix loop output mapping * feat: auto-infer and sync valueType for parallel loop input and output based on referenced array source * fix: refactor parallelRun output type synchronization and improve sub-workflow error handling in dispatch service * feat: enforce parallel concurrency limits and validate against workflow loop constraints * feat: implement retry mechanism for parallel workflow tasks with usage tracking per attempt * fix review * perf: use function * refactor: abstract nested node logic into useNestedNode hook and update parallelRun icon/service logic * fix: type import * refactor: update ParallelRunStatusEnum and i18n labels for improved status clarity * feat: parallel run details and input/output display to chat response modal and service dispatch * fix: config limit error * refactor: optimize parallel run task execution, fix point accumulation, and improve error handling for sub-workflows * fix: include totalPoints in parallel task results * refactor: centralize nested input injection and point safety utilities for workflow dispatchers * test: add unit tests for safePoints utility function * refactor: update parallel workflow runtime types and clean up docstring placement in dispatch utils * fix: include all runtime nodes in parallel execution to ensure variable reference accessibility * refactor: update pushSubWorkflowUsage signature to use object parameter for improved consistency --------- Co-authored-by: DigHuang <114602213+DigHuang@users.noreply.github.com> * feat(s3): add proxy transfer mode with tokenized upload/download (#6729) * feat(s3): add proxy transfer mode with tokenized upload/download * wip: switch to proxy mode for upload progress * fix: office mime types * fix(s3): upload MIME validation, multer whitelist, API error status - Treat AVI/MPEG mime aliases (incl. video/mp1s vs video/mpeg) as matching - Optional allowedExtensions on multer for dataset images and localFile - Map S3/business errors to 4xx in jsonRes where appropriate - Align presign max size with team plan; fix dataset import size UX - Add upload validation tests Made-with: Cursor * fix: show clear message when upload frequency limit is exceeded - Reject ERROR_ENUM.uploadFileIntervalLimit from authFrequencyLimit instead of Mongo doc - Add i18n for upload_file_interval_limit (zh-CN/en/zh-Hant) Made-with: Cursor * fix file token validation and upload mime checks * fix: test * fix(s3): treat m4a audio/mp4 and audio/x-m4a as equivalent - Add MIME equivalence group for AAC/M4A container mismatch (mime-types vs file-type) - Add upload validation test for minimal ftyp/M4A buffer - Test env: keep FILE_TOKEN_KEY in vitest test.env and test/setup.ts (drop loadTestEnv file) Made-with: Cursor * fix(chat): 调试区文件类型与编辑态一致,并修复 accept 在 WebKit 下不更新 - ChatTest: 用 getAppChatConfig + getGuideModule 合并画布引导节点与 chatConfig - useChatTest: 依赖 fileSelectConfig 序列化与 chatConfig,避免深层变更未触发预览更新 - useSelectFile: 用 useCallback + input key 替代 useMemoizedFn,确保 accept 变更后重建 input Made-with: Cursor * fix: invalid request * feat: prompt inject (#6757) * feat: resume chat stream (#6722) * fix: openapi schema issue while creating openapi json * feat: resume chat stream * wip: