<!-- qwen-release-notes:v1 -->
## 亮点
- 在 Channels 中添加了所有者作用域的命名会话，以便每个聊天最多管理八个持久化任务。 ([#10198](https://github.com/QwenLM/qwen-code/pull/10198))
- 启用守护进程扩展安装以接受绝对本地路径，同时拒绝相对路径和不受支持的选项。 ([#10131](https://github.com/QwenLM/qwen-code/pull/10131))
- 更新 Web Shell 侧边栏，为绑定到拉取请求的会话显示 GitHub 风格的彩色状态图标。 ([#10258](https://github.com/QwenLM/qwen-code/pull/10258))
- 在 Web Shell 中添加了可选的 composer 添加菜单，将附件、文件、扩展、MCP 服务器和技能分组。 ([#10379](https://github.com/QwenLM/qwen-code/pull/10379))
- 恢复助手消息中粗体文本的可见性，并防止侧边栏会话名称被截断。 ([#10157](https://github.com/QwenLM/qwen-code/pull/10157))
- 确保工具摘要始终可折叠，并将 MCP 应用与其他聚合内容分离。 ([#10231](https://github.com/QwenLM/qwen-code/pull/10231))
## 破坏性变更
暂无已知破坏性变更。
## 完整变更列表
### 功能
- 添加了私有 OpenTelemetry span 属性，用于报告详细的 LLM 上下文使用指标，包括输入类别和剩余容量。 ([#10016](https://github.com/QwenLM/qwen-code/pull/10016)) by @doudouOUC
- 启用守护进程扩展安装以接受绝对本地路径，同时拒绝相对路径和如 ref 之类的不受支持选项。 ([#10131](https://github.com/QwenLM/qwen-code/pull/10131)) by @callmeYe
- 为现有会话回填 PR 绑定，并添加合并状态快照以跟踪链接的拉取请求是打开、已合并还是已关闭。 ([#9729](https://github.com/QwenLM/qwen-code/pull/9729)) by @wenshao
- 为嵌入式 Web Shell 主机添加了可选设置，用于隐藏会话来源切换，将所有侧边栏目录固定到普通任务对话。 ([#10079](https://github.com/QwenLM/qwen-code/pull/10079)) by @dreamWB
- 通过不上传文件而直接隐藏格式错误或过大的内联标记，为保留的 [FILE: ...] 输出标记实现安全投影。 ([#9991](https://github.com/QwenLM/qwen-code/pull/9991)) by @qqqys
- 更新目标提示，在目标变更时仅通知模型一次，同时继续在每一轮中声明当前目标。 ([#10013](https://github.com/QwenLM/qwen-code/pull/10013)) by @qqqys
- 在 Channels 中支持所有者作用域的命名会话，允许每个聊天最多管理八个持久化命名任务。 ([#10198](https://github.com/QwenLM/qwen-code/pull/10198)) by @doudouOUC
- 更新 Web Shell 侧边栏，根据会话绑定拉取请求的打开、已合并或已关闭状态，显示 GitHub 风格的彩色状态图标。 ([#10258](https://github.com/QwenLM/qwen-code/pull/10258)) by @wenshao
- 通过在审查正文工件中添加稳定的不可见标记，使延迟的审查建议可被外部工具恢复。 ([#9761](https://github.com/QwenLM/qwen-code/pull/9761)) by @wenshao
- /review 命令现在通过检测后继链，在子系统的严重问题在轮次间反复重新增长时发出警告。 ([#10010](https://github.com/QwenLM/qwen-code/pull/10010)) by @wenshao
- 恢复了 Tauri shell 的桌面品牌构建技能，使用 brandId 和 logo 输入生成品牌化包。 ([#10164](https://github.com/QwenLM/qwen-code/pull/10164)) by @yiliang114
- 添加了独立的守护进程会话 API，支持对顶级会话进行创建、列出、恢复、归档和删除操作。 ([#10179](https://github.com/QwenLM/qwen-code/pull/10179)) by @doudouOUC
- /review 命令现在会记住每个项目最后一次显式使用的 --effort 级别，以便在后续无参数运行时使用。 ([#9883](https://github.com/QwenLM/qwen-code/pull/9883)) by @tlysanhuo
- 长时间运行的审查轮次现在每 10 分钟更新一次状态评论，包含经过时间和上次代理活动，以区分活跃工作与停滞进程。 ([#9771](https://github.com/QwenLM/qwen-code/pull/9771)) by @wenshao
- 本地审查-修复循环现在将增量轮次锚定到工作树内容更改而非提交，通过避免完整树的重审来减少 token 使用量。 ([#9659](https://github.com/QwenLM/qwen-code/pull/9659)) by @wenshao
- 审查发现现在支持可选的 fixConstraint 字段，用于指定在修复过程中建议的修复不得违反的现有事实或限制。 ([#10168](https://github.com/QwenLM/qwen-code/pull/10168)) by @wenshao
- feat(sdk): 添加独立会话 API ([#10294](https://github.com/QwenLM/qwen-code/pull/10294)) by @doudouOUC
- feat(cli): OpenTUI 迁移基础设施批次 — 锁定依赖、依赖方向架构检查、框架中立流式模型 ([#10134](https://github.com/QwenLM/qwen-code/pull/10134)) by @chiga0
- 添加了可配置的 Mem0 扩展骨架，提供仅检索的 external-context-mem0 stdio 扩展，并支持限制 HTTP 请求。 ([#10149](https://github.com/QwenLM/qwen-code/pull/10149)) by @doudouOUC
- 添加了新的分类门控，当 PR 的更改已被已合并的修复完全解决时，自动识别并关闭重复 PR。 ([#10292](https://github.com/QwenLM/qwen-code/pull/10292)) by @yiliang114
- 为 Web Shell 添加了可选的 composer 添加菜单，将附件、文件、扩展、MCP 服务器和技能分组，便于指向性选择。 ([#10379](https://github.com/QwenLM/qwen-code/pull/10379)) by @ytahdn
- 为 Web Shell 添加了可选的交互式浏览器终端，每个标签页管理独立的 PTY 会话，具有持久化滚动记录和状态处理功能。 ([#9984](https://github.com/QwenLM/qwen-code/pull/9984)) by @ytahdn
- 添加了 qwen review ab-drive 子命令，用于在共享或全新实例上针对 PR 和基础树运行执行级验证脚本。 ([#9740](https://github.com/QwenLM/qwen-code/pull/9740)) by @wenshao
