dsh-super-code 0.3.11 → 0.4.0

package/cordis.patch.yml

@@ -4,10 +4,7 @@
     - id: preset-super-code
       name: '@deepseek-ai/dsh-agent-preset'
       disabled: !!js >-
-        (() => { try {
-          process.getBuiltinModule('node:module').createRequire(process.argv[1]).resolve('@deepseek-ai/dsh-agent-preset')
-          return false
-        } catch { return true } })()
+        !get('pluginPackages')?.packageOf('@deepseek-ai/dsh-agent-preset', baseUrl)
       config:
         id: super-code
         name: Super Code
@@ -17,7 +14,7 @@
           - id: super-code
             name: dsh-super-code/super-code
             config:
-              presetVersion: 0.3.11
+              presetVersion: 0.4.0
               presetInstallation: bundle:dsh-super-code
           - id: compaction
             name: cordis:group
@@ -61,16 +58,11 @@
           # Declarations may be persisted in a profile patch. Resolve the
           # installed package instead of treating that patch as a skill root.
           - id: skill-super-code-bundled
-            name: '@deepseek-ai/dsh-skill-filesystem'
+            name: dsh-super-code/skills
             config:
-              providerName: super-code-bundled
-              includeDefaultRoots: false
-              watch: false
+              layout: builtin
               bundledSkillDir: !!js >-
-                (() => {
-                  const require = process.getBuiltinModule('node:module').createRequire(new URL('package.json', baseUrl))
-                  return require('node:path').join(require('node:path').dirname(require.resolve('dsh-super-code/package.json')), 'presets/super-code/skills')
-                })()
+                get('pluginPackages').packageOf('dsh-super-code', baseUrl).dir + '/builtin/skills'
           - id: tool-skill
             name: '@deepseek-ai/dsh-tool-skill'
           - id: tool-presentation

package/docs/Acknowledgements.md

@@ -31,23 +31,17 @@
 
 内置前端 Skill 按该规范组织：`SKILL.md`、YAML 元数据和 `references/` 目录。平时仅向模型提供名称和简介，需要时才加载正文和参考资料，可与项目及用户自有的 Skill 共存。解析和加载使用 Harness 原生能力。
 
-对应实现：[super-code-frontend](../presets/super-code/skills/super-code-frontend/SKILL.md)
+对应实现：[oil-ui](../builtin/skills/oil-ui/SKILL.md)
 
-## Apple Human Interface Guidelines
+## oil-ui
 
-[布局](https://developer.apple.com/design/human-interface-guidelines/layout) · [无障碍](https://developer.apple.com/design/human-interface-guidelines/accessibility) · [按钮](https://developer.apple.com/design/human-interface-guidelines/buttons) · [反馈](https://developer.apple.com/design/human-interface-guidelines/feedback) · [条款](https://developer.apple.com/support/terms/)
+[源码](https://github.com/oil-oil/oil-ui) · [MIT](../builtin/skills/oil-ui/LICENSE) · [固定版本与改动](../builtin/skills/oil-ui/UPSTREAM.md)
 
-前端 Skill 采纳了其中的原则：以对齐表达关系、布局适应容器和字号、通过控件尺寸和间距降低操作负担、交互可预测、反馈就近且适度。上述原则整理为视觉设计、交互状态和浏览器验证三份参考文档。
+感谢 oil-oil 提供开源前端设计 Skill。本插件内置 oil-ui 0.32.0 的设计流程、参考资料、风格对比页及辅助脚本，通过 Harness 原生 Skill 机制按需加载，替代原来自写的前端 Skill。
 
-仅借鉴原则，产品仍保持自身视觉风格。原生平台的尺寸单位不直接作为 CSS 像素使用，材质、字体和图标也不作为模板。
+内置副本固定到上游 commit `0d64a00f3a5c45bf7a30d35bded9b7c0fdff2128`，仅将独立版本检查和更新入口改为随插件升级，未引入上游安装器或付费插件。MIT 许可证与原作者版权声明随包保留。
 
-## W3C WCAG 2.2
-
-[WCAG 2.2](https://www.w3.org/TR/WCAG22/) · [2.5.8 目标尺寸（最小）](https://www.w3.org/WAI/WCAG22/Understanding/target-size-minimum.html)
-
-前端 Skill 检查 Web 控件时，采用 2.5.8 的 24 × 24 CSS 像素标准及其间距例外，并结合实际点击、键盘和内容检查。尺寸达标不代表整个界面满足无障碍要求。
-
-对应实现：[交互状态](../presets/super-code/skills/super-code-frontend/references/interaction-states.md)、[浏览器验证](../presets/super-code/skills/super-code-frontend/references/browser-validation.md)
+风格对比页还包含 [Instrument Serif](https://github.com/Instrument/instrument-serif)，采用 [SIL OFL 1.1](../builtin/skills/oil-ui/assets/fonts/instrument-serif/OFL.txt)，用于对比页字体。
 
 ## SWE-bench
 
@@ -74,3 +68,21 @@
 ## 许可证
 
 本项目采用 [Apache-2.0](../LICENSE) 许可证。上述项目、规范、设计资料和数据各自保留版权与许可证，以上游的 LICENSE、NOTICE 和说明为准。设计参考不代表依赖关系、上游背书或功能等价。
+
+## OpenSpec
+
+[Source](https://github.com/Fission-AI/OpenSpec) · [MIT](../builtin/skills/openspec/LICENSE) · [Upstream](../builtin/skills/openspec/UPSTREAM.md)
+
+感谢 OpenSpec 贡献者提供完整的规范驱动工作流 Skill。插件保留上游正文和许可证，通过 DSH 原生机制按需加载，不自动安装其 CLI。
+
+## Matt Pocock Skills
+
+[Source](https://github.com/mattpocock/skills) · [MIT](../builtin/skills/mattpocock/LICENSE) · [Upstream](../builtin/skills/mattpocock/UPSTREAM.md)
+
+感谢 Matt Pocock 及贡献者提供完整的工程与生产力 Skill 包。插件内置正式插件清单及其配套参考、模板和调用规则，不在启用时运行 setup。
+
+## PStack
+
+[Source](https://github.com/cursor/plugins/tree/main/pstack) · [MIT](../builtin/skills/pstack/LICENSE) · [Upstream](../builtin/skills/pstack/UPSTREAM.md)
+
+感谢 Lauren Tan 与 Cursor 插件贡献者提供 PStack。插件保留其58个标准 Skill 的正文、调用规则、配套参考与脚本，通过 DSH 原生 Skill 和斜杠命令按需加载；不注册 Cursor 自动化，也不在开启时运行安装或全局配置。部分流程仍依赖 Cursor 或外部工具，详见上游说明。

package/docs/Changelog.md

@@ -4,6 +4,33 @@
 
 版本号格式为「主版本.效果版本.维护版本」。第二位变化表示编码效果或效率有提升；第三位表示修复、界面和兼容性更新。各版本的验证范围见[兼容性](Compatibility.md)。
 
+## 0.4.0 · 2026-11-12
+
+**任务与记忆工具**
+
+- 简化 `super_code_task` 和 `super_code_memory` 的调用结构，按具体操作提供字段说明、调用示例和错误后的处理方法，减少试错。
+- 支持不创建或切换当前任务，直接读取指定主题的项目记忆；写入仍保留来源核对和并发修改检查，历史信息不代替当前代码验证。
+
+**预置 Skill 与工作流**
+
+- 新增独立的 grill-me，以及 OpenSpec、oil-ui、Matt Pocock Skills、PStack，分为 Skill 和工作流两组；五项默认全部关闭，保留用户已保存的选择。
+- 开启后将完整资源同步到 DSH 的 Skill 目录，已有对话实时生效；每项显示版本和一句使用介绍，提供打开目录的快捷按钮。保护用户已有文件和修改，冲突时提示并支持重试。
+- 沿用原生 `/` 菜单和真实 Skill 名称，支持通过 `/pstack`、`/mattpocock` 等包名搜索并选择具体命令；关闭后不再提供该包的命令和正文，已经加载到上下文的内容不会被撤回。
+- OpenSpec 随插件准备命令环境；开关不会初始化项目、安装项目依赖或修改项目文件，只有用户要求初始化时才执行。
+- PStack 保留上游工作流与资源，并针对 DSH 的工具、子 Agent、审批和图片处理方式进行适配；流程适配面向 Harness `0.2.1-alpha.1` 及之后版本。
+- 内置及适配后的源码统一公开于 `builtin/skills`，保留上游版本、来源和许可证。
+
+**界面、兼容与修复**
+
+- 完善设置自动保存、跨标签草稿保留、服务重连、并发修改和保存失败后的提示与恢复。
+- 改进任务、执行记录、工具状态和文件变更的展示，保留圆圈 Agent 执行树及旧版记忆页面。
+- 修复预设安装与卸载恢复、MCP 连接与审批、删除保护、Windows 命令环境等功能在取消、重连和多次启停下的问题。
+- 更新新版 Harness 的兼容记录；各版本、平台、SDK 检查与完整宿主验收范围分别见[兼容性记录](Compatibility.md)，不将接口测试等同于全部工作流或全平台验收。
+
+- README 增加版本、许可证、npm 总下载量徽章，以及每日自动刷新的 Star 与 npm 下载趋势图。
+
+效果对比及统计限制见 [README](../README.md) 和[评测说明](EvaluationResults.md)；不将 Skill 安装、界面或兼容性测试当作编码效果证明。
+
 ## 0.3.11 · 2026-11-06
 