chat status and read status * feat: sync chat side bar status * fix: allow reassignment of variables in chatTest handler Made-with: Cursor * feat(chat): stream resume hardening, resume modules in @fastgpt/service, stale generating cron - Move stream resume mirror + resumeStatus into packages/service; update API imports - chatTest: ensurePendingChatRoundItems, default responseChatItemId; zod default import for client - useChatTest + HomeChatWindow: enableAutoResume and sync init chatGenerateStatus - ChatContext: safe no-op defaults without provider - Cron: clean MongoChat stuck in generating >30min; timer lock cleanStaleGeneratingChat Made-with: Cursor * fix(chat): address stream-resume PR review (zod/mongoose enum, legacy status, upsert, UI race) - Zod: use z.nativeEnum(ChatGenerateStatusEnum); mongoose chatGenerateStatus enum as [0,1,2] only - Init APIs: default missing chatGenerateStatus to done before read/unread logic - ensurePendingChatRoundItems: unique index + upsert; rename ChatGenerateStatusEnum - ChatBox auto-resume: guard by chatId; sidebar sync via targetChatId - Tests: chat history/feedback APIs pass with schema fixes Made-with: Cursor * fix(chat): expose resume at /api/v2/chat/resume; openapi + review tidy - Move handler from v1/stream to v2/chat/resume (pairs with v2 completions + Redis mirror) - Update fetch, OpenAPI AIPath, comments; remove slim projects/app global chat api - getHistoryStatus default chatGenerateStatus; team init + chatTest notes; ChatItem tweak Made-with: Cursor * fix(chat): fix resume JSON parse catch shadowing; drop unused resumeChatStream Made-with: Cursor * docs(chat): comment closed+stream mirror write path in workflow dispatch Made-with: Cursor * refactor: unify resumable stream mirroring * fix: keep v1 chat completions out of resume flow * refactor: make prepared chat rounds transactional * fix: handle resume stream terminal errors * fix: rerank max token * feat(workflow): extend variable update node with Number/Boolean/Array operations (#6752) * feat(workflow): extend variable update node with Number/Boolean/Array ops * feat: math operator icons and refactor variable update renderers for improved layout and consistency * chore(workflow): clean up variable update types and restore icon cleanup * feat: add test * fix:md_ascii_bug (#6755) * md_ascii_bug * md_ascii_bug * md_ascii_bug * md_ascii_bug * md_ascii_bug * perf: test --------- Co-authored-by: archer <545436317@qq.com> * doc * del dataset * perf: date auto coerce * doc * add test * perf: channel setting * doc * fix: chat resume stream (#6759) * refactor(api): move stream resume to /api/core/chat/resume Relocate resume handler from pages/api/v2 to pages/api/core, update OpenAPI paths, frontend streamResumeFetch URL, tests, and comments. Made-with: Cursor * fix: remove stray conflict markers; use z.nativeEnum for chatGenerateStatus Made-with: Cursor * fix: use enum instead of nativeEnum * fix(chat): address resume review suggestions * fix(chat): require sse when resuming generating chats * revert(chat): keep chatitem dataId index non-unique * fix: ts * fix doc * fix(chat): gate stream resume mirror by header (#6760) * fix: remove stray