DSH Session Insights 0.5.2

0.5.2 continues the 0.5.x repair series for DSH RC.2. It prevents resume from selecting a different project or privacy scope, bounds analysis resources and fixes single-card report widths. It requires DSH >=0.2.0-rc.2; RC.1 is no longer supported. Start a new analysis run after upgrading from 0.5.1. Existing reports remain on disk; separately installed Python CLI and Skill copies require their own update.

• Resume checks the recorded project, window, privacy, depth and locale, and refuses conflicting or ambiguous requests.
• Each plugin instance permits two active analyses and eight cancellable queued requests. Worker memory limits do not cap host RSS.
• The optional Python reader checks decoding budgets incrementally and detects truncated final Zstandard frames, with explicit diagnostics instead of fallback to older logs.
• RC.2 asynchronous question replies and scheduled prompts remain injected context, excluded from human-work and acceptance inference.
• Completed --no-open output is a bare report path. Reports remain self-contained HTML with companion JSON; single narrative cards fill their row and narrow-screen actions wrap.

The canonical package is bound to commit ba38259075527a2efd6de2ee9bf4a3d6420b3388, SHA-256 b85630bc46516a4ad15347089dcb296f2270541c4075ec040a2704f4e3e04e1f (148198 bytes, 25 files). All 13 producer-CI jobs passed, including strict lockfile installation without a legacy peer flag.

Native installation, no-op, reinstall, loading and populated bilingual report-version checks are recorded separately per platform and host. Unchanged runtime, GUI, model and page behavior reuses qualified candidate evidence with explicit input hashes. The real-provider positive case was schema-guided; the real Headless negative case deliberately submitted invalid output and one still-invalid repair before a marked fallback. HTTP/static observations and the user's offline local-file observations remain separate. Declared admission of future DSH versions is not future-host qualification. Attachments record the precise subjects and remaining limits; no daily consumer deployment is included.

Full changelog

中文

0.5.2 延续 0.5.x 的修复系列，适配 DSH RC.2，避免恢复选中不同项目或隐私范围，限制分析资源消耗，并修复单张报表卡片的宽度。要求 DSH >=0.2.0-rc.2，不再支持 RC.1。从 0.5.1 升级后请新建分析运行；已有报告保留。单独安装的 Python CLI 或 Skill 需要分别更新。

• 恢复会核验记录的项目、时间窗、隐私、深度和语言，拒绝矛盾请求或多运行歧义。
• 每个插件实例最多同时分析两个运行，另有八个可取消的排队名额；Worker 内存限制不等于宿主 RSS 上限。
• 可选 Python 读取器在增量解码时检查预算，并检测被截断的 Zstandard 末帧；拒绝的文件会给出诊断，不回退旧日志。
• RC.2 异步问答回复与定时消息仍作为注入上下文，不计入人工工作或验收推断。
• --no-open 的已完成结果只返回报表路径；报告仍为自包含 HTML 和配套 JSON。单张叙述卡片铺满整行，窄屏操作按钮自动换行。

唯一制品绑定提交 ba38259075527a2efd6de2ee9bf4a3d6420b3388，SHA-256 为 b85630bc46516a4ad15347089dcb296f2270541c4075ec040a2704f4e3e04e1f（148198 字节、25 文件）。13 项 producer CI 全部通过，其中严格 lockfile 安装无需 legacy peer 参数。

原生安装、严格幂等、重装、加载及双语非空报告版本检查按平台和宿主分别记录。未变化的运行逻辑、GUI、模型和页面行为，按输入摘要明确复用已资格化候选的证据。真实 provider 正向用例使用 schema 指导；真实 Headless 负向用例故意提交无效内容，并做一次仍无效的修复，随后生成标记为降级的报告。HTTP/静态观察与用户的离线文件观察分别保留。声明范围准入未来 DSH 版本，不代表未来宿主已实测。附件记录精确对象和残余限制；本次不包含日常消费者部署。

完整更新日志

DSH Session Insights 0.5.1

DSH Session Insights 0.5.1

中文

本版适配 DeepSeek Harness 0.2.0-rc.1。工具调用被中断后，宿主补写的恢复结果会单独计数，避免把没有最终结果的调用误算为已确认失败。

变化

• JSON 总量和会话摘要新增 tool_recovery: {outcome_unknown, not_started}。两种恢复状态都不计入工具失败、验证失败、权限或重试计数；同一调用 ID 在不同轮次的证据分别保留。
• TOOL_OUTCOME_UNKNOWN 表示日志记录了调用，但未记录最终结果；TOOL_NOT_STARTED 表示日志没有该调用开始的记录。它们都不能证明工具执行过或没有执行。重试可能改变数据的操作前，应先核实外部状态。