- 已确认的严重问题（Critical issues）现包含方向轴和基线轴，以区分误认证、失败、回归和新暴露面。([#10301](https://github.com/QwenLM/qwen-code/pull/10301)) by @wenshao
- 将 Aone 打造为会话 PR 绑定的首选平台，支持通过 a1 CLI 实现分支到 MR 的自动映射及状态刷新。([#10287](https://github.com/QwenLM/qwen-code/pull/10287)) by @wenshao
### 缺陷修复
- 修复了 Web Shell 的问题，即使文件内容此前被截断，也能正确恢复已保存的会话差异并显示准确的行数。([#10093](https://github.com/QwenLM/qwen-code/pull/10093)) by @ytahdn
- 确保钉钉富文本消息在跨会话时保留所有图片作为有序附件，并优雅地处理上传失败的情况。([#9922](https://github.com/QwenLM/qwen-code/pull/9922)) by @BenGuanRan
- fix: 修复 Windows 测试通道的持续失败问题 ([#10064](https://github.com/QwenLM/qwen-code/pull/10064)) by @wenshao
- fix(core): 在 send_message 中报告广播投递失败 ([#10081](https://github.com/QwenLM/qwen-code/pull/10081)) by @yiliang114
- 将自动修复扫描通道移至持久化自托管 Runner 池，以防止 GitHub 托管 Runner 积压阻塞路由。([#10055](https://github.com/QwenLM/qwen-code/pull/10055)) by @wenshao
- 恢复助手消息中粗体文本的可见性，并防止侧边栏中的会话名称被注意力提示（attention pills）截断。([#10157](https://github.com/QwenLM/qwen-code/pull/10157)) by @wenshao
- 防止分页完成后大型对话轮次自动折叠，确保内容保持可见直至手动关闭。([#10086](https://github.com/QwenLM/qwen-code/pull/10086)) by @ytahdn
- 通过等待恢复的数据块而非使用固定时间延迟，使 delta-resume 重连测试变为确定性测试。([#10203](https://github.com/QwenLM/qwen-code/pull/10203)) by @qwen-code-dev-bot
- 在锚定发布说明前验证 git 中是否存在上一版本标签，若缺失则回退至最新的稳定标签。([#10167](https://github.com/QwenLM/qwen-code/pull/10167)) by @yiliang114
- 在非 Linux 环境的 E2E 分片测试中豁免致命的未处理 vitest RPC 超时错误，同时保持 Linux 环境下的严格错误处理。([#10112](https://github.com/QwenLM/qwen-code/pull/10112)) by @qwen-code-dev-bot
- 确保工具摘要始终可折叠，防止子工具自动展开，并将 MCP 应用与其他聚合内容分离。([#10231](https://github.com/QwenLM/qwen-code/pull/10231)) by @ytahdn
- 确保无头（headless）和 ACP 主机中的工具结果均带有 Goal 轮次许可标记，以便作为 Goal 推理的可引用证据。([#10175](https://github.com/QwenLM/qwen-code/pull/10175)) by @qqqys
- 移除 telemetry-swap 测试中重复的 getToolRegistry mock 属性，以修复由重复对象键引起的 TypeScript 构建错误。([#10252](https://github.com/QwenLM/qwen-code/pull/10252)) by @wenshao
- 通过从遥测交换测试配置 mock 中移除重复的 getToolRegistry 属性，修复 TypeScript 构建失败问题。([#10243](https://github.com/QwenLM/qwen-code/pull/10243)) by @AaronZ345
- 将 qwen3.8-flash 和 qwen3.8-plus 加入模态模式，以启用原生的图像和视频处理，替代通过视觉桥接路由。([#10195](https://github.com/QwenLM/qwen-code/pull/10195)) by @yiliang114
- 将 goal-runtime 启动等待时间限制为 5 秒以防止挂起，并在 Bun 环境下跳过不必要的内存重启步骤。([#10128](https://github.com/QwenLM/qwen-code/pull/10128)) by @chiga0
- 在 POSIX 和 Windows 系统上为 ACP 子进程实现了健壮的流程树回收机制，以防止在关机或出错时产生孤儿进程。([#10142](https://github.com/QwenLM/qwen-code/pull/10142)) by @doudouOUC
- 确保针对 qwen3.8-max 的模型专属推理控制在会话执行前、中、后均保持可用且正确同步。([#10244](https://github.com/QwenLM/qwen-code/pull/10244)) by @callmeYe
- 修复了命令钩子清理逻辑，确保在钩子超时或被取消时，能在 POSIX 和 Windows 上可靠地回收整个进程树。([#10100](https://github.com/QwenLM/qwen-code/pull/10100)) by @doudouOUC
- 自动检测功能现能正确识别 deepseek-v4-flash-vision-exp、glm-4.6v、glm-5v-turbo 和 glm-5.3-flash 为支持视觉的模型。([#10278](https://github.com/QwenLM/qwen-code/pull/10278)) by @yiliang114
- 将自动修复代理的时间预算增加至修复轮次 60 分钟、Runner 默认 90 分钟，以防止任务过早终止。([#10217](https://github.com/QwenLM/qwen-code/pull/10217)) by @wenshao
- 更新了流空闲超时错误提示，引导用户通过环境变量和显式设置进行解决。([#9896](https://github.com/QwenLM/qwen-code/pull/9896)) by @yu-xin-c
- Web Shell 新建会话欢迎页标题现采用优化后的排版、间距、淡入动画，以及与 Qwen 标志匹配的渐变品牌色。([#10230](https://github.com/QwenLM/qwen-code/pull/10230)) by @wenshao
- 附件的 updatedAt 时间戳和 sizeBytes 现会在内容变更时立即更新，确保 Web Shell 预览和缓存反映最新的文件状态。([#9929](https://github.com/QwenLM/qwen-code/pull/9929)) by @zjgzx1988
- 扩展技能现区分标准标识符与显示名称，确保匹配一致性，同时在 CLI 和 Web Shell 中展示友好的本地化标签。([#10296](https://github.com/QwenLM/qwen-code/pull/10296)) by @callmeYe
