DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

Windows Workspace Guard — DeepSeek Harness 插件(DSH Plugin)
← Plugins
W

dsh-windows-workspace-guard

Windows Workspace Guard

在 Windows 上,通过工作区、凭据、审批和审计策略保护 DeepSeek Harness 的 PowerShell、文件、图像和搜索工具。

插件会安装到这里;不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add github:julescules/dsh-windows-workspace-guard#9f14466057bc7785e81a999e5877a970996f3960
README兼容性版本

兼容性与来源证明

Windows Workspace Guard 以 dsh-windows-workspace-guard 发布,当前版本为 1.1.1。Plugin Hub 会校验它的 manifest,并保存精确安装来源,便于复现安装结果。

DSH 兼容范围
*
运行环境
web
发布来源
github
Registry 更新时间
2026/9/5

版本

1.1.1stable
2026/9/5
1.1.0stable
2026/9/5
1.0.0
stable
2026/9/4
查看其余 5 个版本收起版本
0.9.0stable
2026/8/28
0.8.0stable
2026/8/24
0.7.0stable
2026/8/24
0.5.0stable
2026/8/22
0.4.0stable
2026/8/20

相关插件

正在加载相关插件…

最新版
1.1.1
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
未提供
文件数
未提供
Surface
web
许可证
MIT
发布源
github
GitHub
★ 0
周下载
0
最近提交
2026/9/5
查看源码 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

验证 package.json 声明的 GitHub 仓库,即可管理这个公开页面。认领后,Hub 会优先安排当前版本的安全扫描,并在通过后公开展示结果。

认领这个 Plugin →
报告问题
DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录

相关插件

继续浏览 security-access 分类下经过校验的插件。

Doctor@linxin666/dsh-doctorDSH 配置档案的事务性救援模式,配备受监督的启动器、隔离的恢复容器、确定性修复、健康监控以及本地 Web 恢复控制台Pocketdsh-pocket把 DeepSeek Harness 装进你的口袋:一个包、一个设置页,手机扫码即同步访问电脑上的 DSH(局域网 + 公网,实时同屏)。DSCODE@toddzheng024/dscode-bundle完整的 DeepSeek 编码代理,支持持久化 shell、Ultra 协作和自动权限审查。Auto Reviewdsh-auto-review针对 DeepSeek Harness 审批请求的第二模型 AI 自动审查:只读审查子代理在审批应答链上决定允许或拒绝,并采用故障关闭回退机制和完整的会话日志审计。

README

Windows Workspace Guard for DeepSeek Harness

This release caps audit files at 32 MiB, rejects invalid limits, and refreshes the visual walkthrough.

Understand what your Windows agent is about to change—and why a call was blocked.

Your problemAsk the agent
A command looks risky“Check this command without executing it.”
Protection seems incomplete after an upgrade“Run windows_workspace_guard_doctor and explain uncovered tools.”
Too many blocks, no clear explanation“Run windows_workspace_guard_audit_summary and show the most frequent rule IDs.”

New in v1.1.1: make audit logs useful

The new windows_workspace_guard_audit_summary tool reads the configured audit log and returns decision totals and frequent rule IDs. It excludes commands, working directories, and target paths. Malformed rows and the 100,000-record limit are reported explicitly so partial evidence is visible.

Set auditPath to a project-owned JSONL path in the plugin settings to collect future decisions. Logging is opt-in; the tool cannot reconstruct old activity. You can also run node audit-summary.js D:\project\logs\guard.jsonl from the package directory without a model or API key.

Use the summary to identify a rule, then inspect that specific case with the dry-run tool before changing policy.

中文 | English

[!IMPORTANT] Unofficial community plugin. Independently developed and maintained; not reviewed or endorsed by DeepSeek.

Stop a Windows agent before it deletes originals, escapes the workspace, destroys Git recovery paths, reads credentials, or changes system state. PowerShell and both official filesystem suites receive a clear PASS, ASK, or HARD BLOCK before dispatch.

Illustrated guard workflow

Start in 30 seconds

dsh plugin --profile web add github:julescules/dsh-windows-workspace-guard#v1.1.1
dsh --profile web --dump-config
dsh --profile web

Restart DSH, then ask:

Run windows_workspace_guard_doctor, then use windows_workspace_guard_check
to inspect this command without executing it:
Get-Content -LiteralPath $env:DSH_HOME\.credentials.yaml

Why v1.1.1

Verified multi-tool boundary

The default protected set follows the official rc.1 Windows/file surface: pwsh, read, extensionless-capable read_image, write, edit, glob, grep, and Minimal's str_replace_editor. Each adapter reads only the official path and operation fields, omits file content, search text, and replacement strings from policy previews, and applies the same workspace, immutable-path, existing-link, and sensitive-path rules.

Configured tools without a verified argument adapter fail closed. Doctor reports every configured tool as covered or unsupported; the plugin never treats an unknown schema as PowerShell and calls it protected.

Upgrades preserve live settings. If an older profile already stored a shorter toolNames list, add read_image, glob, and grep in the settings card; Doctor will show the effective coverage.

