dsh-skills-panel

Skills Panel

DeepSeek Harness 的 Skills 设置面板：浏览全局和项目级技能，从 skills.sh 搜索并安装，按技能切换模型调用，并在无需 GitHub token 的情况下检查更新。

目标 Profile

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:mathangler/dsh-skills-panel#62cab67ca056d6dbe2626642b47cd806bb9e6c9e

兼容性与来源证明

Skills Panel 以 dsh-skills-panel 发布，当前版本为 0.1.3。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: 2026/9/11

版本

0.1.3stable

2026/9/11

0.1.1stable

2026/9/10

README

dsh-skills-panel

English | 中文

A Skills page for the DeepSeek Harness settings dialog: see every skill DSH can load — global and per project — search and install new ones from skills.sh, control which skills are auto-injected into the model context, and get told when an installed skill changes upstream.

Unofficial. This is a third-party plugin. It is not affiliated with or endorsed by DeepSeek. It uses only public DSH plugin surfaces.

What it does

Installed tab

Lists the effective skill set for a scope, with a source badge per skill (user-dsh, project-dsh, a preset, an agent preset, …) and a marker when a project skill overrides a global one of the same name.
A switch per skill: on means the skill stays in the model-visible catalog, off writes disable-model-invocation: true into the skill's frontmatter so it drops out of the catalog but stays callable as /name. That is DSH's own mechanism, not a private convention.
Skills that are junctions into another tool, or that DSH does not own, are shown read-only — the panel will not rewrite a file it shares with someone else.
View expands the skill's SKILL.md inline.
Remove deletes the skill folder, or only the link when the skill is a junction (a plain rmdir cannot follow a junction into its target).
Check all for updates compares every panel-installed skill against its repository and marks the ones that changed.

Find & install tab

Searches skills.sh and gives each result a GitHub and a skills.sh entry point, so you can inspect it yourself before committing. The panel downloads nothing on its own: a skill you have not confirmed never lands on this machine, not even a byte.
Results you already have are badged already installed and offer Reinstall instead of Install.
Installs to the global skills root or to the current project (.dsh/skills).

Import tab

Brings in a skill folder that already exists on disk, either as a junction (the original folder stays the source of truth) or as a copy.
Imported skills are deliberately excluded from update checks: they have no upstream to compare against.

Updates

Once per page load, in the background, every panel-installed skill is checked against its repository. Changed skills get a dot and a chip, and the tab title gains a count.
Update re-checks first and does nothing when the skill has not changed.

Install

Requires DSH with the web profile and pnpm on PATH.

dsh plugin --profile web add github:mathangler/dsh-skills-panel

Then restart the web app:

dsh web

dsh plugin runs pnpm inside ~/.dsh/profiles/web and then appends this package to dsh.profile.bundles automatically, because the package declares dsh.bundle. You do not need to edit any YAML.

Open Settings → Skills.

Update

dsh plugin --profile web update dsh-skills-panel

Uninstall

dsh plugin --profile web remove dsh-skills-panel

How installing a skill works

Installing does not use the GitHub REST API, so it needs no token and is not subject to the 60-requests-per-hour anonymous limit:

The archive is fetched from codeload.github.com/<repo>/tar.gz/<ref> — trying HEAD first, so no git is required, then main, then master.
tar extracts it into a short-lived cache in the system temp directory.
The skill directory is located by finding the folder named after the skill id that actually contains a SKILL.md.
fs.cp copies it out whole — byte for byte, so scripts, images and other binary assets come along intact.

Install and check-for-updates take that pipeline; nothing else does. Both are actions you asked for by name. Searching sends no request for a skill's content, which is the point: a stranger's repository should not appear on your disk before you have decided anything.

The pipeline needs only codeload.github.com, never raw.githubusercontent.com. That matters on networks where raw is unreachable while codeload works.

Each installed skill is recorded in .skills-panel.json in its skills root with its repository, branch, in-repo path and a content hash. The manifest is what makes update checks and clean removals possible; delete it and the panel will still show the skill, but will not offer updates for it.