- 代理视图现能在切换标签页时保留共享上下文中的排队跟进消息，防止在代理忙碌时队友切换视图导致数据丢失。([#10102](https://github.com/QwenLM/qwen-code/pull/10102)) by @yiliang114
- 闲置团队回收现在具备代际安全性，如果自闲置检查以来团队配置已发生变更，将中止删除操作以防止意外数据丢失。 ([#10236](https://github.com/QwenLM/qwen-code/pull/10236)) by @yiliang114
- fix(core): 在事件桥接附加前报告队友结果已结算 ([#10245](https://github.com/QwenLM/qwen-code/pull/10245)) by @yiliang114
- fix(cli): 恢复因目标运行时等待重构而损坏的无目标启动 ([#10290](https://github.com/QwenLM/qwen-code/pull/10290)) by @yiliang114
- fix(test): 在 Docker 中挂载频道插件测试夹具 ([#10318](https://github.com/QwenLM/qwen-code/pull/10318)) by @yiliang114
- fix(ci): 修复共享池主机导致全绿 E2E 运行变红的问题 (#10325) ([#10329](https://github.com/QwenLM/qwen-code/pull/10329)) by @qwen-code-dev-bot
- 集成测试现在使用隔离的全局 qwen 目录运行，以防止主机配置和已保存的记忆影响测试结果。 ([#10337](https://github.com/QwenLM/qwen-code/pull/10337)) by @wenshao
- E2E 工作流现在包含针对 Linux docker 阶段临时沙箱镜像构建失败的限制重试机制，以提升 CI 稳定性。 ([#10359](https://github.com/QwenLM/qwen-code/pull/10359)) by @qwen-code-dev-bot
- 将归档操作移至会话下拉菜单，并针对运行中的会话禁用该操作，以防止 UI 重叠和意外归档。 ([#10302](https://github.com/QwenLM/qwen-code/pull/10302)) by @wenshao
- 技能切换端点现在直接将请求的名称持久化到工作区设置中，而无需针对运行时技能目录进行验证。 ([#10077](https://github.com/QwenLM/qwen-code/pull/10077)) by @callmeYe
- fix(test): 将工具控制 E2E 测试与共享状态隔离 ([#10340](https://github.com/QwenLM/qwen-code/pull/10340)) by @yiliang114
- 收尾交接标记现在仅在轮次成功交付给模型时才会添加，以确保预算处理正确。 ([#10260](https://github.com/QwenLM/qwen-code/pull/10260)) by @qqqys
- 直接 DWS 最终响应现在使用标准直接消息传递和恢复路径，以确保消息不会静默丢失。 ([#10274](https://github.com/QwenLM/qwen-code/pull/10274)) by @qqqys
- 通过确保在并发测试夹具中取消和分离响应后 HTTP 连接正确关闭，修复了间歇性的 Java 11 测试失败问题。 ([#10365](https://github.com/QwenLM/qwen-code/pull/10365)) by @yiliang114
- 更新 ACP 定时测试夹具，使其根据提示词内容而非请求顺序路由响应，防止由并行测试干扰导致的失败。 ([#10374](https://github.com/QwenLM/qwen-code/pull/10374)) by @yiliang114
- 配置 Linux E2E 分片使用 /var/tmp 下隔离的磁盘支持临时目录，以防止测试执行期间 /tmp 耗尽错误。 ([#10376](https://github.com/QwenLM/qwen-code/pull/10376)) by @yiliang114
- 更新分类工作流，使其在重新运行期间 PR 被关闭时静默退出，防止出现关于缺少机器人审查的不必要警告。 ([#10334](https://github.com/QwenLM/qwen-code/pull/10334)) by @yiliang114
- 更新分类工作流，当响应指示模型层 API 错误时正确使检查步骤失败，而不是将其视为成功。 ([#10333](https://github.com/QwenLM/qwen-code/pull/10333)) by @yiliang114
- ack-review-request 作业现在直接在触发命令下方重新发布确认信息，并添加表情反应以提高可见性。 ([#10381](https://github.com/QwenLM/qwen-code/pull/10381)) by @qqqys
- 通过引入真实的 AbortSignal 超时、类型化错误、授权重连和一致的前台交付，增强了 Computer Use 会话的健壮性。 ([#10319](https://github.com/QwenLM/qwen-code/pull/10319)) by @LaZzyMan
- 通过引入 settings.tools.eager 控制模式可见性而不影响自动批准，将 permissions.allow 与工具注册解耦。 ([#10098](https://github.com/QwenLM/qwen-code/pull/10098)) by @yiliang114
### 文档
- 从代理上下文文件中移除冗余且非操作性的文本，在不更改项目规则或命令的情况下减少 token 消耗。 ([#10137](https://github.com/QwenLM/qwen-code/pull/10137)) by @DragonnZhang
- docs(external-context): 设计可配置的 Mem0 提供程序扩展 ([#10113](https://github.com/QwenLM/qwen-code/pull/10113)) by @doudouOUC
### 内部变更
- 通过向 Config 模拟对象添加缺失的 getToolRegistry 方法修复 telemetry-swap 测试，以防止初始化错误。 ([#10220](https://github.com/QwenLM/qwen-code/pull/10220)) by @wenshao
- 在移除 Electron 后，从 CI 工具和文档中清除了对已删除 packages/desktop 目录的陈旧引用。 ([#10021](https://github.com/QwenLM/qwen-code/pull/10021)) by @yiliang114
- 将包含旧版 Gemini 前缀的内部标识符重命名为通用名称，同时保留已弃用的别名以维持向后兼容性。 ([#9900](https://github.com/QwenLM/qwen-code/pull/9900)) by @yiliang114
- 集中化工作树配置派生逻辑，以确保在所有启动路径中一致地处理工作区路径和上下文。 ([#9915](https://github.com/QwenLM/qwen-code/pull/9915)) by @yiliang114
- refactor(core,cli): 将通用 Gemini 标识符重命名为 Llm ([#10124](https://github.com/QwenLM/qwen-code/pull/10124)) by @yiliang114