 **卸载恢复与预设管理**

package/docs/Compatibility.md

@@ -2,11 +2,13 @@
 
 中文 · [English](en/Compatibility.md)
 
-Web 已验证 Harness `0.1.5-alpha.1` ~ `0.2.1-rc.4`，其中 `rc.4` 为下述增量验收；官方桌面已验证至 `0.2.1-rc.4`（x64 包在 Rosetta 下），NEXT 已验证 `2.0.19-next`（内置 `0.2.1-rc.2`）。
+支持 Harness `0.1.5-alpha.1` ～ `0.2.2-rc.2`，包括 `0.2.2-alpha.3`。0.4.0 已逐版验证这一区间的 27 个完整 Web 宿主，并补验官方 macOS arm64 Desktop `0.2.2-rc.2` 和 NEXT `2.0.22-next`（内置 `0.2.2-rc.1`），本轮范围和限制见下文。
 
-0.3.10 的内置 Windows Bash 另在官方 Desktop `0.2.1-rc.5` 和 NEXT `2.0.21-next`（内置 `0.2.1-rc.4`）完成 Windows 完整进程专项验收，范围见下文。
+Windows 的完整进程验证目前至官方 Desktop `0.2.1-rc.5`、NEXT `2.0.21-next`（内置 `0.2.1-rc.4`）；最新 macOS/Web 结果不扩大 Windows 的已验证范围。
 
-0.3.11 的修复按阶段补充了至 `0.2.2-alpha.3` 的版本契约和官方前端验证，完整宿主范围单独记录，见下方 0.3.11 记录。
+以下按批次记录实际代码和测试层级，不将组件、静态前端夹具或早期开发包的结果视为所有版本与平台的完整复验。
+
+新增 PStack 流程适配以 `0.2.1-alpha.1` 为最低版本，不再适配更早宿主。下列旧版本记录属于此前插件功能的历史验证范围。
 
 ## 已验证的宿主
 
@@ -55,6 +57,67 @@ Web 已验证 Harness `0.1.5-alpha.1` ~ `0.2.1-rc.4`，其中 `rc.4` 为下述
 
 `zod`、`js-yaml`、`@modelcontextprotocol/sdk` 为运行依赖，不打包进插件。Windows Shell 另随包携带独立进程运行的第三方程序，其来源、许可证和对应源码见[运行时说明](../runtime/windows/README.md)。
 