Platform notes

Windows, macOS and Linux. There is no shell script and no PowerShell: the whole pipeline is Node — fetch, fs.cp, fs.symlink, fs.rm — which behaves the same on all three. A Windows junction and a POSIX symlink are both created with fs.symlink, so neither needs elevation.
tar is the only external tool. It ships with macOS and Linux, and with Windows 10 (1803) and later. You do not need git, curl or PowerShell.
The update check compares SKILL.md content only. If upstream changes only an auxiliary file, the panel reports "up to date". This is a deliberate trade-off to keep update checks cheap.

License

MIT — see LICENSE.

README

dsh-skills-panel

English | 中文

Unofficial. This is a third-party plugin. It is not affiliated with or endorsed by DeepSeek. It uses only public DSH plugin surfaces.

What it does

Installed tab

Lists the effective skill set for a scope, with a source badge per skill (user-dsh, project-dsh, a preset, an agent preset, …) and a marker when a project skill overrides a global one of the same name.
A switch per skill: on means the skill stays in the model-visible catalog, off writes disable-model-invocation: true into the skill's frontmatter so it drops out of the catalog but stays callable as /name. That is DSH's own mechanism, not a private convention.
Skills that are junctions into another tool, or that DSH does not own, are shown read-only — the panel will not rewrite a file it shares with someone else.
View expands the skill's SKILL.md inline.
Remove deletes the skill folder, or only the link when the skill is a junction (a plain rmdir cannot follow a junction into its target).
Check all for updates compares every panel-installed skill against its repository and marks the ones that changed.

Find & install tab

Searches skills.sh and gives each result a GitHub and a skills.sh entry point, so you can inspect it yourself before committing. The panel downloads nothing on its own: a skill you have not confirmed never lands on this machine, not even a byte.
Results you already have are badged already installed and offer Reinstall instead of Install.
Installs to the global skills root or to the current project (.dsh/skills).

Import tab

Brings in a skill folder that already exists on disk, either as a junction (the original folder stays the source of truth) or as a copy.
Imported skills are deliberately excluded from update checks: they have no upstream to compare against.

Updates

Once per page load, in the background, every panel-installed skill is checked against its repository. Changed skills get a dot and a chip, and the tab title gains a count.
Update re-checks first and does nothing when the skill has not changed.

Install

Requires DSH with the web profile and pnpm on PATH.

dsh plugin --profile web add github:mathangler/dsh-skills-panel

Then restart the web app:

dsh web

dsh plugin runs pnpm inside ~/.dsh/profiles/web and then appends this package to dsh.profile.bundles automatically, because the package declares dsh.bundle. You do not need to edit any YAML.

Open Settings → Skills.

Update

dsh plugin --profile web update dsh-skills-panel

Uninstall

dsh plugin --profile web remove dsh-skills-panel

How installing a skill works

Installing does not use the GitHub REST API, so it needs no token and is not subject to the 60-requests-per-hour anonymous limit:

The archive is fetched from codeload.github.com/<repo>/tar.gz/<ref> — trying HEAD first, so no git is required, then main, then master.
tar extracts it into a short-lived cache in the system temp directory.
The skill directory is located by finding the folder named after the skill id that actually contains a SKILL.md.
fs.cp copies it out whole — byte for byte, so scripts, images and other binary assets come along intact.

The pipeline needs only codeload.github.com, never raw.githubusercontent.com. That matters on networks where raw is unreachable while codeload works.

Platform notes

Windows, macOS and Linux. There is no shell script and no PowerShell: the whole pipeline is Node — fetch, fs.cp, fs.symlink, fs.rm — which behaves the same on all three. A Windows junction and a POSIX symlink are both created with fs.symlink, so neither needs elevation.
tar is the only external tool. It ships with macOS and Linux, and with Windows 10 (1803) and later. You do not need git, curl or PowerShell.
The update check compares SKILL.md content only. If upstream changes only an auxiliary file, the panel reports "up to date". This is a deliberate trade-off to keep update checks cheap.

License

MIT — see LICENSE.

Skills Panel

兼容性与来源证明

版本

相关插件

Skills Panel

README

dsh-skills-panel

What it does

Install

Update

Uninstall

How installing a skill works

Platform notes

License

兼容性与来源证明

版本

相关插件

相关插件

相关插件

README

dsh-skills-panel

What it does

Install

Update

Uninstall

How installing a skill works

Platform notes

License