- 修复发布分类器测试桩，使其在 Node 将无扩展名可执行文件解释为 CommonJS 或 ESM 时能正确工作。 ([#10402](https://github.com/QwenLM/qwen-code/pull/10402)) by @ytahdn
**完整更新日志**: https://github.com/QwenLM/qwen-code/compare/v0.22.2...v0.22.3
<!-- qwen-release-notes:v1 -->
## Highlights
- Added owner-scoped named sessions in Channels to manage up to eight persistent tasks per chat. ([#10198](https://github.com/QwenLM/qwen-code/pull/10198))
- Enabled daemon Extension installs to accept absolute local paths while rejecting relative paths and unsupported options. ([#10131](https://github.com/QwenLM/qwen-code/pull/10131))
- Updated Web Shell sidebar to display GitHub-style colored state icons for sessions bound to pull requests. ([#10258](https://github.com/QwenLM/qwen-code/pull/10258))
- Added an opt-in composer add menu in Web Shell grouping attachments, files, extensions, MCP servers, and skills. ([#10379](https://github.com/QwenLM/qwen-code/pull/10379))
- Restored bold text visibility in assistant messages and prevented sidebar session names from being truncated. ([#10157](https://github.com/QwenLM/qwen-code/pull/10157))
- Ensured tool summaries are consistently collapsible and separated MCP Apps from other aggregations. ([#10231](https://github.com/QwenLM/qwen-code/pull/10231))
## Breaking Changes
No known breaking changes.
## Complete Change List
### Features
- Added a private OpenTelemetry span attribute to report detailed LLM context usage metrics including input categories and remaining capacity. ([#10016](https://github.com/QwenLM/qwen-code/pull/10016)) by @doudouOUC
- Enabled daemon Extension installs to accept absolute local paths while rejecting relative paths and unsupported options like ref. ([#10131](https://github.com/QwenLM/qwen-code/pull/10131)) by @callmeYe
- Backfills PR bindings for existing sessions and adds merge-state snapshots to track whether linked pull requests are open, merged, or closed. ([#9729](https://github.com/QwenLM/qwen-code/pull/9729)) by @wenshao
- Adds an opt-in setting for embedded Web Shell hosts to hide the session source switch, pinning all sidebar catalogs to ordinary task conversations. ([#10079](https://github.com/QwenLM/qwen-code/pull/10079)) by @dreamWB
- Implements safe projection for reserved [FILE: ...] output markers by redacting malformed or oversized inline markers without uploading files. ([#9991](https://github.com/QwenLM/qwen-code/pull/9991)) by @qqqys
- Updates Goal prompts to notify the model exactly once when the objective changes, while continuing to assert the current objective on every turn. ([#10013](https://github.com/QwenLM/qwen-code/pull/10013)) by @qqqys
- Added support for owner-scoped named sessions in Channels, allowing users to manage up to eight persistent named tasks per chat. ([#10198](https://github.com/QwenLM/qwen-code/pull/10198)) by @doudouOUC
- Updated Web Shell sidebar to display GitHub-style colored state icons for sessions bound to pull requests based on their open, merged, or closed status. ([#10258](https://github.com/QwenLM/qwen-code/pull/10258)) by @wenshao
- Made deferred review suggestions recoverable by external tooling via a stable invisible marker in the review body artifact. ([#9761](https://github.com/QwenLM/qwen-code/pull/9761)) by @wenshao
- The /review command now warns when a subsystem's Critical issues repeatedly regrow across rounds by detecting successor chains. ([#10010](https://github.com/QwenLM/qwen-code/pull/10010)) by @wenshao
- Restored the desktop brand builder skill for the Tauri shell to generate branded packages using brandId and logo inputs. ([#10164](https://github.com/QwenLM/qwen-code/pull/10164)) by @yiliang114