+## 2026-11-12 · 0.4.0 兼容复验
+
+保持原有 peer 依赖范围和 `0.1.5-alpha.1` 兼容下限，使用同一个 0.4.0 候选包逐版复验，未为通过测试修改插件运行代码。
+
+| 验证层级 | 范围 | 结果 |
+|---|---|---|
+| 完整 Web 宿主 | 27 个精确发布版本，`0.1.5-alpha.1` ～ `0.2.2-rc.2` | 每版 12 组，共 324 组通过 |
+| 真实 SDK 与磁盘存储 | 同上 27 版 | 每版 6 组，共 162 组通过 |
+| 原生 Skill 加载 | 同上 27 版 | 每版 24 项、98 次正文加载通过；同名 Skill 按优先级去重 |
+| 原生菜单契约 | 同上 27 版 | 通过；`0.1.5` 原生接口不支持按包名扩展搜索，仍可按具体 Skill 名称选择 |
+| 浏览器交互 | 同上 27 版的官方 React、控件和主题，Remote 服务使用夹具 | 每版 15 组，共 405 组通过，无页面脚本错误 |
+| 官方 macOS arm64 Desktop | `0.2.2-rc.2` | 13 组完整宿主检查、4 组实际界面操作通过 |
+| DSH NEXT · macOS arm64 | `2.0.22-next`，内置 `0.2.2-rc.1` | 13 组完整宿主检查、4 组实际界面操作通过 |
+
+- **完整宿主检查**：实际安装候选包，验证预设版本、预算和删除保护设置互不覆盖、五项默认关闭的 Skill 在已有对话实时启停、MCP 默认配置、手动记忆增删与并发冲突、通过原生工具或 PTC 调用任务与记忆工具，以及完整进程冷重启、原对话续用和记忆恢复。
+- **界面检查**：覆盖六个设置标签、Skill 版本与开关、保存失败、预算校验、连接设置、中英文、深浅主题、窄布局、万条任务虚拟滚动、执行树动画和旧记忆维护。两款真实桌面另验证设置加载、版本一致、五项 Skill 默认关闭、点击自动保存，以及 macOS 上命令环境开关禁用。
+- **来源与基础检查**：官方桌面安装包通过更新源 SHA-512 和深度签名校验，333 个 DSH 组件均为 rc.2；NEXT 签名有效，330 个 DSH 组件均为 rc.1。1154 项单元测试、类型检查和构建打包通过。修正测试脚本对并发完成顺序、上游主题打包格式和实际页面结构的假设，原失败记录保留。
+- **边界**：完整宿主使用本地确定性模型适配器，外部模型调用为零；不代表模型效果、58 个 PStack 工作流或每项功能所有分支均通过。本轮未重跑 Windows、Linux 或所有历史原生桌面版本，也未重测原生目录选择器和物理卸载恢复；这些结果仍按各历史批次单列。PStack 完整流程的适配下限仍为 `0.2.1-alpha.1`。
+
+## 2026-11-11 · PStack DSH 适配
+
+- 第六批移除适配层自行追加的禁命令、强制API检查和快照复审流程，四份角色模板恢复为上游原文，并修复委派段转换丢失上游只读语义的问题。只读角色可用无副作用的命令，仍不得写文件或修改外部状态。命令取证另修复稀疏argv漏检和提交后参数变动；925项回归、13版143项资源检查通过，最低/最新Web真实取证专项各9项独立检查通过。真实流程仍观察到只读子任务重定向写临时文件，不能宣称只读范围问题已解决。
+- 适配范围为 `0.2.1-alpha.1` 起的13个 npm 已发布版本，至当前最新 `0.2.2-rc.1`（含 `0.2.2-alpha.3`）。逐版通过原生 Skill 目录、58项正文加载、斜杠注入、调用权限及同名优先级检查；第五批加入命令取证资源加载，共143项通过；任务工具枚举诊断另39项通过。原生子 Agent 参数契约65项、浏览器已有对话实时开关169项通过。浏览器使用真实前端模块，外部服务边界使用夹具，不是13个完整桌面进程。
+- 2026-11-11 在隔离的 macOS arm64 Web `0.2.2-rc.1`、`tokenhub/deepseek/评测模型 A` 上逐项实跑58个入口，每项一个独立场景，均确认原生斜杠注入及真实模型调用。另对12项重跑原场景，并对两项追加澄清；不把重复运行累计到58项分母。
+- 综合实际产物、必需阶段和范围约束，首轮最终36项通过、17项附说明通过、4项失败、1项部分完成。`how`、`teach`、`create-formal-verification-skill` 的只读子任务仍写入全局临时文件；`principle-outcome-oriented-execution` 两次保留了 `add`，明确目标为 `sum({left,right})` 后才通过，原题措辞与验收契约的歧义需单独核对。`ketchup-video` 生成并解码真实视频，但测试路由漏配图片输入声明，未完成其自身视觉复核。不能将58项运行结束写成58项全部通过。
+- 图片归因更正：这不是模型不支持图片。补齐隔离测试路由的 `input: ["text", "image"]` 后，`0.2.1-alpha.1` 与 `0.2.2-rc.1` 的直接图片附件和原生 `read_image` 均让真实模型读出仅存在于图中的随机验证码和图形数量，提供方请求确认包含图片内容。文本模型配置负对照在发送前拒绝图片。rc.1 中文 Web 另通过原生选图、预览、移除、粘贴、拖放及历史回显；本轮未实跑各平台原生桌面的图片交互。
+- 后续对22个非完整通过入口逐项复测。视频流程这次完成14次实际读图，包含每页成片解码帧，独立复核8组 CLI 输出及四页画面；`statechart` 完整流程也通过真实 TLC 与代码重放。基础独立检查56/57通过，唯一失败为原 API 题与固定目标契约的歧义，另5项扩展检查通过。仍观察到只读子任务写全局临时文件、清理使用 `rm` 和按进程名终止编码器；这些问题未被传图成功或基础检查覆盖，不能宣布58项全部验收通过。
+- 再对只读与审计相关3项补测，7条基础检查通过；形式验证独立副本的正常、错误参数、负变异/恢复另3项通过。审计工作流本次reviewer无写入，实际独立评审通过并披露快照覆盖范围；`teach` 的调查员和合成员仍写全局临时文件。因此提示规则尚不构成只读硬隔离，未宣称相关问题全部解决。
+- 本轮修正了三处全局临时路径示例和子任务范围传递指南，并修复 `statechart` 安装于 `node_modules` 后 Node 拒绝执行 TypeScript 的问题：构建时生成 JavaScript，保留上游原件；含中文和空格的安装路径已直接运行真实 TLC、授权反例和代码重放。MCP 网页真实验证新建/复用会话、稳定请求ID、重启、拒绝分支、分页历史及连接清理；审批等待和远程部署分支未实跑。
+- 转换发行保留58个入口、23个 playbook 和558份原件，当前生成592个文件（含审计快照、稳定请求接纳及项目内命令取证helper）；原件哈希、生成物和脚本权限经过校验。第五批全量919项测试、类型检查和构建通过；此前第四批909项及后补19项定向回归记录保留。首次58项场景另有155条独立检查，其中 API 迁移断言失败，其余154条通过，不能用这些检查替代阶段与范围验收。
+- 第三批再跑8个重点入口，20条基础独立检查通过；修复 `why` 转换器误删代码锚点，将只读契约放进实际角色模板。视频重新解码3页及2个转场并亲看，16条独立检查通过，本批无永久 shell 清理或共享进程名终止。新增审计脚本的追加/篡改/CRLF/中文路径等8项通过，完整23行记录另有12项来源与行为检查及逐行语义复核；不把结构校验当作claim为真的证明。
+- MCP后补真实审批等待与取消，6项检查通过，浏览器实际显示 `waiting_confirmation → cancelling → cancelled`；探针命令未获执行。另发现生成示例后端的同ID变更缓存问题，验收者手工修后13项通过，产品工作流补入指纹、串行接纳和原始请求重试要求。这不是底层服务新发现的bug，也不是模型一次自主通过；远程部署及全部错误状态的实时注入未测。
+- 最低版 `0.2.1-alpha.1` 用最终运行候选重新执行7文件how场景，14条独立检查通过；根节点直接取原生命令结果，源码和工作树未变，两个子任务仅read/glob/grep/send_message。本批仍有模型遵循偏差：teach/show各有只读cat/ls；最低版重复派一次explainer且尝试写捕获文件，因ENOENT未写成，最终文字还少报一个child。随后使用包含这条指引的最终包在最低版补测受管与同步两条返回分支，15项检查通过：每例一次委派、无捕获文件/写入尝试或重复派发；同步结果在同一工具调用提取最终文本。早期重复与少报记录保留；同步回复仍超出要求字数。不宣称硬隔离或所有提示规则均已遵守。
+- 第四批为 `make-bot-ui` 提供无依赖的稳定请求接纳helper，覆盖同ID并发、内容冲突、重启、持久化失败及返回值隔离，避免重新实现缓存。真实从零生成的示例完成92项自测；独立发现 `interrupted` 误作结束及故障transport未DELETE，原模型会话收到诊断后修复，91项确定性夹具检查、13项界面及2项连接恢复检查通过。原会话随后触发验收器累计100次模型请求上限，未完成最终交接，不能写成一次自主完整通过。最终打包helper另在真实MCP/模型上通过18项独立复验，包含新建、显式复用、每请求旧/新结果、拒绝与连接释放；夹具状态不冒充真实网络故障。
+- 第四批审计补审全部1–28行工作及更正决策，最终30行中29–30仅记录评审元数据，25项快照、实际CLI及来源检查通过；原证据曾被覆盖、计数和因果叙述有误，独立reviewer仍为PASS WITH FLAGS，历史问题保留。teach的4child无写入但仍有一次禁shell契约下的ls。bot早期曾写4个全局临时响应并使用含rm的备用清理命令，分支是否执行无法判定；追加项目内取证指引后的窄场景中，3个真实HTTP拒绝响应正确，但状态探针仍写全局临时文件，范围检查失败保留。不能把功能检查通过当作范围遵循通过。
+- 第五批使用项目内唯一取证目录，最新Web审计的34组输出均可复验，14项独立来源/CLI检查通过；手抄哈希和行数错误经原会话追加更正，审阅角色仍执行只读shell，现有禁shell契约失败保留。MCP示例从修后checkpoint完成96项真实链路检查、91项明确故障夹具及最终交接，两份命令证据不再覆盖或写全局临时路径；综合20项独立检查19通过，唯一失败即上述角色契约。最低版另跑真实how/取证专项，16项通过：中文空格路径、字面argv、0/7退出、原始输出、证据复验、源码不变及独立角色仅read/search。未把这些专项写成58项全部通过。
+- 本轮58项只在上述 macOS Web 和单一模型上执行，未重跑 Windows/Linux、全部桌面进程或23个 playbook 的所有分支。更早的最低版配置、委派和回传记录仍单独保留；本批how验证不覆盖该版本的所有流程。GitHub 发布、可选 Bun watcher、跨模型比较和远程部署不在本轮验证范围；结果不代表效果或性能提升。
+
+- 完整性复核补回 setup 的8阶段与17类角色、reflect 独立合成、增量历史、模拟器/缓存分支，以及按原因重试/迟到结果处理。推理预算、候选席位与并发数分别处理；无公开模型发现能力时继承宿主，不扫描私有配置。长任务不假定存在定时器或私有 Agent store。
+
+- 复核后真实模型补验：最新版 setup 完成完整角色配置并遵守公开能力发现边界；reflect 实际运行三位独立评审和第四位合成，结构化提案移入 Backlog，审批前不改 Skill。最低版 `0.2.1-alpha.1` 再次验证新 setup 的17类角色、预算/席位/并发分离和原备注保留。保留首轮越界探测、分类未落实及验收脚本格式假设的失败记录；不把完成标记当作验收证据。
+
+## 2026-11-09 · 最新宿主复验
+
+本轮使用 `main@8d28e47`，包含已替换的 oil-ui `0.32.0`。固定测试包 SHA-256 为 `3b19a6c04b960202d87e13bb17e59085f7378503aaf8316a9c0d78220f6c50d5`；它不是 v0.3.11 Release 的原始发布包，未重新发布或变更版本号。
+
+查询时官方 [npm](https://www.npmjs.com/package/@deepseek-ai/dsh) 的 `latest`、`next` 与 [GitHub 最近 Release](https://github.com/deepseek-ai/deepseek-harness/releases/tag/dsh-v0.2.2-rc.1) 均为 `0.2.2-rc.1`（2026-11-04 发布），`alpha` 为 `0.2.2-alpha.3`，没有查到更晚版本。rc.1 调整了 Loader、工具执行和会话投影接口，但仍保留插件使用的旧投影注册入口；在以下实测范围内无需修改插件运行代码。
+
+| 完整宿主 | 实际版本 | 结果 |
+|---|---|---|
+| 官方 Web | `0.2.2-rc.1` | 通过 |
+| 官方 Web · alpha 渠道 | `0.2.2-alpha.3` | 通过 |
+| 官方 Desktop · macOS arm64 | `0.2.2-rc.1` | 通过 |
+| DSH NEXT · macOS universal 包 | `2.0.22-next`，内置 `0.2.2-rc.1` | 通过 |
+
+- **完整链路**：同一测试包通过真实 Loader、Agent、工具、PTC 与会话存储，检查 standard/PTC/minimal/Super Code/改名副本、安装与重装备份、预算和删除保护设置、主子会话、压缩、旧记忆增改删与冲突/隔离、冷重启、插件禁用/恢复，以及物理卸载后历史保留和官方标准续聊。Super Code 的续聊提示只出现一次，minimal 不再带入已撤销的插件规则。oil-ui 通过真实 Skill 工具按需加载，并核对其正文、提供方、相对资源路径和已安装版本。
+- **MCP 边界**：各已通过的完整宿主另检查端口冲突、凭证轮换、目录授权，以及任务接纳、持久化失败、审批生命周期、目录变更和存储锁共 39 项边界；包含取消、断连、撤销、幂等与重启后不重放。
+- **前端与基础检查**：`0.1.5-alpha.1`、`0.2.2-rc.1`、`0.2.2-alpha.3` 的官方 React、控件和主题配合夹具服务，各通过 14 组交互检查，覆盖五个设置页、自动保存与失败、自定义下拉、中英文、主题、440px 设置布局、万条虚拟列表、圆圈动画、恰好一项展开和旧记忆 CRUD，无页面脚本错误。这不是三个完整宿主。560 项单元测试、类型检查和构建均通过。
+- **来源与版本**：官方 arm64 ZIP 匹配其 nightly 更新源 SHA-512，并通过原始应用的深度签名验证；实际 321 个 DSH 组件均为 rc.1。NEXT 下载频道指向 `2.0.22-next`，实际 330 个 DSH 组件均为 rc.1，原始应用深度签名验证通过；下载 API 未提供公开摘要，本地 SHA-256 只用于固定本轮样本。Web 两个依赖图各有 320 个 DSH 组件，均为目标版本。
+- **验证边界**：宿主使用本地确定性模型适配器，未调用外部模型，不是模型效果评测。本轮没有重跑全部历史完整宿主、最新 Windows/Linux 原生进程或原生目录选择器。rc.1 宿主本身要求 Node `>=24.9`，更新 Web 宿主时需满足；插件为旧宿主保留的 Node 兼容范围没有因此提高。
+
 ## 0.3.11 · 2026-11-06
 