升级

只支持 DSH 0.2.0-rc.1，Node.js 要求 ^22.19.0 || >=24.0.0。Bundle 无需 Python；单独安装的 Python CLI/Skill 需另行更新到 0.5.1。升级后请新建分析运行，旧 manifest 和缓存不能继续使用；已有 HTML、JSON 和 Markdown 报告保留。旧原始日志需先由上游 DSH 迁移为 V4。

dsh plugin --profile web add dsh-session-insights@0.5.1

源码检查、安装包和实际宿主的验证范围分别记录，详见随 Release 提供的 release-acceptance.md。冻结候选记录保留发布前的状态。

完整更新日志 · 安装与使用

English

This patch targets DeepSeek Harness 0.2.0-rc.1. Recovery results written after an interrupted tool call are counted separately, so a missing final result does not become a confirmed failure.

Changes

• JSON totals and session summaries add tool_recovery: {outcome_unknown, not_started}. Neither recovery status increases tool-failure, failed-verification, permission or retry counts. Reused call IDs retain separate evidence across turns.
• TOOL_OUTCOME_UNKNOWN means the log recorded the call but has no final result. TOOL_NOT_STARTED means it has no record of that call starting. Neither proves whether the tool executed. Check external state before retrying an operation that can change data.

Upgrade

Only DSH 0.2.0-rc.1 is supported. Node.js must satisfy ^22.19.0 || >=24.0.0. The Bundle needs no Python; update a separately installed Python CLI/Skill to 0.5.1 independently. Start a new analysis run after upgrading: earlier manifests and caches cannot resume. Existing HTML, JSON and Markdown reports are retained. Migrate older raw logs to V4 through upstream DSH first.

dsh plugin --profile web add dsh-session-insights@0.5.1

Source checks, package checks and actual-host results are recorded separately in the attached release-acceptance.md. The frozen candidate record preserves the pre-publication state.

Full changelog · Installation and usage

DSH Session Insights 0.5.0

DSH Session Insights 0.5.0

中文

本版适配 DeepSeek Harness 0.1.7-rc.2，把新会话格式、程序化工具调用和审批结果纳入复盘。

变化

• Bundle 和可选 Python CLI 只读取 V4 会话。旧原始日志需先通过上游 DSH 迁移；升级后请新建分析运行，已有报告保留。
• 记录 PTC 内层工具执行，并分别展示内外层失败。权限拒绝不再误计为验证命令执行失败，工具注册通知和定时注入不计作人工请求。
• 使用官方消息构造与可取消的会话观察接口；读取完成后释放观察资源，卸载插件时等待自身任务结束。

升级

只支持 DSH 0.1.7-rc.2，Node.js 要求为 ^22.19.0 || >=24.0.0。Bundle 无需 Python；单独安装的 Python CLI/Skill 需另行更新至 0.5.0。

70 项 Node、70 项 Python 测试及 13 个 CI 任务通过；同一安装包已在 macOS rc.2 宿主完成非空真实模型语义流程和 notice 持久化读回。Windows rc.2 隔离宿主已通过安装、注册、中文空格路径筛选、非空确定性报告及取消后重跑检查；中英文离线报告的显示、筛选重置、导出和控制台检查由维护者人工验收通过。Windows 未重复模型语义测试；活跃 Agent 自动工具增删和完整进程重启恢复仍未原生验证。最终验收结果见随 Release 附带的 release-acceptance.md；源码验收记录保留冻结候选时的状态。

完整更新日志 · 安装与使用

English

This version adapts session retrospectives to DeepSeek Harness 0.1.7-rc.2, including its current session format, programmatic tool calls and approval outcomes.

Changes

• Read V4 sessions in the Bundle and optional Python CLI. Migrate older raw logs through upstream DSH and start a new analysis run after upgrading. Existing reports are retained.
• Include PTC inner executions and show inner and outer failures separately. Permission denials do not count as failed verification commands; tool-registration notices and scheduled injections do not count as human requests.
• Use the official message helper and cancellable session observations. Release observations after reading and wait for owned tasks when the plugin unloads.