- Added a standalone daemon session API supporting create, list, resume, archive, and delete operations for top-level sessions. ([#10179](https://github.com/QwenLM/qwen-code/pull/10179)) by @doudouOUC
- The /review command now remembers the last explicitly used --effort level per project for subsequent runs without arguments. ([#9883](https://github.com/QwenLM/qwen-code/pull/9883)) by @tlysanhuo
- Long-running review rounds now update their status comment every 10 minutes with elapsed time and last agent activity to distinguish active work from stalled processes. ([#9771](https://github.com/QwenLM/qwen-code/pull/9771)) by @wenshao
- Local review-fix loops now anchor incremental rounds to working tree content changes instead of commits, reducing token usage by avoiding full tree re-reviews. ([#9659](https://github.com/QwenLM/qwen-code/pull/9659)) by @wenshao
- Review findings now support an optional fixConstraint field to specify existing facts or limits that suggested fixes must not violate during the repair process. ([#10168](https://github.com/QwenLM/qwen-code/pull/10168)) by @wenshao
- feat(sdk): Add standalone session APIs ([#10294](https://github.com/QwenLM/qwen-code/pull/10294)) by @doudouOUC
- feat(cli): OpenTUI migration infra batch — locked deps, dependency-direction arch check, framework-neutral streaming model ([#10134](https://github.com/QwenLM/qwen-code/pull/10134)) by @chiga0
- Added a configurable Mem0 extension skeleton providing a retrieval-only external-context-mem0 stdio Extension with bounded HTTP request support. ([#10149](https://github.com/QwenLM/qwen-code/pull/10149)) by @doudouOUC
- Added a new triage gate that automatically identifies and closes duplicate PRs when their changes are already fully resolved by merged fixes. ([#10292](https://github.com/QwenLM/qwen-code/pull/10292)) by @yiliang114
- Added an opt-in composer add menu to Web Shell that groups attachments, files, extensions, MCP servers, and skills for easier pointer-oriented selection. ([#10379](https://github.com/QwenLM/qwen-code/pull/10379)) by @ytahdn
- Added an opt-in interactive browser terminal to Web Shell that manages independent PTY sessions per tab with persistent scrollback and state handling. ([#9984](https://github.com/QwenLM/qwen-code/pull/9984)) by @ytahdn
- Adds the qwen review ab-drive subcommand to run execution-grade verification scripts against PR and base trees with shared or fresh instances. ([#9740](https://github.com/QwenLM/qwen-code/pull/9740)) by @wenshao
- Confirmed Critical issues now include direction and baseline axes to distinguish between false certifications, failures, regressions, and new surfaces. ([#10301](https://github.com/QwenLM/qwen-code/pull/10301)) by @wenshao
- Makes Aone a first-class platform for session PR bindings, enabling automatic branch-to-MR mapping and state refresh via the a1 CLI. ([#10287](https://github.com/QwenLM/qwen-code/pull/10287)) by @wenshao
### Bug Fixes
- Fixed Web Shell to correctly restore saved session diffs with accurate line counts even when file bodies were previously truncated. ([#10093](https://github.com/QwenLM/qwen-code/pull/10093)) by @ytahdn
- Ensured DingTalk rich-text messages preserve all images as ordered attachments across sessions and handle upload failures gracefully. ([#9922](https://github.com/QwenLM/qwen-code/pull/9922)) by @BenGuanRan
- fix: repair the Windows test lane's standing failures ([#10064](https://github.com/QwenLM/qwen-code/pull/10064)) by @wenshao