conflict markers; use z.nativeEnum for chatGenerateStatus Made-with: Cursor * fix: use enum instead of nativeEnum * fix(chat): address resume review suggestions * fix(chat): require sse when resuming generating chats * feat(chat): gate stream resume mirror by header * refactor(chat): decouple resume mirror header parsing * perf: dataset queue * fix: multipleselect * perf: workflow bug * doc * doc * perf: deploy yml;fix: child nodes watch * adapt embedding model defaultconfig * install shell * add mcp zod check * feat: http tool zod schema * Feat/batch UI (#6763) * feat: aggregate parallel run results into task-specific virtual nodes and update UI to support i18n arguments for module names * style: update workflow node card padding and table styling for improved layout consistency * feat: implement parallel run workflow node with documentation and i18n support * style(modal): WholeResponseModal UI and layout styling * chore: improve chat resume UX (#6764) * fix: remove stray conflict markers; use z.nativeEnum for chatGenerateStatus Made-with: Cursor * fix: use enum instead of nativeEnum * fix(chat): address resume review suggestions * fix(chat): require sse when resuming generating chats * feat(chat): gate stream resume mirror by header * refactor(chat): decouple resume mirror header parsing * feat: improve stream resume fallback * feat: block duplicate chat generation * feat: polish resume unavailable recovery * test: stabilize resume stream timeout * fix: harden resume wait flow * fix: get mcp tool raw schema * style: update UI styling and layout for LLM request detail and response modals * perf: http tool * fix: test * fix: http raw schema * fix: test * deploy yml * deploy yml --------- Co-authored-by: DigHuang <114602213+DigHuang@users.noreply.github.com> Co-authored-by: Ryo <whoeverimf5@gmail.com> Co-authored-by: YeYuheng <57035043+YYH211@users.noreply.github.com>
312 lines
9.0 KiB
Markdown
312 lines
9.0 KiB
Markdown
# FastGPT Code Sandbox
|
||
|
||
基于 Bun + Hono 的代码执行沙盒,支持 JS 和 Python。采用进程池架构,预热长驻 worker 进程,通过 stdin/stdout JSON 协议通信,消除每次请求的进程启动开销。
|
||
|
||
## 架构
|
||
|
||
```
|
||
HTTP Request → Hono Server → Process Pool → Worker (long-lived) → Result
|
||
↓
|
||
┌──────────────┐
|
||
│ JS Workers │ bun run worker.ts (×N)
|
||
│ Py Workers │ python3 worker.py (×N)
|
||
└──────────────┘
|
||
stdin: JSON task → stdout: JSON result
|
||
```
|
||
|
||
- **进程池**:启动时预热 N 个 worker 进程(默认 20),请求到达时直接分配空闲 worker,执行完归还池中
|
||
- **JS 执行**:Bun worker 进程 + 安全 shim(禁用 Bun API、冻结 Function 构造器、require 白名单)
|
||
- **Python 执行**:python3 worker 进程 + `__import__` 拦截 + resource 资源限制
|
||
- **网络请求**:统一通过 `SystemHelper.httpRequest()` / `system_helper.http_request()` 收口,内置 SSRF 防护
|
||
- **并发控制**:请求数超过池大小时自动排队,worker 崩溃自动重启补充
|
||
|
||
## 性能
|
||
|
||
进程池 vs 旧版 spawn-per-request 对比(SANDBOX_POOL_SIZE=20):
|
||
|
||
| 场景 | 旧版 QPS / P50 | 进程池 QPS / P50 | 提升 |
|
||
|------|----------------|------------------|------|
|
||
| JS 简单函数 (c50) | 22 / 1,938ms | 1,414 / 7ms | **64x** |
|
||
| JS IO 500ms (c50) | 22 / 2,107ms | 38 / 1,005ms | 1.7x |
|
||
| JS 高 CPU (c10) | 9 / 1,079ms | 12 / 796ms | 1.3x |
|
||
| JS 高内存 (c10) | — | 13 / 787ms | — |
|
||