Upgrade

DSH 0.1.7-rc.2 is the only supported host version. Node.js must satisfy ^22.19.0 || >=24.0.0. The Bundle needs no Python; update a separately installed Python CLI/Skill to 0.5.0 independently.

All 70 Node tests, 70 Python tests and 13 CI jobs passed. The same tarball completed a nonempty real-model semantic workflow and persisted-notice readback in the macOS rc.2 host. An isolated Windows rc.2 host passed installation, registration, Unicode-and-space path filtering, nonempty deterministic reporting and cancellation followed by a successful rerun. The maintainer manually verified display, filter/reset, exports and browser-console behavior in both offline report languages. Windows model semantics were not retested; active-Agent automatic tool updates and recovery after a full process restart were not tested natively. See the attached release-acceptance.md for final acceptance. The source acceptance record preserves the earlier candidate-freeze status.

Full changelog · Installation and usage

DSH Session Insights 0.4.0

DSH Session Insights 0.4.0

中文

原生 Bundle 现在直接用 Node.js 生成复盘报告，无需再安装 Python。

变化

• 移除原生 Python 子进程及其环境变量、模块搜索路径继承。
• 收紧报告目录和文件访问，拒绝符号链接、硬链接及越界路径；模型输出通过校验后才写入。
• 新增运行产物清理工具，先预览再确认删除。不再写入跨运行共享语义缓存。
• 修复工具明确失败时，被输出中的“成功”文字覆盖的问题。

升级

需要 Node.js 20+（或 DSH 要求的更高版本），DSH 依赖仍为 0.1.5-rc.2。旧 Python 运行记录请通过原 CLI 完成，或重新生成报告。可选 Python CLI 继续保留。

65 项 Python 测试、42 项 Node 测试及 macOS、Linux、Windows CI 已通过，打包后检查覆盖中英文报告。真实 DSH 宿主和模型往返验收待完成。

完整更新日志 · 安装与使用

English

The native Bundle now generates session reports directly in Node.js, without a Python installation.

Changes

• Remove native Python subprocesses and inherited environment and module search paths.
• Restrict report storage, reject symlinks, hard links and escaping paths, and validate model output before writing it.
• Add run cleanup with a preview and explicit confirmation. Native runs no longer write a shared semantic cache.
• Preserve explicit tool failures when output text contains success-looking messages.

Upgrade

Node.js 20+ is required, or the higher version required by DSH. The exact DSH dependency remains 0.1.5-rc.2. Finish legacy Python runs through the CLI or generate a new report. The optional Python CLI remains available.

All 65 Python tests, 42 Node tests and macOS, Linux and Windows CI checks passed. Packaged-report checks cover both languages. Real DSH host and model round-trip acceptance remains pending.

Full changelog · Installation and usage

DSH Session Insights 0.3.2

v0.3.2 - DSH market compatibility metadata

English

The npm package now declares DSH 0.1.5-rc.2 consistently in every field used by DSH plugin markets and installers. Market cards will therefore stop advertising the retired 0.1.0-rc.8 threshold after their catalog refresh.

This release changes compatibility metadata only. Runtime behavior is unchanged from 0.3.1, whose source, CI, and macOS native workflow evidence remains applicable. The 0.3.2 package, installation, registry, GitHub Release, and market display will be verified as separate publication surfaces.

English changelog

简体中文

npm 软件包现在会在 DSH 插件市场和安装器读取的全部字段中统一声明 DSH 0.1.5-rc.2。市场目录刷新后，卡片将不再显示已经停止支持的 0.1.0-rc.8 门槛。

本次仅修改兼容元数据，运行时行为与 0.3.1 相同；该版本的源码、CI 和 macOS 原生流程证据仍然适用。0.3.2 软件包、安装、registry、GitHub Release 和市场显示将作为不同发布面分别验证。

中文变更记录

DSH Session Insights 0.3.1

v0.3.1 - DSH 0.1.5 session compatibility

English

Upgraded DSH sessions now appear in reports. The reader selects one current log generation per session, supports compressed and plaintext generations 0–3, and reports unreadable or unsupported generations without falling back to stale history.

Human-work counts exclude injected context. System messages and unsuccessful model attempts are tracked separately; replaced conversation text leaves semantic evidence, while historical tool and token statistics remain. The bridge also uses the same caller-supplied time window for both report modes.