 本版包含 11 月 4–6 日多轮修复。以下结果对应各轮实际测试的代码，不将早期开发包的完整宿主结果视为最终发布包的全部平台复验。

package/docs/EvaluationResults.md

@@ -1,10 +1,12 @@
 # 评测结果
 
+模型名称已脱敏，字母标签用于区分不同的评测模型。
+
 中文 · [English](en/EvaluationResults.md)
 
 ## Pro100：连续复用环境对比
 
-统一使用 Harness `0.2.1-rc.2`、`deepseek-v4.1-flash-1017`、high、旧预算，各组内部连续复用环境。以下为完整 100 题汇总，按通过数量降序排列。
+统一使用 Harness `0.2.1-rc.2`、`评测模型 B`、high、旧预算，各组内部连续复用环境。以下为完整 100 题汇总，按通过数量降序排列。
 
 | 方案 | 通过/有效 | 平均总 Token（M） | 平均耗时（分钟） | 平均模型调用 | 平均工具调用 | 缓存率 |
 |---|---:|---:|---:|---:|---:|---:|
@@ -38,8 +40,8 @@ SWE-bench Pro hard-100（固定标识 `swebench-pro-hard-100-ddd22492f1c38839`
 
 ### 历史评测过程
 
-1. **旧宿主基线**：在 DSH `0.1.5-rc.2` 与 `deepseek-v4.1-flash` 上评测官方各预设和插件各版本，Super Code `0.3.0` 候选通过 56 题，高于官方最好的标准 + high（51 题）。
-2. **宿主升级后效果下降**：升级至 DSH `0.2.1-rc.2` 与 `deepseek-v4.1-flash-1017` 后，默认预算下标准降至 46 题，Super Code `0.3.4` 降至 41 题，低于标准。
+1. **旧宿主基线**：在 DSH `0.1.5-rc.2` 与 `评测模型 B` 上评测官方各预设和插件各版本，Super Code `0.3.0` 候选通过 56 题，高于官方最好的标准 + high（51 题）。
+2. **宿主升级后效果下降**：升级至 DSH `0.2.1-rc.2` 与 `评测模型 B` 后，默认预算下标准降至 46 题，Super Code `0.3.4` 降至 41 题，低于标准。
 3. **定位原因**：宿主探针显示新版收紧了上下文预算（见[旧预算的内容](#旧预算的内容)），由此判断预算是退化原因之一。
 4. **恢复旧预算后复测**：双方均改用旧预算。标准仍为 46 题；Super Code `0.3.5` 提升至 49 题。据此，`0.3.6` 起插件默认使用旧预算。
 5. **跨模型验证**：在 `glm-5.3-flash` 上与原生 ZCode 对比，检验方案是否依赖特定模型。
@@ -56,7 +58,7 @@ SWE-bench Pro hard-100（固定标识 `swebench-pro-hard-100-ddd22492f1c38839`
 
 ## 新版 DSH 与预算
 
-DSH `0.2.1-rc.2` · 模型 `deepseek-v4.1-flash-1017`（调用标识 `deepseek/deepseek-flash`）· high。
+DSH `0.2.1-rc.2` · 模型 `评测模型 B`（调用标识 `deepseek/评测模型 A`）· high。
 
 | 指标（每题平均） | 标准 · 默认预算 | Super Code 0.3.4 · 默认预算 | 标准 · 旧预算 | Super Code 0.3.5 · 旧预算 |
 |---|---:|---:|---:|---:|
@@ -99,7 +101,7 @@ DSH `0.2.1-rc.2` · 模型 `deepseek-v4.1-flash-1017`（调用标识 `deepseek/d
 
 ## 官方预设
 
-DSH `0.1.5-rc.2` · `deepseek-v4.1-flash`。
+DSH `0.1.5-rc.2` · `评测模型 B`。
 
 | 指标（每题平均） | minimal + high | minimal + max | 标准 + high | 标准 + max | PTC + high | 创造 + high |
 |---|---:|---:|---:|---:|---:|---:|
@@ -126,7 +128,7 @@ DSH `0.1.5-rc.2` · `deepseek-v4.1-flash`。
 
 ## 历史版本
 
-DSH `0.1.5-rc.2` · `deepseek-v4.1-flash` · high。耗时采用历史原始口径（见[口径](#口径)），不能与上文净耗时直接比较。
+DSH `0.1.5-rc.2` · `评测模型 B` · high。耗时采用历史原始口径（见[口径](#口径)），不能与上文净耗时直接比较。
 
 | 指标（每题平均） | minimal + high | 0.0.9 | 0.1.2 | 0.2.0 分层记忆 | 0.3.0 Batch.3 |
 |---|---:|---:|---:|---:|---:|

package/docs/UsageGuide.md

@@ -22,6 +22,12 @@
 
 子 Agent 默认使用父 Agent 的模型。权限和工具由 Harness 管理。
 
+### 图片输入
+
+在 Harness「设置 → 模型」中，为确实支持识图的模型在「输入类型」勾选「图片」，然后在对话中选择该模型。可从输入框的「＋」添加图片，也可粘贴截图或拖入图片；发送前可预览、移除。Super Code 复用宿主的附件、缩放和发送流程。
+
+模型查看项目中的图片时使用原生 `read_image`，支持 PNG、JPEG、WebP、GIF。截图和视频帧必须实际读图，文本路径或文字摘要不代表完成视觉检查。若提示模型未声明图片输入，请先核对所选模型的能力配置；该提示不代表底层模型一定不支持图片。勿为纯文本模型勾选图片能力。
+
 ## 找不到预设
 
 请先确认插件已启用，再对照[兼容性](Compatibility.md)核实宿主版本，然后按宿主版本处理：
@@ -256,6 +262,8 @@ Windows 宿主默认在 Super Code 预设中使用内置 GNU Bash、BusyBox 和
 
 「允许调用方审批」默认关闭。开启后，指定 MCP 任务的一次工具审批可交给调用方，通过交互表单或 `dsh_approve` 决定；宿主权限策略和目录申请的用户确认保持有效。
 
+MCP 的任务和连接设置保存在 DSH 配置目录中。该目录应位于支持原子硬链接的本机文件系统（如 APFS、NTFS、ext4），不能跨机器共享；exFAT、FAT 不支持此锁机制，启动日志会提示迁移配置目录。此要求针对配置目录，允许访问的工作目录不受此文件系统限制。
+
 目录说明：
 
 - 每次进入连接页都会重新读取 DSH 工作区。首次配置时会预选已有工作区，之后新增的工作区**不会**自动获得授权。
@@ -327,17 +335,63 @@ Windows 宿主默认在 Super Code 预设中使用内置 GNU Bash、BusyBox 和
 - 上限：200 条任务记录、每个任务 128 次补充、最近 200 条事件、输出保留最后 12,000 字符。目录申请历史最多 100 条，仅保存在内存中。
 - 任务记录已满时会明确拒绝新任务，不会自动删除旧记录。
 