| Python 简单函数 (c50) | 14.7 / 2,897ms | 4,247 / 4ms | **289x** |
|
||
| Python IO 500ms (c50) | 14.2 / 3,066ms | 38 / 1,003ms | 2.7x |
|
||
| Python 高 CPU (c10) | 3.1 / 2,845ms | 4 / 2,191ms | 1.3x |
|
||
| Python 高内存 (c10) | — | 11 / 893ms | — |
|
||
|
||
资源占用(20+20 workers):空闲 ~1.5GB RSS,压测峰值 ~2GB RSS。
|
||
|
||
## 快速开始
|
||
|
||
```bash
|
||
# 安装依赖
|
||
bun install
|
||
|
||
# 开发运行
|
||
bun run src/index.ts
|
||
|
||
# 运行测试
|
||
bun run test
|
||
```
|
||
|
||
## Docker
|
||
|
||
```bash
|
||
# 构建
|
||
docker build -f projects/code-sandbox/Dockerfile -t fastgpt-code-sandbox .
|
||
|
||
# 运行
|
||
docker run -p 3000:3000 \
|
||
-e SANDBOX_TOKEN=your-secret-token \
|
||
-e SANDBOX_POOL_SIZE=20 \
|
||
fastgpt-code-sandbox
|
||
```
|
||
|
||
## API
|
||
|
||
### `POST /sandbox/js`
|
||
|
||
执行 JavaScript 代码。
|
||
|
||
```json
|
||
{
|
||
"code": "async function main(variables) {\n return { result: variables.a + variables.b }\n}",
|
||
"variables": { "a": 1, "b": 2 }
|
||
}
|
||
```
|
||
|
||
### `POST /sandbox/python`
|
||
|
||
执行 Python 代码。
|
||
|
||
```json
|
||
{
|
||
"code": "def main(variables):\n return {'result': variables['a'] + variables['b']}",
|
||
"variables": { "a": 1, "b": 2 }
|
||
}
|
||
```
|
||
|
||
### `GET /health`
|
||
|
||
健康检查,返回进程池状态。
|
||
|
||
```json
|
||
{
|
||
"status": "ok",
|
||
"version": "5.0.0",
|
||
"jsPool": { "total": 20, "idle": 18, "busy": 2, "queued": 0 },
|
||
"pythonPool": { "total": 20, "idle": 20, "busy": 0, "queued": 0 }
|
||
}
|
||
```
|
||
|
||
### 响应格式
|
||
|
||
成功:
|
||
|
||
```json
|
||
{
|
||
"success": true,
|
||
"data": {
|
||
"codeReturn": { "result": 3 },
|
||
"log": "console.log 输出内容"
|
||
}
|
||
}
|
||
```
|
||
|
||
失败:
|
||
|
||
```json
|
||
{
|
||
"success": false,
|
||
"message": "错误信息"
|
||
}
|
||
```
|
||
|
||
## 环境变量
|
||
|
||
### 服务配置
|
||
|
||
| 变量 | 说明 | 默认值 |
|
||
|------|------|--------|
|
||
| `SANDBOX_PORT` | 服务端口 | `3000` |
|
||
| `SANDBOX_TOKEN` | Bearer Token 认证密钥 | 空(不鉴权) |
|
||
|
||
### 进程池
|
||
|
||
| 变量 | 说明 | 默认值 |
|
||
|------|------|--------|
|
||
| `SANDBOX_POOL_SIZE` | 每种语言的 worker 进程数 | `20` |
|
||
|
||
### 资源限制
|
||
|
||
| 变量 | 说明 | 默认值 |
|
||
|------|------|--------|
|
||
| `SANDBOX_MAX_TIMEOUT` | 超时上限(ms),请求不可超过此值 | `60000` |
|
||
| `SANDBOX_MAX_MEMORY_MB` | 内存上限(MB) | `256` |
|
||
|
||
### 网络请求限制
|
||
|
||
| 变量 | 说明 | 默认值 |
|
||
|------|------|--------|
|
||
| `SANDBOX_REQUEST_MAX_COUNT` | 单次执行最大 HTTP 请求数 | `30` |
|
||
| `SANDBOX_REQUEST_TIMEOUT` | 单次 HTTP 请求超时(ms) | `60000` |
|
||
| `SANDBOX_REQUEST_MAX_RESPONSE_MB` | 最大响应体大小(MB) | `10` |
|
||
| `SANDBOX_REQUEST_MAX_BODY_MB` | 最大请求体大小(MB) | `5` |
|
||
|
||
## 项目结构
|
||
|
||
```
|
||
src/
|
||
├── index.ts # 入口:Hono 服务 + 进程池初始化
|
||
├── env.ts # 环境变量校验(zod)
|
||
├── config.ts # 配置导出
|
||
├── types.ts # 类型定义
|
||
├── pool/
|
||
│ ├── process-pool.ts # JS 进程池管理
|
||
│ ├── python-process-pool.ts # Python 进程池管理
|
||
│ ├── worker.ts # JS worker(长驻进程,含安全 shim)
|
||
│ └── worker.py # Python worker(长驻进程,含安全沙箱)
|
||
└── utils/
|
||
└── semaphore.ts # 信号量(通用并发控制)
|
||
|
||
test/
|
||
├── unit/ # 单元测试(进程池、信号量)
|
||
├── integration/ # 集成测试(API 路由)
|
||
├── boundary/ # 边界测试(超时、内存限制)
|
||
├── security/ # 安全测试(沙箱逃逸防护)