The minimum DSH declaration remains 0.1.0-rc.8. This update targets 0.1.5-rc.2, with macOS native workflow verification and the repository's three-platform CI. Windows/Linux full native model workflows are not repeated for this shared-parser change; future DSH versions are not claimed as verified. Upgrade the Bundle through your existing DSH installation flow and restart your host manually when required.

English changelog

简体中文

升级 DSH 后的新会话现在能够进入报告。读取器为每个会话只选择一个当前日志代际，支持第 0–3 代压缩与明文日志；无法读取或版本不支持时明确报告，不退回过时日志。

用户工作统计排除系统注入内容，系统消息和未形成回复的模型尝试单独统计。被替换的对话不再进入语义证据，历史工具与 Token 统计仍保留；两种报告模式也使用相同的调用方时间窗口。

最低 DSH 版本声明仍为 0.1.0-rc.8。本次针对 0.1.5-rc.2 完成 macOS 原生流程验证，并运行仓库现有三平台 CI。此次共用解析器修复不重复 Windows/Linux 完整原生模型流程，也不声称未来 DSH 版本已验证。沿用现有 DSH 安装流程升级 Bundle，需要时手动重启宿主。

中文变更记录

DSH Session Insights 0.3.0

v0.3.0 - Dashboard overhaul, usage cards, and skill/plugin metrics

English

What you can now do

• See your most-used Skill and plugin/MCP servers as dedicated cards alongside session types, tools, file types, and active hours.
• Read the daily active-time chart in hours with complete date labels and KPIs ordered around time and token usage.
• Use scroll-aware navigation for Session details and Evidence entries, with the active section following document order.

Key changes

• Unified card treatment, 28 px spacing rhythm, equal-height rows with aligned actions, and one-column collapse below 850 px.
• Rebuilt the filter toolbar as a six-column form grid with aligned buttons.
• Added status and per-card color accents, plus locale-aware number formatting such as 1.4B and 14亿.
• Deterministic analyzers now record per-session skill_calls and export Skill/plugin usage aggregates.
• Refreshed the README feature table and light-theme screenshots.

Validation and limits

• Release commit d6065402dcae195eba6f6cd8cdd453bc8129b9f1; annotated tag v0.3.0.
• 37 Python tests and 11 Node.js tests passed, together with fixture, public-tree, package, and three-platform CI checks.
• macOS passed the isolated deterministic workflow. Windows passed the real-model English and Chinese semantic workflows, rendered-English-DOM check, and complete report readback.
• Interactive /session-insights dispatch inside a live plugin-loaded DSH Web profile remained unverified on both platforms. The concurrent-home, removal/rescan, and resource-path-alias edges from v0.2.0 were not repeated.

English acceptance record

简体中文

现在可以做什么

• 在使用情况区域查看最常使用的 Skill 与插件/MCP 服务卡片，并与会话类型、工具、文件类型和活跃时段一起比较。
• 以小时为单位阅读每日活跃时间图，查看完整日期标签，以及围绕时间和 Token 用量重新排列的关键指标。
• 使用跟随滚动位置的导航进入“会话明细”和“证据”部分，当前区块按文档顺序切换。

主要变化

• 统一卡片样式、28 px 间距节奏、等高行和底部对齐操作，并在 850 px 以下切换为单列布局。
• 将筛选工具栏重构为六列表单网格，按钮保持同行对齐。
• 增加状态与卡片配色，并按报告语言格式化数字，例如 1.4B 与 14亿。
• 确定性分析器现在记录每个会话的 skill_calls，并导出 Skill 与插件使用汇总。
• 更新 README 功能表和浅色主题截图。

验证与限制

• Release commit 为 d6065402dcae195eba6f6cd8cdd453bc8129b9f1，annotated tag 为 v0.3.0。
• 37 项 Python 测试、11 项 Node.js 测试，以及 fixture、公开树、软件包和三平台 CI 检查均通过。
• macOS 通过隔离的确定性流程；Windows 通过真实模型中英文语义流程、英文 DOM 渲染检查和完整报告回读。
• 两个平台都未验证已加载 DSH Web profile 中的交互式 /session-insights 分发；v0.2.0 的并发 home、移除/重扫和资源路径别名边界也未在本版本重复测试。

中文验收摘要

DSH Session Insights 0.2.0

dsh-session-insights v0.2.0

v0.2.0 adds a native DeepSeek Harness Bundle experience to the existing local-first session retrospective engine. Install it from npm, run /session-insights, and receive a deterministic or model-assisted bilingual workflow report without copying raw session snapshots into the insights workspace.