+## 预设 Skill 包
+
+在「设置 → Super Code 模式 → 预设 Skill」中，按 Skill 和工作流分组控制，默认全部关闭，已有的手动选择保持不变，无需逐项勾选工作流内的内容：
+
+| 分类 | 名称 | 用途 | 默认 |
+|---|---|---|---|
+| Skill | grill-me | 逐轮追问，梳理需求、检验方案与明确决策。 | 关闭 |
+| Skill | OpenSpec | 用规范组织需求、方案、任务、实现与验收。 | 关闭 |
+| Skill | oil-ui | 设计、改进与评审前端界面。 | 关闭 |
+| 工作流 | Matt Pocock Skills | 覆盖需求追问、规划、开发、测试、排错与代码评审的工程工作流。 | 关闭 |
+| 工作流 | PStack | 架构、开发、验证与评审的协作工作流。 | 关闭 |
+
+grill-me 独立开关复用 Matt Pocock 1.3.2 中的原始 `grill-me` 入口和配套 `grilling` 正文，无需开启全套工作流。开启 Matt Pocock Skills 后，其完整内容仍包含这两项；同时开启不会重复注册。
+
+Matt 包内置上游正式插件的完整发行内容（1.3.2，共27项），不按部分清单裁剪；OpenSpec 内置官方 skills.sh 静态发行的完整12项。PStack 是 Cursor 插件包，内含58个标准 `SKILL.md`（0.15.36），这里接入其全部 Skill、参考资料和辅助脚本，不注册 Cursor 插件、Agent 或自动化。各包的固定来源、文件清单和许可证见 `builtin/skills/` 中的 `UPSTREAM.md` 与 `distribution.json`。
+
+修改自动保存，保存成功后已有对话中的 Skill 目录、后续读取和原生斜杠菜单实时更新，无需刷新页面、新建对话或切换预设。关闭不会撤回已经加载进对话的指令；需要干净上下文时请新建对话。项目和用户自有 Skill 不受开关影响，同名优先级保持不变。保留上游调用规则：Matt 的部分 Skill 需要用户明确调用，不由模型自动选择。
+
+开启的包完整同步到 DSH 用户目录下的 `skills/dsh-super-code/`，保留参考资料和脚本的目录关系；使用自定义 `DSH_HOME` 时跟随该目录。页面上的「打开 Skill 目录」可在 DSH 所在设备的文件管理器中打开 `skills/`。无桌面环境时按钮不可用。关闭会移除未修改的插件副本；用户修改过的内容或已有同名目录会保留并提示，检查后可重试同步。插件卸载后这些嵌套资源不再注册到 Skill 目录。
+
+在使用 Super Code 预设的主对话输入 `/`，即可从 DSH 原生菜单选择 Skill；也可以输入 `/oil-ui`、`/grill-me`、`/openspec-propose` 或 `/poteto-mode` 等命令，再加任务内容。关闭的包不再提供命令和正文。输入 `/pstack`、`/mattpocock` 等包名可查找包内已启用的 Skill，选中后插入真实的 `/skill-name`。包名只用于菜单搜索，不能作为独立命令发送。插件通过 DSH 公开接口补充包名搜索并刷新原生菜单，不重启宿主，不修改草稿或模型调用。正在生成的模型请求及已经加载的上下文不会被撤回。
+
+保留原生同名优先级：项目/用户 Skill 优先；Matt 与 PStack 同时开启时，`/tdd`、`/teach` 使用先注册的 Matt 版本，关闭 Matt 后可使用 PStack 版本。
+
+PStack 使用针对 **DSH `0.2.1-alpha.1` 及之后版本**的转换发行版，保留58个入口和23个 playbook。转换层处理子 Agent 调度、模型选择、历史读取、Skill 路径与长任务检查点；`/setup-pstack` 只写项目内可选的 `.dsh/pstack.md`，`/make-bot-ui` 使用插件的 MCP 连接。默认继承宿主模型，依赖阶段由主 Agent 调度；缺少历史接口时使用明确提供的交接或项目资料，不读取宿主私有记录。原上游文件保留，便于升级核对，不改全局提示词。
+
+配置保留上游全部17类角色。推理预算、评审席位数与并发数分别设置；不支持逐次选择模型或思考强度时，明确继承宿主。reflect 保留三位评审与独立合成阶段。没有定时唤醒能力的宿主只能保存检查点，不能保证离开对话后继续定时运行。
+
+浏览器、Git/GitHub、模型检查器等流程仍需实际工具。辅助脚本可能需要 Bun、已登录的 gh、mise/Java/TLC 或 Pele，应按实际环境与用户授权使用；开启开关不运行脚本或安装工具。缺少能力会明确说明，不会把顺序自查当作独立 Agent 或跨模型验证。
+
+目录同步由插件统一管理，不需要重装预设。各预设仍须提供宿主原生 Skill 工具，才能在该对话中加载和调用；Super Code 预设已包含这些工具。
+
+每个预置包旁展示随插件内置的独立版本号。OpenSpec CLI 与 Skill 固定为匹配的版本，随插件安装准备好；启用后可在 Super Code 命令环境直接使用 `openspec`，无需全局安装。开关、列举和加载 Skill 均不运行命令，不改项目文件、依赖或配置；只有用户明确要求后才执行 `openspec init`。Matt 的工单流程可能需要已登录的 `gh`/`glab`，其 setup 也只在用户明确要求时运行。开关不会授予命令执行、提交、远程写入或付费服务权限；资源随插件更新。
+
 ## 前端 Skill
 
-插件内置 `super-code-frontend`，用于页面、组件、视觉调整、响应式布局、交互状态和浏览器验证。重点检查：
+插件内置 [oil-ui](https://github.com/oil-oil/oil-ui) **0.32.0**，替代原来的 `super-code-frontend`，用于页面、App、后台和组件的设计、改进与评审。主要包括：
 
-- 元素是否对齐到同一基准线
-- 可见范围与可点击范围是否一致
-- 控件位置是否稳定，操作距离和输入负担是否合理
+- 梳理用户任务、设计方向、视觉层级和关键交互
+- 按需制作多风格对比页，保留本地资源以供离线查看
+- 根据真实截图和交互结果评审、打磨页面
 
-它沿用项目自身的视觉风格，不限定框架，也不新增工具。平时只向模型提供名称和简介，任务匹配时才加载正文和相关参考。也可在请求中直接说明「用 super-code-frontend」。
+先在「预设 Skill」设置中开启 oil-ui，再通过 Harness 原生 Skill 机制按需加载，沿用项目的框架与实现约定。平时只向模型提供名称和简介，任务匹配时才加载正文和相关参考。也可在请求中直接说明「用 oil-ui」。
 
-- 0.3.3 起新安装或重装的预设才包含该 Skill。已有预设请先备份自定义内容，再重新安装。
+- 已安装的目录预设不会被自动覆盖。更新到包含 oil-ui 的插件后，请备份自定义内容，再在设置中重新安装预设；声明式预设直接读取当前插件包中的资源。
 - 项目或用户目录中存在同名 Skill 时，按宿主优先级覆盖内置版本。
-- 浏览器验证部分仅为方法指导，不强制防止循环，也不会自动写入长期记忆。
+- 内置副本随插件升级，不单独检查更新或运行上游安装器。来源、改动和许可证见 [来源说明](../builtin/skills/oil-ui/UPSTREAM.md)。
+- 设计指引不要求额外依赖。可选的风格对比生成器需要 Python 3.10+；截图和图片处理脚本需要 Node.js 22+ 及 Chrome、Chromium 或 Edge，部分素材操作还需要 ffmpeg/ffprobe。缺少这些环境时可使用宿主已有工具，并说明未完成的验证。
+- 可选素材检索会访问相应提供方；oiloil 生图需要账户和明确的积分消费授权，不会因插件启动而调用。上游 Pro 推荐脚本仅在使用 Skill 时按其说明写入一次本地提示标记。
 - 能否提升前端成品质量需通过单独的实际任务评测验证，安装测试和 SWE 成绩无法证明这一点。
+
+## 直接调用与按动作帮助
+
+两个状态工具均支持 `{"action":"help","forAction":"read"}`；一次帮助返回字段合同、完整示例、结果与恢复说明，不读取状态或重试持久化。示例来源、身份和版本必须替换为实际看到的值。
+
+`super_code_task {"action":"read"}` 读取当前聚焦任务；没有焦点时明确报错，不猜最新任务。更新仍必须提供已观察的 expectedRevision。纯输入格式在进入写屏障前校验，错误给出简短修复例子；不自动修改别名、类型、来源或 revision。
+
+`super_code_memory {"action":"read","topic":"build"}` 可在没有 task 时直接读取项目主题，不改变焦点和自动注入主题。省略 topic 仍使用原聚焦主题；显式读另一个主题不借用当前任务的代码版本。返回 details 保留 scope/topic，补 id 即可读取该主题的条目。topic 仅用于 project read，remember/forget 继续要求原聚焦主题、来源和并发检查。所有记忆仍是历史信息。
+
+纯格式拒绝不会追加本次业务变更；持久化失败可能已追加事件，应根据错误 list/read 确认，不能自动重放写入。旧记录与工具名不变。接口验证不代表模型效果或 Token 改善。

package/docs/en/Acknowledgements.md

@@ -31,23 +31,17 @@ Design reference only, not a dependency. The UI is built on Harness's public API
 
 The bundled frontend Skill follows this format: `SKILL.md`, YAML metadata, and a `references/` directory. The model sees only the name and summary until the Skill is needed, then loads the body and relevant references. It coexists with project and user Skills. Parsing and loading use Harness's native support.
 
