刚刚，OpenClaw最猛升级！底层架构大换血，全网等了9天

Core Insights - OpenClaw has released a significant update with version 3.22, featuring a complete overhaul of its plugin architecture and the introduction of GPT-5.4, alongside enhanced security measures [1][2][7] Plugin System Overhaul - The plugin system has undergone a major reconstruction, with the old API being completely removed and replaced by a new modular interface, openclaw/plugin-sdk [4][15] - ClawHub has been established as the preferred distribution channel for plugins, enhancing the ecosystem's integrity by ensuring higher quality control compared to npm [5][19] - All existing third-party plugins using the old API must migrate to the new system, indicating a decisive shift in OpenClaw's plugin ecosystem [16][20] Security Enhancements - The update addresses critical security vulnerabilities, including the prevention of Windows credential leaks and the hardening of execution environments [6][22] - Over ten security patches have been implemented, targeting issues such as environment variable injection and Unicode character exploitation [24][30] - The update is deemed essential for users deploying OpenClaw in public environments, emphasizing the urgency of these security fixes [32] Model Ecosystem Expansion - The default model has been upgraded to GPT-5.4, with backward compatibility for gpt-5.4-mini and gpt-5.4-nano [35] - MiniMax's default model has been upgraded from M2.5 to M2.7, simplifying the configuration process by merging API and OAuth into a single plugin [36] - Anthropic Vertex has been integrated, allowing direct access to Claude models via Google Vertex AI, which is particularly beneficial for teams operating on Google Cloud [37][38] Multi-Platform Experience Improvements - The Android version now supports a system-wide dark mode, enhancing user experience across various interfaces [42] - Telegram has introduced automatic topic renaming for DM forums, improving usability by generating meaningful topic tags [43] - The update also includes enhancements in Feishu (Lark) with structured interaction approval cards and real-time reasoning display [44] Agent Engine Enhancements - The long conversation compression mechanism has been iterated to improve efficiency, with the default agent timeout extended from 600 seconds to 48 hours [52] - A new command, /btw, allows users to insert side questions during conversations without disrupting the context [53] - These improvements reflect a shift towards a more reliable AI agent platform, indicating a maturation in OpenClaw's development approach [57][60]