MewuAI 0.3.0 · 喵呜AI 0.3.0

本次更新汇总 v0.2.6 之后的多渠道 AI 接入、模型选择重构和对话修复。应用、安装器、源码标签与下载包统一为 0.3.0。

新增 AI 接入

• ChatGPT Work / Codex： 通过官方本机 codex app-server 接入，沿用官方客户端登录；设置页读取真实模型与思考程度，支持文字、图片及交给 Agent 自行处理的视频。不提供 Chat 页，用量按官方账号规则计算。
• WorkBuddy： 通过桌面客户端自带 CLI 的官方 ACP 接口接入，复用本机登录并读取真实模型和思考选项。视频由 Agent 使用本机已有工具分析。
• MiniMax Code 桌面版： 增加独立配置页，支持发现桌面登录、打开官方客户端、刷新状态和测试连接，无需另外安装或登录 CLI。当前通过桌面会话兼容接入，和 Codex / WorkBuddy 的 Agent 工具接口不同；不宣称为官方公开第三方 SDK。
• API、Hermes、Codex、WorkBuddy、MiniMax Code 使用统一的顶部配置标签与表单布局。移除尚未适配的 OpenClaw、Claude Code 占位页；本版没有豆包工作桌面接入。

多渠道与模型切换

• API 接入改为列表，可保存多个接入点、模型和独立凭据；切换配置页保留未保存草稿，保存后保留各渠道配置。
• 取消用设置页单选项控制所有渠道的旧逻辑；已配置且满足当前使用条件的渠道都进入对话选择列表，兼容旧 WorkBuddy 开关关闭但已有配置的情况。
• 模型入口位于对话条上传按钮之后，使用同款圆形图标。点击直接展开一层列表，点选即切换，当前项显示勾号；支持滚动和键盘导航，修复弹层边框、阴影裁切。
• API 列表只列实际配置的接入点与模型，避免把推荐模型或账户目录当成已配置模型。
• 默认使用上一次选择的渠道及其模型，重启后保留；主页状态随实际渠道显示，修复已选 Codex 却仍显示 MiniMax 的问题。不同渠道和模型的历史分开保存。

对话与连接修复

• 修复没有截图、录屏或上传附件时仍附带识图和视觉批注提示词的问题；文字请求不自动创建隐式全屏截图。
• 修复 Agent 回答直接显示 annotationProtocol 等原始 JSON；统一提取正文并校验可执行批注。
• WorkBuddy「测试连接」改为协议握手、会话建立和模型/思考选项检查，不发送推理对话；等待有独立的 30 秒上限，超时明确结束。该检查通过不等于所有模型推理能力已验证。
• 修复 WorkBuddy 思考迟迟不结束及动态模型/思考选项被误判不可用；完成、取消、失败按当前请求处理，不把工具活动或仅有思考的内容当作最终答案。
• API 与 MiniMax Code 的连接结果改为设置页内的绿色成功文字，统一各页状态呈现。两者手动测试仍发送简短验证消息，可能使用服务额度。

界面、许可与发布

• 主页移除最小化按钮，隐藏到托盘入口改用关闭图标；全局截图快捷键支持按 Delete 清空并停用。
• 项目自有源码正式采用 MPL-2.0，发布包携带 LICENSE 与 SOURCE.md；具体许可范围和第三方条款见仓库对应文件。
• 补齐社区行为准则、贡献指南、安全策略、Issue / PR 模板；README 更新全部当前渠道和使用方法，并新增完整版本索引。
• README 演示保留 GIF，移除仓库源演示 MP4。新 Release 只附安装 EXE 和便携 ZIP，校验使用 GitHub 资产原生 SHA-256，不附演示视频或 SHA256SUMS.txt。

升级与兼容范围

• v0.2.5 及更早客户端依赖旧校验文件，自动更新可能失败；请从本仓库 Release 手动下载安装器升级。v0.2.6 起支持 GitHub 原生 digest。
• Codex / WorkBuddy 已有本机文字、合成图片和合成视频验证记录；这不代表所有账户、模型或其他电脑均通过。Agent 视频能力取决于本机已有工具，喵呜AI不捆绑或自动安装视频解码工具。
• MiniMax Code 桌面会话、客户端版本和所选模型会影响可用性；未登录、会话过期或额度不足时需在官方客户端处理。

全部提交变化 · 完整版本记录

---

This release brings together the AI integrations, channel picker redesign, and conversation fixes developed after v0.2.6. The application, installer, source tag, and packages use version 0.3.0.

