Skip to content

v5.3.14

Choose a tag to compare

View all tags
@github-actions github-actions released this 22 Oct 15:15
· 483 commits to main since this release
v5.3.14
220071c

v5.3.14 — 2026-10-22

主要改进

  • AI 回复支持展示图片：回复里的 Markdown 图片此前只有部分来源能渲染。本次打通来源解析——/workspace、/var/minis、minis://、file:// 经 LinuxGuestPathResolver 折算到宿主（与 read_image 同一映射），Linux 工具环境里生成的图片无需拷贝即可展示；无 scheme 的相对路径按工作区根解析；行内渲染与画廊预览的判定统一到同一来源白名单；提示词在 read_image 段落补充了贴图语法，模型被明确告知可以贴图。

  • 端侧设备动作模型：接入 FunctionGemma 270M Mobile Actions——ModelScope / Hugging Face 双源下载（校验、取消、令牌管理），注册 device_actions 工具，把英文指令变成手电筒、联系人、邮件、地图、Wi-Fi、日历六类真实设备动作；设置页管理模型与推理后端（Auto/GPU/CPU），独立试用页可直接下发指令。R8 保留反射面与 LiteRT-LM 原生库声明。

  • 回答后自动收起思考与工具步骤，并升级三档分级：一轮回复的正文一出现，同轮排在它前面的思考与工具组整组收进「已完成 N 个步骤」卡片，点一下仍可展开；手动展开过的组不再被自动收起。折叠行为从布尔开关升级为三档（激进折叠 / 普通折叠 / 不折叠）：激进档按真实用户消息划轮，本轮最终回复之前的中间回复与痕迹并进同一张工作卡片，问答卡、建议芯片与流式回复保持可见，整轮没有正文则不收。

  • 媒体检索批量操作与阈值偏好：检索页补多选（全选 / 反选 / 退出多选）与批量移动 / 复制（目录选择器 + SAF 授权持久化 + 二次确认，部分失败按条数汇总）；新增 set_threshold 工具与相似度阈值滑杆共用同一偏好，search 省略 min_score 时按偏好过滤；新增 batch_act 工具支持按 source_keys 或现场检索后批量执行。

  • 语音设置改个人音色子页结构：删除旧 SpeechSettingsScreen，「我的声音」内豆包 / Mimo 音色独立子页（复用 NavHost 入栈 / 出栈转场）；豆包音色表单样式与点按触感统一；语音入口策略、模式控制器补充联动开关。

  • 子代理设置改预设详情页：预设编辑从行内改成独立详情页，设置页只留列表与入口；资料配置弹窗拆出任务分工 / 思考档位 / 分辨率 / gpt 速度子页。

  • 输入区与跟底动效：输入条工具按钮收敛进两级选择器，只留常用操作；跟底时新长行的上提从逐帧限速改为缓动追赶，新行从静止线下方滑入，不再整行硬跳。

  • 模型备份自动发现端侧模型目录：除语音 / 向量清单外，自动发现 noBackupFilesDir/<kind>/<id>/<revision>/ 三级布局下的完整模型目录（如动作模型），导出时现场计算哈希；半下载目录跳过，符号链接不打包；恢复只刷新真正被写入的类型。

  • 宽屏界面：侧栏改标准 side sheet 固定宽度并排，不再提供覆盖抽屉形态；新建 / 选择会话不再无条件收起侧栏；二级页内容列居中限宽，汉堡按钮改独立 toggle 回调。

  • 图片组改横向轮播：图片行改用整条聊天区宽度等比适配，每张一样大、整组两端完整；缩到 64dp 下限仍放不下才横向滚动，与文件行解耦。

  • 安全收紧：网络安全配置默认禁明文，仅放行本地回环（localhost / 127.0.0.1 / ::1），保住 OAuth loopback 回调与本地 Linux 服务——行为变更：显式配置的远程 http:// 模型端点将不可用，需改用 https；OAuth 登录页禁用混合内容。

  • i18n 硬编码清理：枚举与 provider 层只保留协议值和 @StringRes 引用，展示文案一律在 UI 边界按当前语言解析；媒体源推荐 label、Root、Model ID、压缩摘要标签、会话导出系统状态等硬编码全部改为字符串资源，新增 27 个 key，简 / 繁 / 英三语言同步落地。

  • 浏览器设置子页转场：设置页与浏览器共用同一个标签池不能拆路由，改为页内覆盖层——设置页侧滑进入、底层浏览器四分之一屏视差，时长与 NavHost 一致，支持 RTL 镜像。

Improvements

  • feat: Markdown images in AI replies now resolve across all sources. /workspace, /var/minis, minis:// and file:// paths are mapped to the host via LinuxGuestPathResolver (same mapping as read_image), so images produced inside Linux tool sessions render without copying; schemeless relative paths resolve against the workspace root; inline rendering and gallery preview share one source whitelist; the system prompt gained explicit image syntax guidance.

  • feat: On-device action model (FunctionGemma 270M Mobile Actions) — dual-source download (ModelScope / Hugging Face) with checksum verification and cancellation, a device_actions tool turning English instructions into six real device actions (flashlight, contacts, email, map, Wi-Fi, calendar), settings management with Auto/GPU/CPU backend selection, and a standalone try-out page.

  • feat: Thinking and tool groups of a turn auto-collapse into a "N steps done" card as soon as the reply body appears (manually expanded groups stay open); folding upgraded from a boolean to three levels (aggressive / normal / off), with aggressive mode merging intermediate replies and traces of the turn into one work card while keeping Q&A cards, suggestion chips and streaming replies visible.

  • feat: Media search gains multi-select with batch move/copy (directory picker, persisted SAF grant, per-item error summary), a shared similarity-threshold preference (set_threshold tool + slider), and a batch_act tool accepting explicit source_keys or an inline query.

  • feat: Voice settings restructured — Doubao/Mimo voice catalogs moved into independent sub-pages with real NavHost transitions; form styling and tap haptics unified.

  • feat: Sub-agent presets moved to a dedicated detail page; the profile config dialog splits tier / reasoning effort / resolution / gpt-speed into sub-pages.

  • refactor: Composer tools moved into a two-level picker; the bottom-follow tail lift switched from per-frame rate limiting to eased pursuit, so new rows slide in from the rest line instead of jumping a full row.

  • feat: Model backup auto-discovers complete model directories under noBackupFilesDir/<kind>/<id>/<revision>/ (e.g. the action model) and hashes them at export time; half-downloaded directories and symlinks are skipped; restore refreshes only the kinds actually written.

  • feat: Wide-screen layout uses a standard fixed-width side sheet (no overlay drawer), keeps the sidebar open when creating/selecting conversations, and centers secondary pages in a capped content column.

  • fix: Image rows now adapt to the full chat width with uniform sizes, falling back to horizontal scrolling only below the 64dp floor, decoupled from file rows.

  • fix(security): Cleartext traffic is now denied by default with only local loopback exempted (preserving OAuth callbacks and local Linux services) — breaking change: explicitly configured remote http:// model endpoints stop working, use https; the OAuth login page blocks mixed content.

  • refactor(i18n): Enums and providers now carry only protocol values plus @StringRes references; all display text resolves at the UI boundary in the current language. 27 new keys land in Simplified/Traditional Chinese and English together.

  • fix(browser): Browser settings became an in-page overlay (side-slide with quarter-screen parallax, RTL-mirrored) since the settings page and browser share one tab pool and cannot be separate routes.

Assets 3
Loading