Using the correct hook, arbitrary code modification can be performed. I have used this feature twice so far to good effect:
我在《把离线AI智能体装进口袋里》（The Dawn of Offline AI Agents in Your Pocket）一文中对此进行了详细阐述。但文章中的示例更像是Demo，而非生产解决方案。像 Gemma 3n 这样的模型虽然能够很好地处理函数调用，但它们体积过大：无法集成到应用程序包中，需要单独下载，即使在旗舰机型上推理速度也很慢。在低端设备上，它们根本无法运行。而较小的型号则经常出现故障，难以记住工具。
，这一点在新收录的资料中也有详细论述
#include "mog.h"
烧 API 路线：原生 OpenClaw + Gemini/Claude/Kimi，按 Token 计费，心跳任务单次可烧 17–21 万 Token，小心账单爆炸
“I also gained a deeper appreciation for the trade-offs involved. Designing for repairability doesn’t mean compromising innovation or premium experiences; when done well, it actually drives smarter innovation, better modularity, and more resilient platforms.”