AI integrations

• ChatGPT Work / Codex: connect through the official local codex app-server, reuse the official login, and read real models and reasoning options. Text, images, and videos handed to the Agent are supported. There is no Chat tab; usage follows the account's official rules.
• WorkBuddy: connect through its bundled CLI's official ACP interface, reuse the local login, and read real model and reasoning options. The Agent inspects videos using tools already installed on the PC.
• MiniMax Code desktop: add a settings page to discover the desktop session, open the official client, refresh status, and test a connection without a separate CLI installation or login. This desktop-session compatibility path is different from the Codex / WorkBuddy Agent interfaces and is not presented as an official public third-party SDK.
• Unify the API, Hermes, Codex, WorkBuddy, and MiniMax Code settings layouts. Remove the unimplemented OpenClaw and Claude Code placeholders. Doubao Work desktop is not included in this release.

Channels and model selection

• Save multiple API endpoints with separate models and credentials. Preserve configuration drafts when navigating settings and retain each channel's saved configuration.
• Make configured, usable channels available together instead of enabling only the selected settings tab. Recognize saved WorkBuddy configurations even when a legacy enable flag is off.
• Place a round model button immediately after Upload. Open a single scrollable list, switch with one click, mark the current item, support keyboard navigation, and fix clipped popup borders and shadows.
• Show only configured API endpoints/models, rather than treating suggested or account-listed models as configured entries.
• Remember the last channel/model across restarts, display it on the home page, and keep history scoped to its channel/model. Fix the home page showing MiniMax after selecting Codex.

Conversation and connection fixes

• Keep text-only messages free of automatic screen attachments and visual-analysis instructions.
• Parse Agent responses into readable answers instead of displaying raw annotation-protocol JSON; validate executable annotations separately.
• Check WorkBuddy through ACP initialization, session creation, and model/reasoning configuration with a separate 30-second timeout. The check sends no inference prompt and does not certify every model's inference capability.
• Fix lingering WorkBuddy thinking and incorrect rejection of dynamic model/reasoning options. Distinguish completed answers from tool activity, reasoning-only output, cancellation, and failure.
• Show consistent inline connection results, including green success text for API and MiniMax Code. Their manual tests still send a short verification message and may use service quota.

Interface, licensing, and distribution

• Replace the home-page tray-hiding icon with Close, remove Minimize, and allow Delete to clear and disable the capture shortcut.
• Adopt MPL-2.0 for project-owned source and include LICENSE / SOURCE.md in packages. See the repository for scope and separate third-party terms.
• Add community, contribution, security, issue and PR guidance; update both READMEs and add a complete release-history index.
• Keep the README GIF and remove the source demo MP4. Releases attach only the installer and portable ZIP, using GitHub's native asset SHA-256 without a separate checksum file or demo video.

Upgrade and compatibility

• v0.2.5 and earlier updaters require the old checksum file and may fail to update automatically. Download the installer manually from this repository's Releases. Native GitHub digests are supported starting with v0.2.6.
• Existing local Codex / WorkBuddy text, synthetic-image, and synthetic-video checks do not verify every account, model, or PC. Agent video analysis depends on locally installed tools; MewuAI does not bundle or install decoders.
• MiniMax Code availability depends on its desktop session, client version, and selected model; resolve expired login or quota issues in the official client.

Full comparison · Release history

MewuAI 0.2.6 · 喵呜AI 0.2.6

本次更新改善标注编辑与录屏导出，修复 Hermes 安装后启动、历史复制和高 DPI 菜单问题。

• 取色提示可见时按 C 复制当前像素的 RGB 色值，文字输入和文本复制保持原有操作。
• 移除独立移动工具。使用画笔、高亮或形状工具时，可直接点击已有对象并拖动；矩形、椭圆等支持四角缩放，箭头支持分别调整起点和终点，文字双击重新编辑。调整支持撤销、重做，编辑控制点不会进入导出图片。
• 区域录屏默认录制电脑声音，麦克风可单独开启；默认导出 MP4，也可选择 MP3 音频或 GIF 动图。GIF 格式本身不包含声音。
• 修复安装或更新后 Hermes 后台可能报“代码 1”的启动环境继承问题；修复更新关闭设置窗口时可能出现的窗口状态异常。
• 历史记录支持拖选、Ctrl+C 和右键复制完整内容；修复鼠标仍在对话条内时，因下方截图工具条重叠而误收纳的问题。
• 复用历史控件与取色缓存，限制后台框选探测并发，减少重复布局和鼠标移动时的多余工作。
• 修复高分辨率和高 DPI 下托盘右键菜单文字偏上；主页未配置可用 AI 时显示红色状态点，设置入口直接进入 AI 标签。
• 更新器优先使用 GitHub 安装包资产自带的 SHA-256。此迁移版本仍提供旧客户端自动升级所需的校验兼容文件；演示仅保留 README 内嵌 GIF，不再上传演示 MP4。

---

This update improves annotation editing and recording exports, and fixes Hermes startup after installation, history copying, and high-DPI tray menus.

