v1.1.4 (2026-11-12)
- Refreshed the desktop interface with a unified sidebar, searchable settings, compact ACP Agent details, and a message minimap with scrollable previews
- Moved conversation search into global search, prioritized matching titles, and fixed accidental selection when confirming input method text
- Select an Agent in the new-chat composer while preserving Agent-specific drafts, models, and workspace defaults. Open the workspace and browser before sending the first message, and copy session references between chats
- Added opt-in interactive HTML visualizations in chat and moved Markdown infographics into an isolated renderer
- Added per-Agent tool mode and proactive collaboration settings, with proactive collaboration enabled for new Agents. Default output budgets now use the model's output limit within the context budget. Fixed subagent model selection, resumed execution settings, stop reason reporting, and access from Code Mode
- Separated the built-in Agent runtime from desktop dependencies and verified persistence, permissions, cancellation, and recovery in standalone Node.js. Unified MCP settings across windows and consolidated chat session state
- Improved provider model discovery and refresh, connection-test model selection, and GitHub Copilot sign-in. Added providers including Atlas Cloud, LiteLLM, Wally, Standard Compute, UberLLM, OpenLLM, and Wiro AI
- Fixed MCP OAuth issuer and callback handling, and added optional MCP presets for search, media generation, and hosting, including Parallel Search, Webz, EvoLink, deed.page, Bowmark, and Sofya
- Improved file watcher recovery and reduced unnecessary Git watching. Fixed question dialogs lingering after an answer and preserved literal braces in HTML chat exports
- Completed French interface and native menu translations, updated Electron and bundled runtimes, removed built-in RTK command rewriting, and refreshed the model and ACP catalogs
- 刷新桌面界面，统一侧栏，支持搜索设置，精简 ACP Agent 详情，并为消息地图增加可滚动预览
- 将会话搜索整合到全局搜索，优先展示标题匹配的会话，并修复使用输入法确认文字时误选结果的问题
- 可在新对话输入框直接选择 Agent，保留各 Agent 的草稿、模型和工作区默认值。发送首条消息前即可打开工作区和浏览器，也可在对话间复制会话引用
- 对话支持按需启用交互式 HTML 可视化，Markdown 信息图在隔离环境中渲染
- 支持为各 Agent 设置工具模式和主动协作，新建 Agent 默认开启主动协作。默认输出预算在上下文预算内采用模型的输出上限。修复子代理模型选择、恢复执行时的配置、停止原因上报，以及 Code Mode 调用子代理的问题
- 将内置 Agent 运行时与桌面依赖解耦，在独立 Node.js 环境验证持久化、权限、取消与恢复。统一各窗口的 MCP 设置，并整合会话状态
- 改善服务商模型发现与刷新、连接测试中的模型选择，以及 GitHub Copilot 登录。新增 Atlas Cloud、LiteLLM、Wally、Standard Compute、UberLLM、OpenLLM 和 Wiro AI 等服务商
- 修复 MCP OAuth 授权服务地址匹配与回调处理，新增 Parallel Search、Webz、EvoLink、deed.page、Bowmark、Sofya 等可选 MCP 预设，覆盖搜索、媒体生成和托管
- 改善文件监听的故障恢复，缩小 Git 监听范围。修复回答后提问对话框滞留的问题，并在导出 HTML 对话时保留原始花括号
- 补全法语界面和原生菜单翻译，升级 Electron 与内置运行时，移除内置 RTK 命令改写，并刷新模型和 ACP 目录