-Implementation: [super-code-frontend](../../presets/super-code/skills/super-code-frontend/SKILL.md)
+Implementation: [oil-ui](../../builtin/skills/oil-ui/SKILL.md)
 
-## Apple Human Interface Guidelines
+## oil-ui
 
-[Layout](https://developer.apple.com/design/human-interface-guidelines/layout) · [Accessibility](https://developer.apple.com/design/human-interface-guidelines/accessibility) · [Buttons](https://developer.apple.com/design/human-interface-guidelines/buttons) · [Feedback](https://developer.apple.com/design/human-interface-guidelines/feedback) · [Terms](https://developer.apple.com/support/terms/)
+[Source](https://github.com/oil-oil/oil-ui) · [MIT](../../builtin/skills/oil-ui/LICENSE) · [Pinned version and adaptations](../../builtin/skills/oil-ui/UPSTREAM.md)
 
-The frontend Skill adopts these principles: alignment expresses relationships, layouts adapt to containers and type size, control size and spacing reduce effort, interactions are predictable, feedback is close and proportionate. They're organized into three references: visual design, interaction states, and browser validation.
+Thanks to oil-oil for this open-source frontend design Skill. The plugin bundles oil-ui 0.32.0's workflow, references, comparison page, and helper scripts, loaded on demand through Harness's native Skill mechanism in place of the previous custom frontend Skill.
 
-Only principles are borrowed; each product keeps its own visual style. Native platform units aren't treated as CSS pixels, and materials, fonts, and icons aren't used as a template.
+The bundled copy is pinned to upstream commit `0d64a00f3a5c45bf7a30d35bded9b7c0fdff2128`. Its only adaptation replaces the standalone update entrypoint with plugin-managed updates; neither the upstream installer nor a paid plugin is included. The MIT license and original copyright notice are preserved in the package.
 
-## W3C WCAG 2.2
-
-[WCAG 2.2](https://www.w3.org/TR/WCAG22/) · [2.5.8 Target Size (Minimum)](https://www.w3.org/WAI/WCAG22/Understanding/target-size-minimum.html)
-
-When checking Web controls, the frontend Skill uses 2.5.8's 24 × 24 CSS pixel criterion and its spacing exceptions, alongside real click, keyboard, and content checks. Passing the numbers does not make the whole interface accessible.
-
-Implementation: [interaction states](../../presets/super-code/skills/super-code-frontend/references/interaction-states.md), [browser validation](../../presets/super-code/skills/super-code-frontend/references/browser-validation.md)
+The comparison page also bundles [Instrument Serif](https://github.com/Instrument/instrument-serif), under [SIL OFL 1.1](../../builtin/skills/oil-ui/assets/fonts/instrument-serif/OFL.txt), for comparison-page typography.
 
 ## SWE-bench
 
@@ -74,3 +68,21 @@ Full list in [package.json](../../package.json); locked versions in [package-loc
 ## License
 
 This project uses [Apache-2.0](../../LICENSE). The projects, specifications, design resources, and datasets above keep their own copyrights and licenses as stated upstream (LICENSE, NOTICE, and dataset docs). A design reference does not imply a dependency, endorsement, or feature equivalence.
+
+## OpenSpec
+
+[Source](https://github.com/Fission-AI/OpenSpec) · [MIT](../../builtin/skills/openspec/LICENSE) · [Upstream](../../builtin/skills/openspec/UPSTREAM.md)
+
+Thanks to the OpenSpec contributors for the complete specification-driven workflow Skill distribution. The plugin preserves the original prompts and licenses and loads them on demand through DSH.
+
+## Matt Pocock Skills
+
+[Source](https://github.com/mattpocock/skills) · [MIT](../../builtin/skills/mattpocock/LICENSE) · [Upstream](../../builtin/skills/mattpocock/UPSTREAM.md)
+
+Thanks to Matt Pocock and contributors for the complete engineering and productivity Skill distribution. The plugin includes the official plugin inventory with its references, templates and invocation rules; it does not run setup on activation.
+
+## PStack
+
+[Source](https://github.com/cursor/plugins/tree/main/pstack) · [MIT](../../builtin/skills/pstack/LICENSE) · [Upstream](../../builtin/skills/pstack/UPSTREAM.md)
+
+Thanks to Lauren Tan and the Cursor plugin contributors for PStack. The plugin preserves its 58 standard Skills, invocation rules, references and scripts, loaded on demand through native DSH Skills and slash commands. It does not register Cursor automations or run setup on activation. Some workflows still require Cursor capabilities or external tools; see the upstream notes.

package/docs/en/Changelog.md

@@ -4,6 +4,33 @@
 
 Versions are `major.performance.maintenance`. The second number changes when coding results or efficiency improve; the third covers fixes, UI, and compatibility. See [Compatibility](Compatibility.md) for what each release was tested on.
 
+## 0.4.0 · 2026-11-12
+
+**Task and memory tools**
+
+- Simplified calls to `super_code_task` and `super_code_memory`, with action-specific field guidance, examples, and recovery instructions to reduce trial and error.
+- Read a named project-memory topic without creating or switching the focused task. Writes retain source checks and revision-based conflict protection; historical information does not replace checking current code.
+
+**Preset Skills and workflows**
+
+- Added standalone grill-me alongside OpenSpec, oil-ui, Matt Pocock Skills, and PStack, grouped into Skills and workflows. All five default off, and saved choices are preserved.
+- Enabled packages sync their complete resources to the DSH Skill directory and take effect in existing conversations. Each entry shows its version and a short usage hint, with a shortcut to open the directory. Existing user files and edits are protected; conflicts are reported and can be retried.
+- Uses the native `/` menu and real Skill names. Search by package names such as `/pstack` or `/mattpocock`, then select a specific command. Disabled packages no longer supply commands or bodies; content already loaded into context is not withdrawn.
+- OpenSpec's command environment ships with the plugin. Toggling a package does not initialize a project, install project dependencies, or modify project files; initialization runs only when requested.
+- PStack retains upstream workflows and resources, with adaptations for DSH tools, child Agents, approvals, and image handling. Workflow adaptation targets Harness `0.2.1-alpha.1` and later.
+- Bundled and adapted sources are published under `builtin/skills`, with upstream versions, provenance, and licenses.
+
+**UI, compatibility, and fixes**
+
+- Improved Settings autosave, draft retention across tabs, reconnection, concurrent edits, and recovery after failed saves.
+- Improved task, execution, tool-status, and changed-file displays while retaining circular Agent nodes and the existing memory page.
+- Fixed cancellation, reconnection, and repeated enable/disable issues in preset installation and uninstall recovery, MCP connections and approvals, deletion protection, and the Windows command environment.
+- Updated compatibility records for newer Harness versions. See [Compatibility](Compatibility.md) for version and platform coverage, distinguishing SDK checks from full-host acceptance; interface tests do not establish that every workflow and platform has passed.
+
+- Added version, license, and total npm download badges, with Star and npm trend charts refreshed daily.
+
+See the [README](../../README.en.md) and [evaluation notes](EvaluationResults.md) for results and statistical limits. Skill installation, UI, and compatibility tests are not evidence of coding performance.
+
 ## 0.3.11 · 2026-11-06
 
 **Uninstall recovery and preset management**

package/docs/en/Compatibility.md

@@ -2,11 +2,13 @@
 
 [中文](../Compatibility.md) · English
 
-Web is verified from Harness `0.1.5-alpha.1` through `0.2.1-rc.4`, with incremental rc.4 coverage detailed below. Official Desktop is verified through `0.2.1-rc.4` (x64 archive under Rosetta), and NEXT at `2.0.19-next` (bundled `0.2.1-rc.2`).
+Supports Harness `0.1.5-alpha.1` through `0.2.2-rc.2`, including `0.2.2-alpha.3`. Version 0.4.0 has been checked against all 27 complete Web hosts in this range, official macOS arm64 Desktop `0.2.2-rc.2`, and NEXT `2.0.22-next` (bundled `0.2.2-rc.1`). This round's scope and limits are recorded below.
 