- fix(core): report broadcast delivery failures in send_message ([#10081](https://github.com/QwenLM/qwen-code/pull/10081)) by @yiliang114
- Moved the autofix scan lane to the persistent self-hosted runner pool to prevent GitHub-hosted runner backlogs from blocking routing. ([#10055](https://github.com/QwenLM/qwen-code/pull/10055)) by @wenshao
- Restores bold text visibility in assistant messages and prevents session names in the sidebar from being truncated by attention pills. ([#10157](https://github.com/QwenLM/qwen-code/pull/10157)) by @wenshao
- Prevents large transcript turns from collapsing automatically after pagination completes, ensuring content remains visible until manually closed. ([#10086](https://github.com/QwenLM/qwen-code/pull/10086)) by @ytahdn
- Makes the delta-resume reconnect test deterministic by waiting for the resumed chunk instead of using a fixed time delay. ([#10203](https://github.com/QwenLM/qwen-code/pull/10203)) by @qwen-code-dev-bot
- Verifies the previous release tag exists in git before anchoring release notes, falling back to the latest stable tag if missing. ([#10167](https://github.com/QwenLM/qwen-code/pull/10167)) by @yiliang114
- Exempts off-Linux E2E shards from fatal unhandled vitest RPC-timeout errors while maintaining strict error handling on Linux. ([#10112](https://github.com/QwenLM/qwen-code/pull/10112)) by @qwen-code-dev-bot
- Ensures tool summaries are consistently collapsible, prevents child tools from opening automatically, and separates MCP Apps from other aggregations. ([#10231](https://github.com/QwenLM/qwen-code/pull/10231)) by @ytahdn
- Ensures tool results in headless and ACP hosts are stamped with Goal turn permits so they can serve as citable evidence for Goal reasoning. ([#10175](https://github.com/QwenLM/qwen-code/pull/10175)) by @qqqys
- Removes a duplicate getToolRegistry mock property in telemetry-swap tests to fix TypeScript build errors caused by duplicate object keys. ([#10252](https://github.com/QwenLM/qwen-code/pull/10252)) by @wenshao
- Fixes TypeScript build failure by removing a duplicate getToolRegistry property from the telemetry swap test configuration mock. ([#10243](https://github.com/QwenLM/qwen-code/pull/10243)) by @AaronZ345
- Adds qwen3.8-flash and qwen3.8-plus to modality patterns to enable native image and video handling instead of routing through the vision bridge. ([#10195](https://github.com/QwenLM/qwen-code/pull/10195)) by @yiliang114
- Bounds the goal-runtime startup wait to 5 seconds to prevent hanging and skips unnecessary memory relaunch steps when running under Bun. ([#10128](https://github.com/QwenLM/qwen-code/pull/10128)) by @chiga0
- Implemented robust process tree reaping for ACP child processes on POSIX and Windows to prevent orphaned processes during shutdown or errors. ([#10142](https://github.com/QwenLM/qwen-code/pull/10142)) by @doudouOUC
- Ensured model-specific reasoning controls remain available and correctly synchronized for qwen3.8-max before, during, and after session execution. ([#10244](https://github.com/QwenLM/qwen-code/pull/10244)) by @callmeYe
- Fixed command hook cleanup to reliably reclaim entire process trees on POSIX and Windows when hooks time out or are cancelled. ([#10100](https://github.com/QwenLM/qwen-code/pull/10100)) by @doudouOUC
- Auto-detection now correctly identifies deepseek-v4-flash-vision-exp, glm-4.6v, glm-5v-turbo, and glm-5.3-flash as vision-capable models. ([#10278](https://github.com/QwenLM/qwen-code/pull/10278)) by @yiliang114
