What's Changed
新增
- 为插件市场新增分类筛选并优化界面组件。(#6762)
- 新增 Kimi Coding Plan Provider(兼容 Anthropic API)。(#6559)
- 统一并重命名
skill.md文件为SKILL.md。(#6757) - 新增 Ollama 关闭思考模式开关。(#5941)
- 新增小米 MiMo TTS 与 STT Provider。(#6643)
- 新增多模态上下文 token 计数能力(图片、音频与思维链)。(#6596)
- 新增按元信息名称安装插件并校验可导入标识符。(#6530)
优化
- 降低
kook_type中StrEnum到(str, Enum)以兼容旧版本 Python。(#6512) - 优化 WebChat 消息处理:提前保存消息并增强流式补全。(#6698)
- 优化 Dashboard 中移动端 persona 选择器布局。(#5907)
- 优化 Dashboard,减少 MDI icon font 范围并自托管 Google Fonts。(#6532)
- 使用 Pydantic 重构 KOOK 适配器数据模型以提升消息结构解析与校验性能。(#5719)
- 调整文档术语与表达(“消息平台”改为“机器人”等)并优化更新说明文字。(#6568)
修复
- 修复 web search tools 中被禁用系统函数时未正确生效的问题。(#6584)
- 跳过空的知识库嵌入批次,避免不必要计算。(#6106)
- 修复 Feishu opus 文件在 Whisper API STT 场景下未正确转换的问题。(#6078)
- 修复 Lark 在工具调用中断时卡片创建与重续流程。(#6743)
- 修复 Windows 环境下 skill 文件编码兼容问题。(#6058)
- 修复 WebChat 单独渲染的 HTML 回复被错误显示的问题。(#6074)
- 修复 Gemini 数组 schema 缺少
items字段的问题。(#6051) - 修复 skills 使用硬编码工作区路径的问题,改为使用缓存中的真实沙箱路径。(#6331)
- 修复 subagent 工具调用超时传递,修正
tool_call_timeout并将默认超时从 60 秒提升到 120 秒。(#6713) - 修复
skills-like下缺失extra_user_content_parts时未注入图片说明的问题。(#6710) - 修复
/stop后继续处理后续 follow-up 消息的问题。(#6704) - 修复核心流程在 stop 时中断子代理工具等待的行为。(#5850)
- 修复 Windows 平台 stdio MCP 服务
PATHEXT丢失的问题。(#5822) - 修复 Telegram 轮询重启机制,失败时自动恢复循环。(#6648)
- 修复
/stop后 follow-up 状态未清理导致后续触发继续运行的问题。(#6656) - 修复 OpenAI 配置日志,改进代理与 API Base 的可读性与排障能力。(#6669)
- 修复长工具调用中截断逻辑误删唯一 user 消息的问题。(#6198)
- 修复截断器移除唯一用户消息导致部分 provider 返回
400的问题。(#6581) - 修复 ID 白名单为空时提示文案不清晰的问题。(#6611)
- 修复 MoonshotAI 官方 API 下 OpenAI token usage 缺失问题。(#6618)
- 提升 Wecom AI Bot 流式可读性与速度(限流处理),改善卡顿与显示体验。(#6610)
- 修复 Wecom 长连接回复消失的问题。(#6606)
- 修复 Windows 打包运行时 pip/native 构建环境配置问题。(#6575)
- 修复仅在显式配置时传递 embedding
dimensions,避免无效参数干扰。(#6432) - 修复
register_agent装饰器导致的NameError。(#5765) - 修复默认 persona 场景下的 subagent 查询失败问题。(#5672)
- 修复 Windows 下 skill 读取失败问题。(#6490)
- 修复
SKILL.md多行 frontmatter 解析问题。(#6460) - 修复 SQLite
database is locked问题,新增繁忙超时处理。(#6474) - 修复工作流限制仅允许上游仓库触发的报错场景。(#6531)
What's Changed (EN)
New Features
- Added category filtering for market plugins and enhanced UI components.
- Normalized and renamed
skill.mdfiles toSKILL.md. - Added a toggle to disable thinking mode in Ollama.
- Added Xiaomi MiMo TTS & STT providers.
- Added Kimi Coding Plan provider with Anthropic API compatibility.
- Added context token counting support for multimodal content (images, audio, and chain-of-thought).
- Added plugin installation by metadata name and validated importable identifiers.
Improvements
- Refactored
StrEnumto(str, Enum)inkook_typefor backward compatibility. - Enhanced WebChat message handling with proactive saving and streaming completion.
- Simplified dashboard persona selector layout for mobile screens.
- Subset MDI icon font and self-hosted Google Fonts in dashboard.
- Implemented Pydantic data models for KOOK adapter for better retrieval and schema validation.
- Updated documentation wording for incomplete updates terminology.
Bug Fixes
- Fixed disabled system functions in web search tools.
- Skipped empty knowledge-base embedding batches.
- Converted Feishu opus files for Whisper API STT.
- Deferred Lark card creation and renewal on tool call break.
- Added fallback for Windows skill file encodings.
- Rendered standalone HTML replies as code in WebChat.
- Ensured Gemini array schemas always include
items. - Fixed skills path resolution to use cached sandbox path instead of hardcoded workspace root.
- Passed
tool_call_timeoutto subagent handsoff/cron/background tasks and increased default timeout to 120 seconds. - Fixed
skills-likere-query missingextra_user_content_partscausing image caption not injected. - Rejected follow-up messages after
/stoprequest. - Interrupted subagent tool waits on stop.
- Preserved
PATHEXTfor stdio MCP servers on Windows. - Added automatic restart for Telegram polling loop on failure.
- Fixed follow-up persistence logic after
/stoptrigger. - Improved OpenAI proxy/API base logging.
- Prevented truncation logic from removing the only user message in long tool-calling conversations.
- Fixed truncation removing the only user message and returning
400for some providers. - Updated ID whitelist empty-configuration hint wording.
- Fixed token usage with MoonshotAI official API.
- Improved Wecom AI Bot streaming readability and speed via throttling.
- Prevented Wecom AI Bot long connection replies from disappearing.
- Set packaged Windows runtime build env for pip native builds.
- Only pass embedding dimensions when explicitly configured.
- Fixed
register_agentdecoratorNameError. - Fixed subagent lookup failure when using default persona.
- Fixed reading skills on Windows.
- Fixed multiline frontmatter parsing in
SKILL.md. - Fixed SQLite
database is lockedwith busy timeout. - Restricted workflows to upstream repo.