-The bundled Windows Bash in 0.3.10 also passes full-process Windows acceptance on official Desktop `0.2.1-rc.5` and NEXT `2.0.21-next` (bundled `0.2.1-rc.4`), within the scope below.
+Full-process Windows coverage currently reaches official Desktop `0.2.1-rc.5` and NEXT `2.0.21-next` (bundled `0.2.1-rc.4`). New macOS/Web results do not extend that Windows scope.
 
-The 0.3.11 fixes add staged version-contract and official-frontend checks through `0.2.2-alpha.3`. Full-host coverage is recorded separately in the 0.3.11 section below.
+The records below distinguish each code baseline and test layer. Component checks, static-frontend fixtures, and earlier development-package results do not establish a complete rerun of every version and platform.
+
+The new PStack workflow adaptation starts at `0.2.1-alpha.1`; earlier hosts are outside its scope. Older-version records below describe historical coverage of other plugin features.
 
 ## Verified hosts
 
@@ -55,6 +57,67 @@ Under npm rules, a range like `>=0.1.5-alpha.1 <0.3.0` excludes prereleases such
 
 `zod`, `js-yaml`, and `@modelcontextprotocol/sdk` are runtime dependencies and are not bundled. The Windows shell also packages separately licensed programs that run as subprocesses; see [runtime notes](../../runtime/windows/README.md) for sources, licenses, and corresponding source.
 
