知识库 · AI 终端 · 数字人多会话 · 界面缩放 · 图片文字识别，以及多项修复
Knowledge base · AI terminal · Multi-session digital humans · Display scaling · OCR, plus fixes
- 知识库：把文件、文件夹、PDF、Office 文档导入知识库，自动整理成可检索的知识，聊天时直接引用并标注出处；支持监视文件夹自动更新、一键清空重学，远程网页端同样可用
- AI 终端：内置终端运行在独立进程中，卡死或崩溃不再影响主界面，入口移到右侧工作区
- 数字人升级：支持多会话同时对话、可把任务接回客户端继续；能按应用单独配置能力、MCP 与技能；停止进行中的 IM 会话不再丢失历史
- 远程访问：支持固定地址的长期隧道与无界面服务器模式，多设备接入自动仲裁，访问码规则放宽
- 界面缩放：设置中可在 50%-150% 之间调节，支持 Cmd/Ctrl 加号、减号、0 快捷键，重启后保持
- 模型与登录：新增 GLM 编程套餐登录、Atlas Cloud 内置源，可按模型单独指定是否支持图片，额度用量直接显示在来源上，首次配置流程简化
- 图片文字识别：内置 OCR，可直接读取截图和扫描件中的文字
- 其他改进：浏览器 User-Agent 可自定义、Canvas 文件标签支持刷新与自动重载、MCP 支持连接检测与多服务器批量导入、企业微信群回复可关闭引用气泡、凭据本地加密存储
- 稳定性修复：修复 OpenAI 兼容接口的用量统计与 max_tokens 传递、Windows 后台进程弹出控制台窗口、图片查看器缓存不刷新、创建空间时中文输入法回车误提交等问题
- Knowledge base: import files, folders, PDFs and Office documents; Halo organises them into searchable knowledge and cites sources inline in chat. Watched folders update automatically, one click clears and relearns, and it all works in the remote web client too
- AI terminal: the built-in terminal now runs in its own process, so a hang or crash no longer affects the main window. Entry point moved to the right-hand workspace
- Digital humans: hold several conversations at once and hand a task back to the desktop client to continue; configure capabilities, MCP servers and skills per app; stopping a running IM session no longer loses its history
- Remote access: long-lived tunnels with a fixed address, a headless server mode, automatic arbitration across multiple devices, and a relaxed access-code policy
- Display scaling: adjustable from 50% to 150% in Settings, with Cmd/Ctrl plus, minus and 0 shortcuts, persisted across restarts
- Models and sign-in: GLM coding plan login, Atlas Cloud as a built-in source, per-model vision override, quota usage shown directly on the source, and a simplified first-run setup
- OCR: read text straight out of screenshots and scanned documents
- Also: customisable browser User-Agent, refresh and auto-reload for Canvas file tabs, MCP connection probing and multi-server import, optional quote bubbles for WeCom group replies, and encrypted local credential storage
- Fixes: usage reporting and max_tokens forwarding for OpenAI-compatible endpoints, console windows popping up from Windows background processes, stale image-viewer cache, and the create-space dialog submitting early during Chinese IME input