V4.25.3

• 更新日志(简体中文)
• Changelog(English)

<a id="chinese"></a>

What's Changed

重点更新

• 修复钉钉 Stream 任务异常退出后无法持续重连的问题，提升钉钉适配器长连接稳定性。(#8610)
• Dashboard 登录新增 TOTP 双因素认证。(#8189)
• Dashboard 侧边栏新增插件 WebUI 入口，并支持 MDI 图标；插件页面现在会同步 Dashboard 主题。(#8569, #8390)
• 新增未来任务 UI，并新增 /name 命令用于命名 umo 且在 UI 中展示。(#8559, #8575)
• MiniMax Token Plan 支持动态获取模型列表，默认模型升级为 M3。(#8475, #8505)
• 优化长期记忆与上下文压缩，提高压缩质量和 KV-cache 命中率。(#8530, #8226)

修复

• 修复 Provider 统计遇到 SQLite lock 时未重试的问题。(commit)
• 修复空 modalities 被当作已配置能力，导致核心图片请求错误路由到视觉 fallback 的问题。(#8451)
• 修复 Reply 组件 toDict 额外字段导致私聊消息发送失败的问题。(#8477)
• 修复带 Reply 组件唤醒机器人且用户提示为空时的处理问题。(#8461)
• 修复 aiocqhttp 适配器无法解析语音的问题，并增加音频文件存在性检查。(#8523, #8529)
• 修复 OpenAI Embedding 在 SiliconFlow 非 Qwen embedding 模型上的无效参数问题。(#8508)
• 修复 Gemini Embedding 批量文本在 gemini-embedding-2 上被折叠的问题。(#8537)
• 修复 WeCom AI 消息链提取纯文本时的空白处理问题。(#8563)
• 修复 Dashboard 在 launcher 下运行时 frame 安全头过于严格的问题。(#8554)
• 修复控制台 pip 安装错误展示，改为使用 toast 提示。(#8462)
• 优化模板配置相关问题。(#8228)

优化与工程

• 增强 Reply 消息链对 Record 组件的处理。(#8527)
• Whisper STT 调用 OpenAI 转写接口后会关闭音频文件句柄。(#8528)
• 防止唤醒命令被记录为群聊上下文。(#8536)
• 修正中文文本中的 token 术语。(#8465)
• 更新 GitHub Actions 依赖。(#8533)

<a id="english"></a>

What's Changed (EN)

Highlights

• Fixed DingTalk Stream task recovery so the adapter keeps reconnecting after unexpected SDK task exits. (#8610)
• Added TOTP two-factor authentication for Dashboard login. (#8189)
• Added plugin WebUI entries to the Dashboard sidebar with MDI icon support, and synchronized Dashboard theme styling into plugin pages. (#8569, #8390)
• Added the future task UI and introduced /name for naming umo entries shown in the UI. (#8559, #8575)
• MiniMax Token Plan can now fetch model lists dynamically, with the default model upgraded to M3. (#8475, #8505)
• Improved long-term memory and context compression quality, including better KV-cache hit rates. (#8530, #8226)

Bug Fixes

• Retried provider stats reads when SQLite is locked. (commit)
• Treated empty modalities as unconfigured so image requests do not incorrectly route to vision fallback. (#8451)
• Fixed private message sending failures caused by extra fields in Reply component toDict output. (#8477)
• Fixed waking the bot with a Reply component and an empty user prompt. (#8461)
• Fixed voice parsing in the aiocqhttp adapter and added an audio file existence check. (#8523, #8529)
• Fixed invalid OpenAI Embedding parameters for SiliconFlow non-Qwen embedding models. (#8508)
• Wrapped Gemini batch embedding texts in Content to avoid collapse on gemini-embedding-2. (#8537)
• Fixed whitespace handling when extracting plain text from WeCom AI message chains. (#8563)
• Relaxed Dashboard frame security headers when running under the launcher. (#8554)
• Displayed console pip-install errors with toast messages. (#8462)
• Optimized template config handling. (#8228)

Improvements and Maintenance

• Enhanced Reply chain handling for Record components. (#8527)
• Closed audio file handles after OpenAI transcription calls in Whisper STT. (#8528)
• Prevented wake commands from being recorded as group chat context. (#8536)
• Fixed Chinese token terminology. (#8465)
• Updated GitHub Actions dependencies. (#8533)