• Press C while the color inspector is visible to copy the current pixel as RGB. Text entry and text copying retain their existing behavior.
• Remove the separate move tool. Click and drag existing annotations while using the pen, highlighter, or shape tools. Resize rectangles, ellipses, and other objects using corner handles, adjust either arrow endpoint, and double-click text to edit it. Adjustments support undo and redo; editing handles stay out of exported images.
• Region recordings capture computer audio by default, with an optional microphone. MP4 is the default export; MP3 audio and GIF animation are also available. GIF does not carry audio.
• Fix an inherited process environment that could make Hermes fail with code 1 after installation or an update. Fix a window-state exception when updates close Settings.
• Select and copy history text with Ctrl+C, or copy the full content from its context menu. Fix the conversation bar hiding while the pointer remains inside it above an overlapping capture toolbar.
• Reuse history controls and color-sampling caches, and bound background snapping probes to reduce repeated layout and unnecessary pointer-movement work.
• Correct vertically misaligned tray-menu text at high DPI. Show a red home-page status indicator when AI is unavailable, and open the AI tab directly from its settings entry.
• Prefer GitHub's native SHA-256 digest for installer verification. This migration release retains the checksum compatibility file needed by older clients for automatic upgrades. Demos remain embedded GIFs in the README; no demo MP4 is attached.

---

维护说明（2026-09-07）：以上为 v0.2.6 原始功能说明。此前误加的后续变更已移到 0.3.0 更新说明。v0.2.6 的标签及安装包未更改。SHA256SUMS.txt 附件于 2026-09-07 移除；v0.2.5 及更早版本的自动升级可能因此失败，请手动下载安装器。完整历史

Maintenance note (2026-09-07): the original v0.2.6 feature notes above are restored. Later changes are recorded in the 0.3.0 notes. The v0.2.6 tag and binaries are unchanged. The checksum attachment was removed on 2026-09-07; users of v0.2.5 or earlier may need to download the installer manually. Full history

MewuAI 0.2.5 · 喵呜AI 0.2.5

本次更新改善长回答阅读与操作响应，加入教学共享模式，并让设置页自动检查更新。

• 设置 → 捕获新增“教学演示模式”。开启后，下次截图的选区、批注和回答可被屏幕共享捕获，相关贴图与原位对话框同步可见；设置与凭据窗口继续防捕获。演示模式下停用自带区域录屏和长截图，避免将覆盖层录入结果。腾讯会议与极域的远端显示仍需在实际教学环境中验证。
• 合并流式增量，复用稳定的 Markdown 段落与截图裁剪，减少长回答逐字扫描、重复排版及鼠标移动时的无效布局；改善历史加载与已删除区域的媒体资源回收。
• 历史展开/收起恢复纯箭头；回到最新回复使用滚动条底部的小箭头，修正与滑块的对齐，不改变左侧正文布局。上翻阅读时保留当前位置，点击箭头后恢复跟随。
• 每次打开设置页自动检查官方 GitHub 更新。发现新版先询问是否下载，选择“稍后”不下载；下载及 SHA-256 校验完成后再确认安装并重启。关闭设置会取消检查和下载。

0.2.4 的正式发布被录屏转 GIF 集成测试拦截，未提供下载包。本版本保留所有功能更新，并让录屏测试等待原生录制就绪后计时，同时校验成片时长。

性能改善基于本机合成回放；不代表所有设备或场景中的卡顿均已消除。公测模型仍为 MiniMax M3，其他模型请自行验证兼容性。

---

This update improves long-answer reading and responsiveness, adds a teaching screen-sharing mode, and checks for updates when Settings opens.