+## November 12, 2026 · 0.4.0 compatibility rerun
+
+The existing peer dependency ranges and minimum `0.1.5-alpha.1` remain unchanged. Every version uses the same 0.4.0 candidate; no plugin runtime changes were needed for these checks.
+
+| Layer | Scope | Result |
+|---|---|---|
+| Complete Web hosts | All 27 exact published versions, `0.1.5-alpha.1` through `0.2.2-rc.2` | 12 groups per version; 324 pass |
+| Actual SDK and disk storage | The same 27 versions | Six groups per version; 162 pass |
+| Native Skill loading | The same 27 versions | 24 checks and 98 body loads per version pass; duplicate names follow provider precedence |
+| Native menu contract | The same 27 versions | Pass; `0.1.5` lacks package-query expansion, but individual Skill names remain selectable |
+| Browser interactions | Official React, controls and themes from all 27 versions, with fixture Remote services | 15 groups per version; 405 pass without page errors |
+| Official macOS arm64 Desktop | `0.2.2-rc.2` | 13 complete-host groups and four actual UI checks pass |
+| DSH NEXT · macOS arm64 | `2.0.22-next`, bundled `0.2.2-rc.1` | 13 complete-host groups and four actual UI checks pass |
+
+- **Complete-host checks**: install the candidate, verify preset versions, independent budget/protection saves, five default-off Skill packages enabling and disabling in existing conversations, default MCP settings, manual memory CRUD and concurrency conflicts, native/PTC task and memory calls, complete process restarts, original-conversation continuation and memory recovery.
+- **UI checks**: six settings tabs, Skill versions and switches, failed saves, budget validation, connection settings, Chinese/English, light/dark themes, narrow layouts, 10,000-row task virtualization, execution-tree animation and legacy memory maintenance. Both actual desktops additionally verify settings loading, matching versions, five default-off packages, switch autosave and disabled Windows command-environment controls on macOS.
+- **Provenance and baseline**: the official archive matches the feed SHA-512 and passes deep signature verification; all 333 DSH components are rc.2. NEXT has a valid signature and all 330 DSH components are rc.1. All 1154 unit tests, type checking and build/pack pass. Acceptance scripts were corrected for concurrent completion order, upstream theme packaging and actual page structure; original failed records remain preserved.
+- **Limits**: complete hosts use a local deterministic model adapter with zero external-model calls. This does not establish model effectiveness, acceptance of all 58 PStack workflows or every feature branch. This round does not repeat Windows, Linux, every historical native desktop, native directory pickers or physical-uninstall recovery; those retain their historical evidence identities. Full PStack workflow adaptation still starts at `0.2.1-alpha.1`.
+
+## November 11, 2026 · PStack DSH adaptation
+
+- Batch six removes adapter-added command bans, mandatory API checks and repeated snapshot audits. Four role templates now match upstream exactly, and delegation translation preserves the upstream read-only requirement. Read-only roles may use commands without side effects, but may not write files or change external state. Command capture also fixes sparse-argv validation and post-submission argument mutation. All 925 regression tests and 143 resource checks across 13 versions pass; real minimum/latest Web capture scenarios each pass nine independent checks. Live workflow children still redirect output into temporary files, so read-only scope adherence remains unresolved.
+- The scope contains all 13 npm releases from `0.2.1-alpha.1` through the current latest `0.2.2-rc.1`, including `0.2.2-alpha.3`. Native catalogs, all 58 bodies, slash injection, invocation permissions and same-name precedence pass, with batch five adding command-capture resources for 143 checks; task enum diagnostics pass another 39 checks. Native subagent contracts pass 65 checks; live switches in existing browser conversations pass 169. Browser checks use real frontend modules with fixtures at service boundaries, not 13 complete desktop processes.
+- On November 11, 2026, all 58 entrypoints ran through native slash injection and real model calls in isolated macOS arm64 Web `0.2.2-rc.1`, using `tokenhub/deepseek/Evaluation model A`. Each entrypoint had an independent scenario. Twelve scenarios were rerun, and two received clarifications; repeats are not added to the 58-entry denominator.
+- First-round acceptance: 36 pass, 17 pass with qualifications, four fail and one partial. Read-only children in `how`, `teach` and `create-formal-verification-skill` still wrote global temporary files. `principle-outcome-oriented-execution` retained `add` twice and passed after the target was clarified as `sum({left,right})`; ambiguity between the original wording and the acceptance contract needs separate review. `ketchup-video` encoded and decoded a real video, but the test route omitted the image-input declaration and the workflow did not complete its visual review. Execution completion is not acceptance.
+- Image diagnosis corrected: the model does support images. With `input: ["text", "image"]` declared in the isolated route, direct image attachments and native `read_image` on both `0.2.1-alpha.1` and `0.2.2-rc.1` let the live model read a random code and count shapes present only in the pixels. Provider requests contained image content. A text-only configuration rejected images before sending. The Chinese rc.1 Web UI also passed native selection, preview, removal, paste, drop and history display; this round did not exercise native desktop image interactions across platforms.
+- All 22 entries without full acceptance were rerun. The video workflow completed 14 actual image reads, including decoded frames from every slide; eight CLI samples and four decoded images were independently checked. The complete `statechart` workflow also passed real TLC and code replay. Independent baseline checks passed 56/57; the remaining check compares an ambiguously worded API task against a fixed contract and needs separate interpretation. Five additional checks passed. Read-only children still wrote global temporary files, and cleanup used `rm` and process-name matching to stop encoders. Image transport and baseline checks do not erase these defects, so all 58 workflows cannot be reported as fully accepted.
+- Three read-only/audit workflows were tested again, passing seven baseline checks. An independent copy of the formal-verification artifact passed three additional normal, invalid-argument and mutation/restoration checks. The audit reviewer made no writes this time, and its independent review passed with the snapshot coverage limits disclosed. The `teach` investigator and synthesizer still wrote global temporary files. Prompt rules therefore do not provide a hard read-only boundary, and these issues are not all resolved.
+- This round corrected three global temporary-path examples and child-scope propagation guidance. It also fixed the installed `statechart` helper: Node refuses TypeScript stripping under `node_modules`, so packaging now emits JavaScript while retaining upstream sources. The actual installed wrapper passed TLC, a concrete authorization counterexample and code replay under a Unicode path with spaces. The MCP bot UI exercised real new/reused conversations, stable request IDs, restart, rejection branches, paged history and session termination; live approval waiting and remote deployment were not exercised.
+- The distribution retains 58 entrypoints, 23 playbooks and 558 originals, with 592 generated files, including audit snapshots, stable submission admission and project-local command capture. Original hashes, generated bytes and helper permissions are checked. All 919 batch-five tests, type checking and build pass; the earlier 909-test batch and 19 focused follow-up checks remain recorded. The initial 58-scenario independent checks total 155: 154 pass and the API migration assertion fails. These checks do not replace phase and scope review.
+- A third batch reruns eight priority entrypoints with 20 passing baseline assertions. The converter now preserves why's code-anchor step, and source-only contracts are inside the actual role templates. Three video slides and two transitions were independently decoded and viewed; 16 checks pass, without permanent shell cleanup or shared-name process termination. Eight snapshot-helper checks cover append/tamper/CRLF/Unicode behavior; the complete 23-row trail receives 12 provenance/behavior checks plus per-row semantic review. Structural validity is not claim verification.
+- Live MCP approval and cancellation now pass six checks; the browser observes `waiting_confirmation → cancelling → cancelled`, and the pending probe never executes. Acceptance found a changed-content cache bug in the generated example backend. After a manual repair, 13 checks pass; the product workflow now specifies fingerprints, serialized admission and original-input retries. This is neither a newly found core-service defect nor an autonomous first-pass success. Remote deployment and live injection of every error status remain untested.
+- Minimum `0.2.1-alpha.1` reruns a seven-file how scenario with the final runtime candidate: 14 independent checks pass, native command results are used directly, the source/tree stays unchanged, and two children use only read/glob/grep/send_message. Model adherence limits remain: teach/show use read-only cat/ls, and the minimum run repeats an explainer and attempts a capture write that fails with ENOENT; its handoff also undercounts children. The final package then reruns managed and synchronous return branches on the minimum host: 15 checks pass, with one delegate per case, no capture/write attempt or redispatch, and same-call final-text extraction for the synchronous result. Earlier duplicate/undercount evidence remains; the synchronous reply still exceeds its requested length. Neither hard isolation nor universal prompt compliance is established.
+- Batch four adds a dependency-free stable submission helper to make-bot-ui, covering concurrent IDs, content conflicts, restarts, persistence failures and independent return values. The generated example completes 92 self-checks. Acceptance then finds interrupted incorrectly treated as settled and failed transports closed without DELETE. After diagnostic continuation in the original model session, 91 deterministic fixture checks, 13 independent UI checks and two recovery checks pass. The session then reaches the acceptance harness's cumulative 100-request limit without its final handoff, so this is not an autonomous full-pass claim. The final packaged helper separately passes 18 independent checks with real MCP/model calls: new and explicitly reused conversations, per-request prior/current results, refusals and session release. Fixture states are not real network-failure injection.
+- The fourth audit covers all work/correction rows 1–28; final rows 29–30 contain review metadata only. Twenty-five snapshot, actual CLI and provenance checks pass. Replaced evidence, incorrect counts and causal descriptions remain disclosed; the reviewer verdict is PASS WITH FLAGS. Teach's four children make no writes but one still uses ls under a narrower no-shell contract. Bot development writes four global temporary responses and issues a cleanup command containing an rm fallback whose actual branch is unknown. A narrow live scenario after project-local evidence guidance gets all three HTTP refusals right but still writes a global status-probe file; its scope check remains failed. Functional checks do not establish universal scope adherence.
+- Batch five retains unique project-local capture directories: all 34 audit captures re-verify and 14 independent provenance/CLI checks pass. The original session appends corrections for mistyped hashes and row counts; its reviewer still executes a read-only shell command, so the existing no-shell contract remains failed. The repaired MCP example checkpoint completes 96 real integration checks, 91 explicitly simulated fault checks and its final handoff, without global temporary responses or overwritten command evidence. Nineteen of 20 independent checks pass; the sole failure is that role contract. A separate real minimum-host how/capture scenario passes 16 checks for Unicode paths with spaces, literal argv, exit 0/7, exact streams, byte verification, unchanged source and a read/search-only child. These targeted checks do not establish full acceptance of all 58 workflows.
+- The 58-scenario run covers only that macOS Web host and model, not Windows/Linux, every desktop release, all 23 playbook branches, cross-model evaluation, GitHub publication or the optional Bun watcher. Earlier minimum-host configuration/delegation evidence is retained separately; the new how run does not cover every minimum-host workflow. No effectiveness or performance improvement is established.
+
+- The completeness review restores setup’s eight stages and 17 role categories, independent reflect synthesis, incremental history, simulator/cache branches, and cause-based retries/late-result handling. Reasoning budget, candidate count and concurrency remain distinct. Missing public model discovery means inheritance, not scanning private configuration. Long runs assume neither a scheduler nor a private agent store.
+
+- Follow-up live-model checks: latest-host setup preserves the full role map and public capability-discovery boundary; reflect runs three independent reviews plus a fourth synthesizer, routes structural proposals to Backlog and makes no Skill edits before approval. Minimum `0.2.1-alpha.1` rechecks the new setup’s 17 roles, separate budget/panel/concurrency values and existing notes. Initial discovery/classification failures and acceptance-script format assumptions remain recorded; a completion marker alone is not acceptance evidence.
+
+## November 9, 2026 · Latest-host rerun
+
+This run uses `main@8d28e47`, including the oil-ui `0.32.0` replacement. The fixed test package has SHA-256 `3b19a6c04b960202d87e13bb17e59085f7378503aaf8316a9c0d78220f6c50d5`. It is not the original v0.3.11 Release archive; no version bump or republication was performed.
+
+At checking time, official [npm](https://www.npmjs.com/package/@deepseek-ai/dsh) `latest`/`next` and the [most recent GitHub Release](https://github.com/deepseek-ai/deepseek-harness/releases/tag/dsh-v0.2.2-rc.1) are all `0.2.2-rc.1`, released November 4, 2026. The alpha tag is `0.2.2-alpha.3`; no later release was found. rc.1 changes Loader, tool execution, and session projections but retains the legacy projection registration used by the plugin. No plugin runtime change is required within the tested scope below.
+
+| Complete host | Actual version | Result |
+|---|---|---|
+| Official Web | `0.2.2-rc.1` | Pass |
+| Official Web · alpha channel | `0.2.2-alpha.3` | Pass |
+| Official Desktop · macOS arm64 | `0.2.2-rc.1` | Pass |
+| DSH NEXT · macOS universal archive | `2.0.22-next`, bundled `0.2.2-rc.1` | Pass |
+
+- **Complete flow**: the same package runs through actual Loader, Agent, tools, PTC, and session storage. Checks cover standard/PTC/minimal/Super Code/renamed-copy presets, installation and reinstall backups, budgets and deletion settings, parent/child histories, compaction, legacy memory CRUD/conflicts/scoping, cold restarts, plugin disable/re-enable, and physical removal followed by preserved history and official-standard continuation. Super Code shows one recovery notice; minimal excludes revoked plugin rules. The real deferred Skill tool loads oil-ui, verifying its body, provider, relative resources, and installed version.
+- **MCP boundaries**: each passing full host additionally checks port conflicts, token rotation, and directory consent, plus 39 checks across admission, persistence failures, approval lifecycle, directory changes, and storage locks. Coverage includes cancellation, disconnect, revocation, idempotency, and no replay after restart.
+- **Frontend and baseline**: official React, controls, and themes from `0.1.5-alpha.1`, `0.2.2-rc.1`, and `0.2.2-alpha.3` each pass 14 interaction groups with fixture services. These cover five settings tabs, autosave/failures, custom dropdowns, Chinese/English, themes, 440px settings layouts, 10,000-row virtualization, circle animation, exactly one open section, and legacy memory CRUD, without page errors. These are not three complete hosts. All 560 unit tests, type checking, and build pass.
+- **Provenance and versions**: the official arm64 ZIP matches its nightly feed SHA-512 and the original application passes deep signature verification; all 321 DSH components are rc.1. The NEXT download channel points to `2.0.22-next`, all 330 DSH components are rc.1, and the original application passes deep signature verification. Its API supplies no public digest; local SHA-256 only pins the tested sample. Both Web graphs contain 320 DSH components, all at their target version.
+- **Limits**: full hosts use a local deterministic model adapter, without external-model calls or effectiveness evaluation. This run does not repeat every historical complete host, latest native Windows/Linux processes, or native directory picking. The rc.1 host requires Node `>=24.9` when upgrading Web; the plugin retains its Node range for older hosts.
+
 ## 0.3.11 · 2026-11-06
 
 This release includes successive fixes from November 4–6. Results below apply to the code tested at each stage; full-host results from an earlier development package are not presented as a complete platform rerun of the final release package.