- Increased autofix agent time budgets to 60 minutes for repair passes and 90 minutes for the runner default to prevent premature task termination. ([#10217](https://github.com/QwenLM/qwen-code/pull/10217)) by @wenshao
- Updated stream idle timeout error messages to guide users to both environment variables and explicit settings for resolution. ([#9896](https://github.com/QwenLM/qwen-code/pull/9896)) by @yu-xin-c
- The Web Shell new-session welcome header now features improved typography, spacing, a fade-in animation, and a brand-aligned gradient matching the Qwen logo. ([#10230](https://github.com/QwenLM/qwen-code/pull/10230)) by @wenshao
- Artifact updatedAt timestamps and sizeBytes now update immediately when content changes, ensuring Web Shell previews and caches reflect the latest file state. ([#9929](https://github.com/QwenLM/qwen-code/pull/9929)) by @zjgzx1988
- Extension Skills now separate canonical identity from display names, ensuring consistent matching while showing friendly localized labels in the CLI and Web Shell. ([#10296](https://github.com/QwenLM/qwen-code/pull/10296)) by @callmeYe
- Agent View now preserves queued follow-up messages in shared context across tab switches, preventing data loss when teammates switch views while an agent is busy. ([#10102](https://github.com/QwenLM/qwen-code/pull/10102)) by @yiliang114
- Stale team reclamation is now generation-safe, aborting deletion if the team configuration has changed since the staleness check to prevent accidental data loss. ([#10236](https://github.com/QwenLM/qwen-code/pull/10236)) by @yiliang114
- fix(core): report teammate result settled before event bridge attach ([#10245](https://github.com/QwenLM/qwen-code/pull/10245)) by @yiliang114
- fix(cli): restore goal-less startup broken by the goal-runtime wait refactor ([#10290](https://github.com/QwenLM/qwen-code/pull/10290)) by @yiliang114
- fix(test): mount channel plugin fixture in Docker ([#10318](https://github.com/QwenLM/qwen-code/pull/10318)) by @yiliang114
- fix(ci): stop shared-pool hosts exiting all-green E2E runs red (#10325) ([#10329](https://github.com/QwenLM/qwen-code/pull/10329)) by @qwen-code-dev-bot
- Integration tests now run with an isolated global qwen dir to prevent host configuration and saved memories from affecting test results. ([#10337](https://github.com/QwenLM/qwen-code/pull/10337)) by @wenshao
- E2E workflows now include a bounded retry for transient sandbox image build failures in the Linux docker leg to improve CI stability. ([#10359](https://github.com/QwenLM/qwen-code/pull/10359)) by @qwen-code-dev-bot
- Moved the archive action to the session dropdown and disabled it for running sessions to prevent UI overlap and accidental archiving. ([#10302](https://github.com/QwenLM/qwen-code/pull/10302)) by @wenshao
- Skill toggle endpoints now persist requested names directly to workspace settings without validating against the runtime Skill catalog. ([#10077](https://github.com/QwenLM/qwen-code/pull/10077)) by @callmeYe
- fix(test): isolate tool-control E2E from shared state ([#10340](https://github.com/QwenLM/qwen-code/pull/10340)) by @yiliang114
- Wind-down hand-off markers are now stamped only when the turn is successfully delivered to the model to ensure correct budget handling. ([#10260](https://github.com/QwenLM/qwen-code/pull/10260)) by @qqqys
- Direct DWS final responses now use standard direct-message delivery and recovery paths to ensure messages are not silently dropped. ([#10274](https://github.com/QwenLM/qwen-code/pull/10274)) by @qqqys
