Pinned Tweet
Full Android running on iOS. No jailbreak. No JIT.
• Boots in 3s
• Hardware GPU acceleration (up to 120FPS)
• Native ProMotion support
Public testing is officially open. Try it out 👇
Forget SuperSU and Magisk—you’re definitely talking to the right guy 😉 KernelSU is coming soon!
Replying to @weishu1024
and please add Root (supersu or Magisk) option and Gapps or what ever can loging account google. Thank you so much ❤️❤️
1
5
728
Here’s the TestFlight link! Note that this is an older build—we’ve already fixed a lot of bugs since then, and a new version is coming soon.
testflight.apple.com/join/Eh…
Everyone is building mobile AI Agents right now, so I built something different: Twoyi.
Instead of needing Root/ADB, cloud phones, or buying a second device, Twoyi runs a full, isolated Android OS inside your phone—exclusively for your AI Agent.
1. Zero Root / ADB required
2. Complete privacy isolation (empty OS by default, you decide which apps to import)
3. True multitasking: use your phone normally while the AI works in the background
4. Human-in-the-loop: 1-tap takeover anytime, strict confirmation for payments/passwords
5. BYOK (Bring Your Own Key, supports DeepSeek, OpenAI, etc.)
weishu retweeted
这篇文章是今天看到最精彩的文章,是一篇技术深度逆向工程实验,成功从 Codex CLI 的 compact() API中提取出了隐藏的系统提示。
非常好地解释了 Codex CLI 在使用自己模型,在长上下文管理上的实际运行方式,几乎把黑箱内部机制扒得一清二楚。
1 当上下文太长时,不是简单丢弃历史,而是调用 compact() API
2 服务器用一个专用的 compactor LLM,把整个对话历史总结成一个精炼的“handoff summary”。
3 这个 summary 被 AES 加密成 blob 返回给客户端
4 下次 responses.create() 时,客户端把 blob 传回去,服务器解密后总结,一起喂给 codex 主模型。
5 模型无缝接力继续工作,不会因为压缩而严重失忆或漂移。
这解释了为什么 Codex 在长任务上连续性特别强。
配图也很棒,非常好的科普文章。









