Android now shares the same version line as iOS — this release jumps from 0.22-preview to 1.12, and drops the -preview suffix. It is still a prerelease build for testing.
本次起 Android 与 iOS 统一版本号：从 0.22-preview 直接对齐到 1.12，并去掉 -preview 后缀。仍属测试版本。
New features / 新功能
Voice & read-aloud / 语音与朗读
- A floating speech player: mute, switch TTS model, and change speed without leaving the chat.
悬浮语音播放器：静音、切换 TTS 模型、调整语速，无需离开对话。 - Read-aloud now announces tool calls as they run, matching iOS.
朗读现在会播报正在执行的工具调用，与 iOS 一致。 - Read-aloud pauses automatically while the microphone is capturing and resumes when it closes, so playback is no longer recorded into your dictation.
开始录音时自动暂停朗读、停止录音后恢复，避免播放内容被一并录入。 - Per-voice system TTS catalog in the picker, so you can choose a specific system voice rather than just the engine default.
选择器中提供系统 TTS 逐条音色列表，可指定具体音色而非仅用引擎默认。 - Speech recognition now uses Silero VAD to segment speech, and accumulates short segments — stop-start dictation works instead of dropping fragments.
语音识别改用 Silero VAD 做断句，并累积短语音片段——断续口述不再丢字。
Chats & organization / 会话与整理
- Session grouping: create groups, file chats into them, change or remove a chat's group, and delete a group together with its chats.
会话分组：创建分组、归档会话、变更或移出分组，并可连同分组内会话一起删除。 - Auto-grouping files a new chat into a matching group when its title is generated (on by default, both platforms).
自动分组：生成标题时将新会话归入匹配的分组（双端默认开启）。 - "New Chat in Group" starts a chat already filed in the group you are viewing.
「在分组中新建会话」直接在当前分组下开始新对话。 - The home list adopts the iOS folder-card look, with accordion state remembered and a quick-navigation mini-bar.
首页列表采用 iOS 文件夹卡片视觉，记忆展开状态，并提供快速导航条。
Models & providers / 模型与供应商
- Vision Group: non-vision models can now read images by delegating to a vision model, and the reply names which model read the image.
视觉分组：非视觉模型可委托视觉模型读图，回复中会注明由哪个模型完成识别。 - Custom thinking rules — create, scope, and edit reasoning settings per model, including via minis-config.
自定义思考规则：按模型创建、限定范围与编辑推理设置，也可通过 minis-config 配置。 - Quick-test any model straight from its row in the picker, with an "Active" badge showing which group member is actually in effect.
在选择器中直接快速测试模型，并以「Active」标记当前实际生效的分组成员。 - Drag to reorder providers and model groups; hide or show a model directly from the provider's model list.
拖拽排序供应商与模型分组；可直接在供应商模型列表中隐藏/显示模型。 - Third-party OpenAI/Anthropic-compatible endpoints can now be used without an API key.
第三方 OpenAI/Anthropic 兼容端点现在允许不填 API Key。 - Image-to-image editing support (
/images/edits).
支持图生图编辑（/images/edits）。
Context management / 上下文管理
- Auto-compaction: a new "Compact & Enable Auto-Compact" action plus an in-chat toggle, so long conversations compact themselves.
自动压缩：新增「压缩并开启自动压缩」操作与对话内开关，长会话可自动压缩。
Text selection / 文本选择
- Long-pressing inside a table now selects the cell rather than the whole row, with haptic feedback; extra actions fold into an overflow menu.
长按表格内容改为选中单元格而非整行，并加入震动反馈；多余操作收进溢出菜单。
Fixes / 修复
Stability / 稳定性
- Fixed a crash and a silent provider wipe in the provider config store — a failed database read could discard your configured providers.
修复供应商配置的崩溃与静默清空问题——数据库读取失败可能导致已配置的供应商被丢弃。 - Fixed a native crash (SIGSEGV) in voice activity detection when the recognizer was released mid-processing.
