v1.1.0-beta.8 (2026-07-26)
- Added fully offline OCR for image attachments (Light-OCR): images are routed to vision or OCR based on the selected model, OCR text is reused across history, retry, edit and compaction, and per-attachment mode, cancellation and cache controls are available in Settings
- Gave every Agent its own Skills root and catalog, replacing the shared catalog and its sync workflow with explicit import from another Agent or from an external DeepChat data directory
- Rebuilt the Agent browser picture-in-picture as a native overlay: open-in-side-panel and close controls, dragging outside the DeepChat window, with the Canvas overlay kept as a fallback
- Added Routerra, Straico, and StepFun Token Plan as built-in provider presets
- Added Linux ARM64 installers, and gave the macOS DMG installer a proper drag-to-install window with a background image and the app and Applications icons aligned to it
- Made long conversations reuse the provider prompt cache: the system prompt no longer changes whenever a summary, handoff state or memory updates, cache markers now reach the request for OpenAI, Anthropic, Bedrock, OpenRouter and Zenmux, and compaction sizes the retained history to the model's context budget
- Retried transient provider failures with bounded, abortable backoff, and never replayed a request once text, reasoning or tool output was committed
- Stopped executing tool calls truncated by
max_tokens: the batch is rejected atomically, matching error results are written, and the request is retried once - Restructured the Agent runtime into a one-directional owner graph behind a harness facade, added a typed tool execution contract that keeps write-capable tools sequential, and rebuilt Hooks on a typed event pipeline with bounded command execution
- Fixed chat search highlights on streamed rows, sidebar search pagination, the invisible pending-assistant row, session-restore scrolling that overrode a search or spotlight jump, and session revision recovery on existing databases
- Cut Settings startup cost by lazy-loading locales (synchronous JS 4.25 MB to 0.9 MB) and reduced Dashboard CPU with cached formatters and polling that pauses in the background
- Fixed auto-update failing with
Invalid input: expected string, received Date: updater metadata is now published with a quoted releaseDate and validated against electron-updater's own parser across all four channel files - Fixed the macOS clean-install Gatekeeper failure caused by the bundled Computer Use helper, and added distribution verification for both the app and the macOS updater payload
- Rebuilt the release pipeline to verify every published package before release, and locked dependencies with a tracked pnpm lockfile for reproducible builds
- 新增完全离线的图片附件 OCR(Light-OCR)：根据所选模型自动在视觉与 OCR 之间路由，OCR 文本在历史、重试、编辑与压缩中复用，并提供逐附件模式、取消与缓存管理
- 每个 Agent 拥有独立的 Skills 根目录与目录管理，取消共享目录与同步流程，改为从其他 Agent 或外部 DeepChat 数据目录显式导入
- 将 Agent 浏览器画中画重构为原生浮窗：提供打开侧边栏与关闭按钮，可拖动到 DeepChat 窗口之外，并保留 Canvas 浮层作为回退方案
- 新增 Routerra、Straico 与 StepFun Token Plan 内置 Provider
- 新增 Linux ARM64 安装包；macOS DMG 安装窗口改为带背景图的拖拽安装界面，应用图标与 Applications 快捷方式按背景对齐
- 让长对话真正复用 Provider 的提示词缓存：系统提示词不再随摘要、handoff 状态与记忆的更新而变化，缓存标记能正确传给 OpenAI、Anthropic、Bedrock、OpenRouter 与 Zenmux，压缩时按模型上下文预算保留近期历史
- 对瞬时 Provider 失败进行有界、可中断的退避重试，并在已产出文本、思考或工具调用后不再重放请求
- 不再执行被
max_tokens截断的工具调用：整批原子拒绝并写入对应错误结果，随后仅自动重试一次 - 将 Agent runtime 重构为 harness facade 之上的单向 owner 图，引入类型化工具执行契约(可写工具强制串行)，并将 Hook 重建为类型化事件管道与有界命令执行
- 修复流式消息的搜索高亮、侧边栏搜索分页、待生成消息行不可见、会话恢复滚动覆盖搜索/聚光跳转，以及已有数据库的 session revision 恢复问题
- 懒加载语言包降低设置页启动开销(同步 JS 4.25 MB 降至 0.9 MB)，并通过缓存 formatter 与后台暂停轮询降低 Dashboard CPU 占用
- 修复自动更新报错
Invalid input: expected string, received Date：发布端强制为 releaseDate 加引号，并用 electron-updater 自身的解析器校验全部四个更新渠道文件 - 修复由内置 Computer Use 助手导致的 macOS 全新安装 Gatekeeper 失败，并对应用与 macOS 更新包补充分发校验
- 重构发布流水线，发布前校验每个产物；纳入 pnpm lockfile 锁定依赖，保证构建可复现