Highlights

• Native /session-insights command plus six bounded model-facing workflow tools.
• Official sessionQuery ingestion over a versioned JSONL stdin bridge.
• Complete deterministic and semantic workflows with validation, bounded repair, resume, cancellation, and explicit fallback.
• Chinese and English Dashboard and semantic output from one report schema.
• Existing file-log CLI and Agent Skill workflows remain available.

Install

Requirements: DeepSeek Harness and Python 3.11 or newer.

dsh plugin --profile web add dsh-session-insights
dsh web

Then run in the DSH composer:

/session-insights --days 30 --locale en

Use --locale zh-CN for Chinese output or --deterministic to skip the model-assisted stage.

Privacy boundary

In native plugin mode, complete raw snapshots stream from sessionQuery to the bundled Python bridge over stdin and are not copied into the run directory. Deterministic analysis can run offline. If semantic analysis is enabled, only bounded evidence cleaned according to the selected analysis-privacy mode is sent to the model provider already configured in DSH.

Validation and limits

• Release commit and annotated tag: 7e9f1e3e50cdf464de59edf3a81cebc7f06c2ed8, v0.2.0.
• The exact npm package contains 17 intended files, no bytecode, and no install/build lifecycle scripts. Registry readback was byte-identical to the local artifact: SHA-256 3bb0dd2a64eece762b0eccfca51f6317fe294cece215971211a3bbce84122515.
• Main release CI 32544825580 and tag CI 32545526300 each passed nine Ubuntu/macOS/Windows jobs across Python 3.11–3.13.
• Isolated native macOS acceptance passed Bundle installation, composed-config readback, Web startup, alias-path repetition, and rendered English semantic output.
• Focused Windows-native acceptance passed Bundle loading, composed config, Web startup, the six-tool lifecycle, resume, cancellation, repeated runs, alias-plus-space paths, native project filtering, and real-model en/zh-CN semantic completion without fallback.
• Deterministic slash dispatch and rendered English DOM were not invoked natively on Windows. Plugin removal and concurrent-home behavior also remain explicitly unverified.
• CI is portability evidence and does not replace the native acceptance scopes above.

Release assets include the exact npm Bundle tarball, compatible Python wheel and source distribution, the Agent Skill zip, and SHA256SUMS.txt.

---

dsh-session-insights v0.2.0（中文）

v0.2.0 在原有本地优先会话复盘引擎上增加了 DeepSeek Harness 原生 Bundle 体验。通过 npm 安装后运行 /session-insights，即可生成确定性或模型辅助的双语工作流复盘；原始会话快照不会被复制到 insights 工作目录。

主要内容

• 原生 /session-insights 命令与 6 个有界模型工具。
• 通过版本化 JSONL stdin bridge 接入官方 sessionQuery。
• 完整确定性与语义流程，包含验证、有界修复、续跑、取消和明确降级。
• 基于同一报告 schema 生成中文或英文 Dashboard 与语义内容。
• 保留既有文件日志 CLI 与 Agent Skill 工作流。

安装

需要 DeepSeek Harness 和 Python 3.11 或更高版本。

dsh plugin --profile web add dsh-session-insights
dsh web

随后在 DSH 输入框运行：

/session-insights --days 30 --locale zh-CN

添加 --deterministic 可跳过模型语义阶段。

隐私边界

在原生插件模式下，完整原始快照从 sessionQuery 经 stdin 流式交给内置 Python bridge，不会被复制到运行目录。确定性分析可以离线运行；启用语义分析时，只有按所选 analysis-privacy 模式清洗并限制范围后的证据，才会交给 DSH 已配置的模型提供方。