修复语音活动检测在处理中被释放导致的原生崩溃（SIGSEGV）。 - A bad skill import can no longer permanently break app startup.
错误的技能导入不再会永久性地导致应用无法启动。 - The sandbox self-heals early startup crashes by retrying without seccomp, and restores the vendored loaders that some ROMs rejected.
沙盒在早期启动崩溃时会自动以无 seccomp 方式重试，并恢复部分 ROM 无法运行的加载器。
Network / 网络
- Raised the streaming first-response timeout from 30s to 120s — complex agent turns, local models, and slow relays were being cancelled as false timeouts.
流式首包超时从 30 秒提高到 120 秒——复杂 Agent 任务、本地模型与慢速中转此前会被误判超时取消。
Voice & read-aloud / 语音与朗读
- Fixed the read-aloud silent-failure chain: playback no longer stops early, dies silently, or loses audio focus.
修复朗读的静默失败链路：不再提前中断、无声失败或丢失音频焦点。 - MiniMax TTS and OpenRouter TTS/ASR now use the correct endpoints and payloads.
MiniMax TTS 与 OpenRouter TTS/ASR 现在使用正确的端点与请求格式。
Chat / 对话
- "Minis is thinking" stays visible when the only streamed content is hidden reasoning.
当流式内容仅为隐藏的推理时，「Minis 正在思考」不再消失。 - Fixed several scroll-follow problems: auto-follow no longer breaks mid-stream, and the scroll-to-bottom button appears reliably.
修复多处滚动跟随问题：流式过程中不再中断自动跟随，回到底部按钮显示更可靠。 - Compaction no longer cuts a tool call and its result in half.
上下文压缩不再把工具调用与其结果截断分开。 - Sessions left untitled by an interrupted title request now recover.
因标题请求被中断而未命名的会话现在可以恢复命名。
Memory & skills / 记忆与技能
- Restored the memory-file delete flow, which was unreachable — memory files can be deleted again.
恢复此前无法触达的记忆文件删除入口——现在可以删除记忆文件。 - Skill entries whose files are gone from disk are pruned instead of lingering.
磁盘上已不存在的技能条目会被清理，不再残留。
Live Updates / 灵动岛通知
- A completed task no longer shows a Stop button, and the progress bar no longer renders as a stalled position tracker.
任务完成后不再显示「停止」按钮，进度条也不再显示为停滞的位置指示器。
Settings / 设置
- The floating-overlay toggle now reflects what you actually set: if the permission is missing it stays on and tells you a permission is needed, instead of silently appearing off.
悬浮窗开关现在如实反映你的设置：缺少权限时保持开启并提示需要授权，而不是静默显示为关闭。 - Wired up the manual mirror-selection screen, which could not be reached.
接通此前无法进入的手动镜像选择页面。 - Fixed markdown formula rendering: inline KaTeX is centered on the text line, scales with system font size, and is no longer clipped.
修复 Markdown 公式渲染：行内 KaTeX 与文本基线居中对齐、随系统字号缩放，且不再被裁切。
Polish / 视觉打磨
- The accent color moves to iOS's blue, desaturated so it doesn't read as glaring on Android.
强调色改为 iOS 蓝，并做降饱和处理以避免在 Android 上过于刺眼。 - Model picker refinements: clearer section headers, neutral expand chevrons, wrapped modality chips, and consistent row anatomy across all pickers.
模型选择器优化：更清晰的分区标题、中性展开箭头、模态标签整块换行，以及各选择器一致的行结构。 - Long-press menus now anchor at your finger, use a softer popup shadow, and vibrate consistently.
长按菜单在手指位置弹出，采用更柔和的阴影，并统一加入震动反馈。 - The rename-group dialog's inputs get a border and padding so they no longer run together.
重命名分组对话框的输入框加上边框与内边距，不再挤在一起。 - Search and input fields share a consistent 42dp height.
搜索与输入框统一为 42dp 高度。
Full changelog: 0.22-preview...1.12