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。
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.