验证与限制

• Release commit 与 annotated tag：7e9f1e3e50cdf464de59edf3a81cebc7f06c2ed8、v0.2.0。
• exact npm 包含 17 个预期文件，无 bytecode，也没有 install/build 生命周期脚本；registry 下载包与本地 artifact 逐字节一致，SHA-256 为 3bb0dd2a64eece762b0eccfca51f6317fe294cece215971211a3bbce84122515。
• main release CI 32544825580 与 tag CI 32545526300 均通过 Ubuntu/macOS/Windows × Python 3.11–3.13 共 9 个任务。
• macOS 隔离原生验收覆盖 Bundle 安装、合成配置回读、Web 启动、别名路径重复运行和完整英文语义渲染。
• Windows 聚焦原生验收覆盖 Bundle 加载、合成配置、Web 启动、六工具生命周期、续跑、取消、重复运行、别名与空格路径、原生项目筛选，以及真实模型 en/zh-CN 无降级完成。
• Windows 未原生触发确定性斜杠命令分发或英文 DOM 渲染；插件移除与并发 home 行为也仍明确未验证。
• CI 仅提供可移植性证据，不能替代上述原生验收范围。

Release 附件包含 exact npm Bundle tarball、兼容 Python wheel 与源码包、Agent Skill zip，以及 SHA256SUMS.txt。

DSH Session Insights 0.1.0

dsh-session-insights v0.1.0

Local-first, evidence-backed workflow retrospectives for DeepSeek Harness sessions. This first release provides a DSH Agent Skill, Python CLI, deterministic JSON/Markdown reports, a self-contained offline HTML dashboard, and an optional fail-closed semantic workflow.

Highlights

• Reads local DSH session.jsonl.zstd files with project and time filtering.
• Defaults to redacted evidence; metrics omits excerpts and semantic batches, while local requires explicit selection.
• Provides deterministic caching, bounded semantic batches, strict validators, and explicit fallback.
• Installs only marked Skill/runtime directories in an isolated Python 3.11+ venv.
• Ships synthetic fixtures, Schema v1, privacy-negative tests, public-tree auditing, and bilingual documentation.

Install

Review the checkout, then install into an isolated DSH home:

python3 scripts/bootstrap.py install --dsh-home /path/to/isolated-dsh-home

On Windows, use py -3.11 scripts\bootstrap.py install --dsh-home <path>.

Validation and limits

• Release commit: 98d1f3c0bd4ef436a455f78fab6214e4ab85ff11; tag: v0.1.0.
• 27 unit tests plus fixture, Schema, privacy, public-tree, and exact share-boundary checks passed.
• GitHub Actions passed on Ubuntu, macOS, and Windows with Python 3.11–3.13.
• Isolated native acceptance passed on macOS and Windows 11 with DSH 0.1.0-rc.8.
• Native Linux acceptance, real-session testing, real-model semantic round trips, and Windows DSH GUI Skill rescanning were not performed.
• Deterministic reports can run offline. Semantic evidence is sent only when you ask the DSH model configured in your environment to process the bounded, privacy-filtered batches; this tool adds no separate upload path.

Release assets include the DSH Skill bundle, Python wheel and source distribution, plus SHA256SUMS.txt.

---

dsh-session-insights v0.1.0（中文）

这是面向 DeepSeek Harness 会话的首个本地优先、证据驱动复盘版本，包含 DSH Agent Skill、Python CLI、确定性 JSON/Markdown 报告、自包含离线 HTML Dashboard，以及可选的 fail-closed 语义流程。

主要内容

• 读取本地 DSH session.jsonl.zstd，支持项目和时间筛选。
• 默认使用脱敏证据；metrics 不保留摘录且跳过语义批次，local 必须显式选择。
• 提供确定性缓存、有界语义批次、严格验证器和明确降级。
• 仅在隔离 Python 3.11+ venv 中管理带标记的 Skill/runtime 目录。
• 包含合成 fixture、Schema v1、隐私负例、公开树审计和双语文档。

验证与限制

• Release commit：98d1f3c0bd4ef436a455f78fab6214e4ab85ff11；tag：v0.1.0。
• 27 项单元测试以及 fixture、Schema、隐私、公开树和精确 share-boundary 检查通过。
• GitHub Actions 覆盖 Ubuntu、macOS、Windows 与 Python 3.11–3.13。
• macOS 与 Windows 11 已在 DSH 0.1.0-rc.8 上分别完成隔离原生验收。
• 未进行 Linux 原生验收、真实会话测试、真实模型语义往返或 Windows DSH GUI Skill 重扫。
• 确定性报告可离线运行。只有当你让当前 DSH 模型处理有界、按隐私设置清洗后的批次时，证据才会交给环境中配置的模型提供方；本工具不增加独立上传通道。

Release 附件包含 DSH Skill bundle、Python wheel、源码包和 SHA256SUMS.txt。