|
||
├── compat/ # 兼容性测试(旧版代码格式)
|
||
├── examples/ # 示例测试(常用包)
|
||
└── benchmark/ # 压测脚本
|
||
```
|
||
|
||
## 添加 JS 包
|
||
|
||
沙盒内的 JS 代码通过 `require()` 加载包,但仅允许白名单内的包。
|
||
|
||
### 当前白名单
|
||
|
||
`lodash`、`dayjs`、`moment`、`uuid`、`crypto-js`、`qs`、`url`、`querystring`
|
||
|
||
### 添加新包步骤
|
||
|
||
1. **安装包**:
|
||
|
||
```bash
|
||
cd projects/code-sandbox
|
||
bun add <package-name>
|
||
```
|
||
|
||
2. **加入白名单**(环境变量 `SANDBOX_JS_ALLOWED_MODULES`):
|
||
|
||
在逗号分隔列表中添加包名:
|
||
|
||
```bash
|
||
SANDBOX_JS_ALLOWED_MODULES=lodash,dayjs,moment,uuid,crypto-js,qs,url,querystring,your-new-package
|
||
```
|
||
|
||
3. **重新构建 Docker 镜像**。
|
||
|
||
### 注意事项
|
||
|
||
- 只添加纯计算类的包,不要添加有网络/文件系统/子进程能力的包
|
||
- 包会被打入 Docker 镜像,注意体积
|
||
- 网络请求统一走 `SystemHelper.httpRequest()`,不要放行 `axios`、`node-fetch` 等网络库
|
||
|
||
## 添加 Python 包
|
||
|
||
### 当前预装包
|
||
|
||
`numpy`、`pandas`(通过 `requirements.txt` 安装)
|
||
|
||
### 添加新包步骤
|
||
|
||
1. **编辑 `requirements.txt`**:
|
||
|
||
```
|
||
numpy
|
||
pandas
|
||
your-new-package
|
||
```
|
||
|
||
2. **加入白名单**(环境变量 `SANDBOX_PYTHON_ALLOWED_MODULES`):
|
||
|
||
在逗号分隔列表中添加包名。如果新包依赖了黑名单中的模块(如 `os`),标准库路径的间接导入会自动放行,无需额外配置。
|
||
|
||
3. **重新构建 Docker 镜像**。
|
||
|
||
### 注意事项
|
||
|
||
- Python 的模块黑名单通过 `__import__` 拦截实现,只拦截用户代码的直接 import
|
||
- 标准库和第三方包的内部间接 import 不受影响
|
||
- 危险模块(`os`、`sys`、`subprocess`、`socket` 等)始终被拦截
|
||
|
||
## 安全机制
|
||
|
||
### JS
|
||
|
||
- `require()` 白名单,非白名单模块直接拒绝
|
||
- `Bun.spawn`、`Bun.write`、`Bun.serve` 等 API 禁用
|
||
- `Function` 构造器冻结,阻止 `constructor.constructor` 逃逸
|
||
- `process.env` 清理,仅保留必要变量
|
||
- `fetch`、`XMLHttpRequest`、`WebSocket` 禁用
|
||
|
||
### Python
|
||
|
||
- `__import__` 黑名单拦截:用户代码无法 import 危险模块(`os`、`sys`、`subprocess` 等)
|
||
- `exec()`/`eval()` 内的 import 同样被拦截(基于调用栈帧检测)
|
||
- `builtins.__import__` 通过代理对象保护,用户无法覆盖
|
||
- `signal.SIGALRM` 超时保护
|
||
|
||
### 网络
|
||
|
||
- 所有网络请求通过 `httpRequest()` 收口
|
||
- 内网 IP 黑名单:`10.0.0.0/8`、`172.16.0.0/12`、`192.168.0.0/16`、`127.0.0.0/8`、`169.254.0.0/16`
|
||
- 仅允许 `http:` / `https:` 协议
|
||
- 单次执行请求数、响应体大小、超时均有限制
|
||
|
||
## 内置函数
|
||
|
||
### JS(全局可用)
|
||
|
||
| 函数 | 说明 |
|
||
|------|------|
|
||
| `SystemHelper.httpRequest(url, opts?)` | HTTP 请求(opts: `{method, headers, body, timeout}`) |
|
||
|
||
### Python(全局可用)
|
||
|
||
| 函数 | 说明 |
|
||
|------|------|
|
||
| `SystemHelper.httpRequest(url, opts?)` | HTTP 请求(opts: `{method, headers, body, timeout}`) |
|
||
|
||
## 测试
|
||
|
||
```bash
|
||
# 全部测试(332 cases)
|
||
bun run test
|
||
|
||
# 单个文件
|
||
bunx vitest run test/security/security.test.ts
|
||
|
||
# 带详细输出
|
||
bunx vitest run --reporter=verbose
|
||
|
||
# 压测(需先启动服务)
|
||
bash test/benchmark/bench-sandbox.sh
|
||
bash test/benchmark/bench-sandbox-python.sh
|
||
```
|
||
|
||
测试配置:串行执行(`fileParallelism: false`),池大小 1(避免资源竞争)。
|
||
|
||
测试覆盖维度:
|
||
|
||
| 分类 | 文件数 | 用例数 | 说明 |
|
||
|------|--------|--------|------|
|
||
| 单元测试 | 4 | 43 | 进程池生命周期/恢复/健康检查、Semaphore 并发控制 |
|
||
| 集成测试 | 2 | 53 | HTTP API 路由、JS/Python 功能验证 |
|
||
| 安全测试 | 1 | 102 | 模块拦截、逃逸攻击、SSRF 防护、注入攻击 |
|
||
| 边界测试 | 1 | 58 | 空输入、超时、大数据、类型边界 |
|
||
| 兼容性测试 | 2 | 39 | 旧版 JS/Python 代码格式兼容 |
|
||
| 示例测试 | 1 | 31 | 常用场景和第三方包 |
|
||
|