• Enable Teaching mode under Settings → Capture to make the next capture overlay, annotations, answers, related pins, and in-place dialogs visible to screen sharing. Settings and credential windows remain protected. Built-in region recording and scrolling captures are disabled in this mode to keep the overlay out of recorded results. Remote viewing through Tencent Meeting and Mythware still requires validation in the actual teaching environment.
• Batch streamed text, reuse stable Markdown blocks and image crops, and reduce per-character scans, repeated layout, and unnecessary work during pointer movement. Improve history loading and release media resources when deleted regions leave retained history.
• Restore arrow-only history controls. Place a small return-to-latest arrow below the answer scrollbar, aligned with its thumb without shifting the answer layout. Scrolling up preserves the reading position; clicking the arrow resumes following the latest response.
• Opening Settings automatically checks official GitHub releases. A new version prompts before downloading; Later leaves it undownloaded. Installation and restart require confirmation after the download passes SHA-256 verification. Closing Settings cancels checks and downloads.

The 0.2.4 release was stopped by a recording-to-GIF integration-test failure and has no download packages. This version retains all feature updates and starts the recording test duration after the native Recording event, with an additional recorded-duration assertion.

Performance improvements were measured with synthetic local replay and do not imply that every latency spike is eliminated on all devices. The beta test model remains MiniMax M3; please verify compatibility with other models.

MewuAI 0.2.3 · 喵呜AI 0.2.3

本次更新简化 AI 接入设置，并修复 Hermes 兼容性与贴图再次截图的问题。

• 提供商精简为 OpenAI 通用、MiniMax、MiniMax (CN)、火山引擎；只有 OpenAI 通用需要填写 URL，其余只需 Key 和模型，不再要求命名或管理连接。
• 根据所选提供商加载模型，支持手动填写模型 ID；已保存 Key 逐字符掩码显示，各提供商的凭据独立保留。
• 请求参数 JSON 与自定义请求头统一收进默认折叠的高级设置，说明通过小问号悬浮显示。支持 MiniMax M3 service_tier: priority，默认不启用；优先服务为标准价格的 1.5 倍。
• 连接测试结果采用统一浅色弹窗；再次截图置顶图片时保留边框和阴影，并限制超大贴图快照的内存占用。
• 改善 Windows Hermes 的启动器、虚拟环境和工具路径兼容性，隔离外部 Python 环境干扰，提供脱敏启动错误提示；不修改 Hermes 全局配置。

本地 Release 全量 736 项测试通过。公测模型：MiniMax M3。Hermes 已通过本机启动、人格与模型列表检查，其他电脑的兼容性仍需实际环境验证。

可在“设置 → 关于 → 检查更新”下载安装。

---

This update simplifies AI setup and fixes Hermes compatibility and recapturing pinned images.

• Choose from OpenAI compatible, MiniMax, MiniMax (CN), or Volcengine. Only OpenAI-compatible endpoints require a URL; the others need just an API key and a model. No connection names or management controls.
• Load models for the selected provider or enter an ID manually. Saved keys appear as masked characters, with credentials kept separate for each provider.
• Request parameters and custom headers now live in a collapsed Advanced settings section, with help on hover. MiniMax M3 supports service_tier: priority; it is off by default and costs 1.5× the standard rate.
• Connection-test results use the app's light dialog style. Recaptured pins retain their borders and shadows, with bounded memory use for large snapshots.
• Improve Windows Hermes launcher, virtual-environment, and tool-path compatibility; isolate conflicting Python environment variables and show privacy-safe startup errors without changing Hermes's global settings.

All 736 local Release tests passed. Beta test model: MiniMax M3. Local Hermes startup, profile-list, and model-list checks passed; compatibility on other PCs still needs validation in those environments.

Install through Settings → About → Check for updates.

MewuAI 0.2.2 · 喵呜AI 0.2.2

0.2.2 公测版集中完善原位标注、长截图、内容引用与国际化体验。 / The v0.2.2 public beta strengthens in-place annotation, scrolling capture, content references, and internationalization.

测试模型 / Tested model：MiniMax M3（MiniMax-M3）。

修复自包含安装包的运行时还原，以及慢机器上录屏启动期间的取消等待。0.2.0 / 0.2.1 未发布安装包，本次提供完整下载。 / Fixed runtime-pack restore and cancellation waits during recording startup on slower machines. The 0.2.0 and 0.2.1 builds did not produce releases; this version provides the complete downloads.