Official grep searches hidden and ignored files. With guardSensitiveData enabled, v1.1.1 requires a narrow include glob such as *.js or *.ks; an unbounded search fails closed before a hidden .env can be returned. Explicit credential-like paths and glob patterns are blocked too.

Monotonic hard blocks

Workspace escapes and unverifiable mutation paths are hard blocks. allowExact remains available for explicit review-level commands, but it cannot authorize a write outside the configured workspace or bypass any non-overridable policy.

Static non-overridable rules are also registered through the official synchronous ctx.tools.guard() seam. They remain denials even if another reorderable tools/pre-execute listener short-circuits its waterfall. Older Harness builds without that API keep the pre-execute fallback.

Live junction/symlink inspection remains asynchronous in tools/pre-execute; it cannot be moved into a synchronous guard and is reported honestly as a separate layer.

Credential and secret boundary

With guardSensitiveData: true (default), guarded pwsh, read, and editor calls block explicit reads or copies of:

  • $DSH_HOME\.credentials.yaml and .env files;
  • user SSH, AWS, Azure, Git, npm, GitHub CLI, and NuGet credential locations;
  • configured sensitivePaths;
  • sensitive environment variables and full Env: enumeration;
  • same-command outbound-network use combined with an explicit sensitive source.

This is a conservative tool boundary, not a general DLP system. It does not inspect arbitrary native-process memory, already-running processes, or tools outside toolNames.

Read-only Windows doctor

windows_workspace_guard_doctor reports facts without changing ACLs or configuration:

  • ctx.tools.guard() availability;
  • configured-tool coverage and adapter type;
  • DSH home and configured workspace/protected path state;
  • existing link metadata for configured roots;
  • audit-path writability and bounded duplicate-runtime checks;
  • credential-file ACL metadata on Windows, without opening credential contents.

Decisions

ResultMeaning
PASSNo matched risk under the active policy.
ASKA reviewable operation needs one host approval in mode: ask.
HARD BLOCKDisk/system mutation, policy bypass, immutable path, link traversal, or sensitive-data access cannot be approved away.

Use windows_workspace_guard_check for a dry run. Select pwsh with command; select read, write, or edit with path; or select str_replace_editor with an operation plus absolute path. It returns machine-readable findings and never executes the call.

Main settings

The DSH Web settings card updates these values live:

enabled: true
mode: block               # block | ask | report
toolNames: [pwsh, read, read_image, write, edit, glob, grep, str_replace_editor]
workspaceRoots: []        # empty = current session cwd
protectedPaths: []
guardExistingLinks: true
guardSensitiveData: true
sensitivePaths: []
auditPath: ''             # optional append-only JSONL
auditIncludeCommand: false
auditFailClosed: false

With an audit path, dispatch writes occur after host approval and monotonic guards. Denied calls are observed after their final result. Commands are hashed and redacted by default.

DSH integration

The package stays outside Harness core and uses the official dsh.bundle.patch, tools/pre-execute, ctx.tools.guard(), tools/execute, tools/result, settings, and typed-tool seams.

Upgrade, disable, uninstall

dsh plugin --profile web add github:julescules/dsh-windows-workspace-guard#v1.1.1
dsh plugin --profile web list
dsh plugin --help

Use the disable/remove command shown by dsh plugin --help for your Harness build, then restart DSH. Profile command names are still changing during developer preview.

Troubleshooting and data

  • Missing settings card: run dsh --profile web --dump-config, confirm windows-workspace-guard, then restart Web.
  • False positive: run the dry-run tool and report finding IDs plus a redacted command/path.
  • The plugin makes no network requests. Doctor reads filesystem/ACL metadata only and never credential values.
  • No audit file is created unless auditPath is configured.
  • Keep workspaceRoots and sensitivePaths absolute and narrow.

Verify

npm run check
npm pack --dry-run
node scripts/build-release-metadata.mjs .\dsh-windows-workspace-guard-1.1.1.tgz .\builds\v1.1.1
.\scripts\verify-release.ps1 -PackagePath .\dsh-windows-workspace-guard-1.1.1.tgz -ChecksumsPath .\builds\v1.1.1\SHA256SUMS

Each release ships a SHA-256 checksum and CycloneDX SBOM. The verifier reads files literally and makes no network request.

Limits

  • Static policy is not an operating-system sandbox.
  • A filesystem TOCTOU window remains between link inspection and execution.
  • Tools outside toolNames are not covered; configured names with unknown argument schemas fail closed and appear as Doctor warnings.
  • ACL warnings are review evidence, not automatic permission repair.

For missed cases, reply in the official community plugin thread with Windows, PowerShell and DSH versions, a redacted command, expected PASS/ASK/BLOCK, and redacted doctor output.

See CONTRIBUTING.md and SECURITY.md.

License

MIT

Compatibility: runtime-validated on DSH 0.1.2-rc.1. Upstream 0.1.3-alpha.1 is published on GitHub; its npm package was unavailable on 2026-09-06, so runtime compatibility with that alpha is not yet claimed.