- Fixed an intermittent Java 11 test failure by ensuring HTTP connections close properly after cancel and detach responses in the concurrency test fixture. ([#10365](https://github.com/QwenLM/qwen-code/pull/10365)) by @yiliang114
- Updated the ACP cron test fixture to route responses based on prompt content instead of request order, preventing failures caused by parallel test interference. ([#10374](https://github.com/QwenLM/qwen-code/pull/10374)) by @yiliang114
- Configured Linux E2E shards to use isolated disk-backed temporary directories under /var/tmp to prevent /tmp exhaustion errors during test execution. ([#10376](https://github.com/QwenLM/qwen-code/pull/10376)) by @yiliang114
- Updated the triage workflow to silently exit when a PR is closed during a re-run, preventing unnecessary warnings about missing bot reviews. ([#10334](https://github.com/QwenLM/qwen-code/pull/10334)) by @yiliang114
- Updated the triage workflow to correctly fail the check step when the response indicates a model-layer API error instead of treating it as success. ([#10333](https://github.com/QwenLM/qwen-code/pull/10333)) by @yiliang114
- The ack-review-request job now reposts acknowledgements directly under the triggering command and adds reactions to improve visibility. ([#10381](https://github.com/QwenLM/qwen-code/pull/10381)) by @qqqys
- Hardens Computer Use sessions with real AbortSignal timeouts, typed errors, authorization reconnection, and consistent foreground delivery. ([#10319](https://github.com/QwenLM/qwen-code/pull/10319)) by @LaZzyMan
- Decouples permissions.allow from tool registration by introducing settings.tools.eager to control schema visibility without affecting auto-approval. ([#10098](https://github.com/QwenLM/qwen-code/pull/10098)) by @yiliang114
### Documentation
- Removes redundant and non-operational text from agent context files to reduce token usage without altering project rules or commands. ([#10137](https://github.com/QwenLM/qwen-code/pull/10137)) by @DragonnZhang
- docs(external-context): Design a configurable Mem0 provider extension ([#10113](https://github.com/QwenLM/qwen-code/pull/10113)) by @doudouOUC
### Internal Changes
- Fixed telemetry-swap tests by adding the missing getToolRegistry method to the Config mock to prevent initialization errors. ([#10220](https://github.com/QwenLM/qwen-code/pull/10220)) by @wenshao
- Removed stale references to the deleted packages/desktop directory from CI tooling and documentation following the Electron removal. ([#10021](https://github.com/QwenLM/qwen-code/pull/10021)) by @yiliang114
- Renamed internal identifiers containing legacy Gemini prefixes to generic names while maintaining deprecated aliases for backward compatibility. ([#9900](https://github.com/QwenLM/qwen-code/pull/9900)) by @yiliang114
- Centralized worktree configuration derivation to ensure consistent handling of workspace paths and context across all launch paths. ([#9915](https://github.com/QwenLM/qwen-code/pull/9915)) by @yiliang114
- refactor(core,cli): rename generic Gemini identifiers to Llm ([#10124](https://github.com/QwenLM/qwen-code/pull/10124)) by @yiliang114
- Fixes the release classifier test stub to work correctly when Node interprets extensionless executables as either CommonJS or ESM. ([#10402](https://github.com/QwenLM/qwen-code/pull/10402)) by @ytahdn
**Full Changelog**: https://github.com/QwenLM/qwen-code/compare/v0.22.2...v0.22.3