• 重构图片与视频标注管线：OCR 与无障碍控件校准目标，气泡全局避让，视频批注按真实呈现帧保持时间轴同步。 / Reworked image and video annotations with OCR/accessibility alignment, global callout placement, and presentation-frame timeline sync.
• 已有标注可继续追问、追加或明确重标；贴图、贴视频、复制与导出均可保留当前标注。 / Existing annotations now survive follow-ups and can be appended, replaced, pinned, copied, or exported.
• 长截图支持向上、向下连续拼接，修复滚轮焦点、图像比例及选区边框混入成品的问题。 / Scrolling capture now stitches in both directions, with fixes for wheel focus, aspect ratio, and selection borders leaking into the output.
• 手工标注支持选择、移动、重新编辑与单对象删除，工具栏按上下文收纳。 / Manual annotations can be selected, moved, re-edited, and deleted individually, with a contextual toolbar.
• 新增 AI 表格识别及一次复制 Excel、Markdown、CSV 与图片格式。 / Added AI table recognition with one-step Excel, Markdown, CSV, and image clipboard formats.
• 智能框选、按钮吸附、全屏对话条收纳和保存时机经过稳定性修复。 / Improved smart selection, control snapping, full-screen composer hiding, and capture-safe saving.
• 关于页支持 GitHub 检查更新、校验下载并静默安装；贴图层级、缩放与双击关闭同步修复。 / About now checks GitHub updates, verifies downloads, and installs silently; pinned-image layering, zoom, and double-click close were also fixed.
• 应用默认跟随 Windows UI 语言，也可在“常规”中固定简体中文或 English；安装程序自动适配系统语言。 / The app follows the Windows UI language by default and can be pinned to Simplified Chinese or English in General; the installer adapts automatically.

本版本支持 Windows 10 2004（build 19041）或更高版本，x64。安装版和免安装 ZIP 均为自包含发布包。 / Requires Windows 10 version 2004 (build 19041) or newer on x64. Both installer and portable ZIP are self-contained.

MewuAI 0.1.0 · 喵呜AI 0.1.0

0.1.0 公测版聚焦于稳定性、隐私与原位标注体验。 / The v0.1.0 public beta focuses on stability, privacy, and in-place assistance.

测试模型 / Tested model：MiniMax M3（MiniMax-M3）。

• 修复非左上角及多区域马赛克在实时显示、图片导出和视频导出中的尺寸与坐标错位。
• 结构化 AI 回复先严格解析再渲染，避免 JSON 协议泄漏到回答区；Hermes 与远程 Provider 保持一致。
• 恢复红色矩形/椭圆批注，移除重复的蓝色框；视频补标请求独立持有附件缓冲。
• OCR 与原位翻译始终使用干净冻结截图；上传附件使用准确 MIME 类型、独立编号和有限预览。
• 托盘菜单选中状态改为整行浅色圆角高亮；快捷键设置改为单一组合键捕获控件，默认 Shift+Alt+S。
• 多模态模型能力按实际模型 ID 判定；纯文本模型可进行文字对话，选择视觉附件时会明确提示能力限制。
• 无显式截图、录屏或上传附件时默认保持纯文字对话，不会自动发送桌面截图；Hermes 使用当前 Profile 保持上下文。
• 关于页面补充作者、仓库、版本和中英双语许可说明。
• 对话条新增默认折叠的“提问与历史 / Prompt & history”抽屉；回答返回前即可查看当前提问，展开时向上增长并保留最近对话。

本版本支持 Windows 10 2004（build 19041）或更高版本，x64。安装版和免安装 ZIP 均为自包含发布包。

MewuAI 0.0.11 · 喵呜AI 0.0.11

新增关于页面（作者、仓库和中英双语许可说明），并修复设置界面左侧菜单边框裁切。

Adds an About page with author, repository and bilingual license information, and fixes clipping in the settings navigation.

MewuAI 0.0.10 · 喵呜AI 0.0.10

本次更新优化附件上传按钮、流式回答布局、Hermes 状态收纳、AI 标注显示、沉浸式译文遮罩和滚动长截图，并修复异常结构化回答显示为 JSON 的问题。

This update improves attachment upload styling, streaming answer layout, Hermes status cleanup, AI annotation display, immersive translation overlays, and scrolling capture, while preventing malformed structured responses from leaking raw JSON.

MewuAI 0.0.9 · 喵呜AI 0.0.9

修复 GitHub Release 已部分创建时的资产补传流程，并包含此前版本的功能改进。

MewuAI 0.0.8 · 喵呜AI 0.0.8

修复 Windows 发布安装器版本号与项目版本不同步的问题，并包含流式思考、标注布局和附件引用改进。

Fixes the Windows installer version mismatch and includes streaming reasoning, annotation layout and attachment reference improvements.