Focus: AI Agent Runtime & Execution Safety
专注方向：AI Agent 运行时与执行安全 —— 让 Agent 跑得起、跑得稳、跑得安全
Recent work · 最近在做：openclaw streaming throughput · OpenSandbox SDK security · loopx Windows coordination
| Direction · 方向 | What I work on · 我在做什么 |
|---|---|
| 🧠 Agent Runtime & Reliability 运行时与可靠性 |
Session / goal lifecycle, scheduling, locks & leases, quotas, streaming stability 会话与目标生命周期、调度、锁与租约、配额、流式稳定性 |
| 🛡️ Sandboxing & Execution Security 沙箱与执行安全 |
Sandbox lifecycle, egress / ingress policy, secret hygiene, multi-language SDK & CLI 沙箱生命周期、网络出入口策略、密钥卫生、多语言 SDK 与 CLI |
| 🖥️ Cross-platform Engineering 跨平台工程 |
Making agent tooling rock-solid on Windows / WSL, CI health & verification gates 让 Agent 工具在 Windows / WSL 上稳定可用，CI 健康与验证门禁 |
Rolling 12 months · 近 12 个月（截至 2026-10-27）
| Contributions ¹ | Commits | PRs Merged | Issues |
|---|---|---|---|
| 3,474 | 2,254 | 146 | 171 |
¹ Includes private contributions · 含私密贡献；GitHub contribution graph, last 12 months
| Project · 项目 | ⭐ | What I did · 我做了什么 |
|---|---|---|
| openclaw/openclaw | 393K | Removed streaming re-projection and reply stalls at HTML, JSX, code-block and list boundaries; contained tool-call text leaks and repeated runtime-context scrubbing 消除 HTML、JSX、代码块与列表边界的流式重投影和回复卡顿；收敛工具调用文本泄漏与重复上下文清洗 |
| bytedance/deer-flow | 84.5K | Native Windows agent spawn / PATH resolution; contained malformed DuckDuckGo and Tavily payloads; hardened oversized-payload and symlink tests 原生 Windows Agent 启动与 PATH 解析；容错异常 DuckDuckGo / Tavily 载荷；加固超大载荷与 symlink 测试 |
| agentscope-ai/QwenPaw | 36.5K | Atomic session-state persistence; aligned channel config types with the built-in registry 会话状态原子化持久化；通道配置类型与内置注册表对齐 |
| opensandbox-group/OpenSandbox | 16.2K | Go / JS / Kotlin / C# SDK hardening: tenant-key isolation, SSE wake retries, idempotent-write handling, file-transfer timeouts and non-UTF-8 CLI output Go / JS / Kotlin / C# SDK 加固：租户密钥隔离、SSE 唤醒重试、非幂等写入处理、文件传输超时与非 UTF-8 输出兼容 |
| loopx-project/loopx | 6.4K | Windows machine identity for task leases, interrupted Goal recovery, UTF-8 report artifacts, CI & canary gates, signed NSIS updates 任务租约的 Windows 机器标识、中断 Goal 恢复、UTF-8 报告产物、CI 与 canary 门禁、NSIS 签名更新 |
| apache/maka | 5.9K | Plugged secret-redaction gaps (fine-grained GitHub PATs, shared redactors); bounded Electron install retries in CI 补齐密钥脱敏漏洞（细粒度 GitHub PAT、共享脱敏器）；CI 中为 Electron 安装增加限时重试 |
| the-open-agent/openagent | 5.7K | Auto chat titles with fallback, usage metadata resilience, a Windows GUI (UIA) automation tool, config fixes 聊天标题自动生成与回退、用量元数据容错、Windows GUI（UIA）自动化工具、配置修复 |
| FireRedTeam/FireRed-OpenStoryline | 3.6K | Session-restore invariants (fail-closed on corrupt rows), config-secret redaction, drag-and-drop upload, typed errors & retries 会话恢复不变量（损坏数据 fail-closed）、配置密钥脱敏、拖拽上传、类型化错误与重试 |
| 0xsline/OpenChatCut | 2.7K | Windows verification-suite isolation, Ctrl-chord shortcuts, whisper.cpp token grouping, local media import Windows 验证套件隔离、Ctrl 组合快捷键、whisper.cpp 分词、本地媒体导入 |
… and 15+ more repositories — HiRoute, agentbench, casdoor, langchain-ai/deepagents and others. 另外还有 15+ 个仓库的贡献。
- Member of @the-open-agent and @effectorHQ · 两个开源组织的成员
- Long-term goal: keep making AI agents run fast, steady and safe in open source
长期目标：持续让 AI Agent 在开源世界里跑得稳、跑得安全 - Off the keyboard: football ⚽, exploring new places 🌍, meeting people with stories ✨
代码之外：足球、到处走走、认识有故事的人 - Always learning — there will always be people better than me, and that's a good thing.
永远在学习——总有人比我更强，而这正是好事。
Happy to talk about agents, sandboxes and runtime reliability — or just to say hi.
欢迎交流 AI Agent、沙箱与运行时可靠性——也欢迎单纯打个招呼。如果我的工作对你有帮助，一颗 ⭐ 或一个关注就是最大的鼓励。