Agent Plugins

dsh-agent-plugins-market

A one-stop skills, subagent, MCP and LSP manager inside DeepSeek Harness (DSH), compatible with Claude Code, Codex, Cursor, Kimi and other agent plugin layouts.

Reuse supported content from Claude Code, Codex, Cursor, Kimi and other recognized layouts, and manage your own skills, commands, agent personas, MCP services and LSP servers in the DSH Web GUI.

If this plugin is useful to you, a ⭐ on GitHub is appreciated.

English | 简体中文 | Documentation | npm

Quick start · Everyday use · Compatibility · FAQ

Pages

Market Add sources, preview suites, install and enable.	Skills Browse skills and author your own.	Commands Manage prompt templates invoked as /name.
Agent personas Roles with an exact provider, model and effort.	MCP services Credentials, authorization and connection status.	LSP servers Language server configuration and runtime status.

What you can do

Enable the experimental Agent preset manager in plugin settings to show its composer entry. It is off by default and does not disable the session runtime. Session extension presets select skills, commands, roles, MCP and LSP capabilities through the Agent extension icon button to the right of Permissions. Each workspace owns a preset library and a default for new sessions. Selecting a preset while a session runtime is active takes effect on that session's next user request. Mid-turn steering and tools do not change. Once the selection commits, the open slash skill menu refreshes without another keystroke or page reload. Editing the selected or pending preset in the current session updates that session. Every other session keeps its own detached copy. Copy and paste a preset to another workspace to create an independent copy. Globally managed resources show their actual state without a session switch. Preset details do not change shared credentials or configuration. After a failed session selection, explicit recovery restores readiness and continues queued user input without another Send.
Ten suite layouts. Claude Code, Codex, Cursor, Kimi Code, ZCode, Qoder CLI, GitHub Copilot CLI, Universal .plugin/, agent-plugins and manifest-less skill collections.
Sources. Add a Git repository, a local directory or an archive (.zip / .tar.gz / .tgz / .tar); adopt a checkout you cloned yourself; refresh on demand; delete a managed checkout when you remove its source.
First-party source. The plugin presets one source record pointing at its own suite collection, so the market lists that repository on first run with no URL to paste. It is an ordinary Git source from there on — refresh it to fetch, then install and toggle its suites like any other.
Downloads that fit your network. Pick a download region — default auto follows the interface language, or choose global / China mainland — and the plugin routes github.com clones through the matching mirror. A proxy and per-invocation tuning live in the host config.
Runtime surfaces. Enabled suites inject into sessions: skills into the catalog and slash menu, commands as /name, agent personas into the subagent catalog, MCP tools with an mcp__ prefix, hooks onto host lifecycle events, and language servers through the lsp tool.
MCP. A built-in bridge runs stdio, Streamable HTTP with OAuth and legacy SSE without a host MCP client. ${VAR} references resolve from the host credential store or the launch environment; per-server overrides disable or patch a declaration without editing the source, set the tool-call and startup timeouts, and switch individual tools off; creation and editing share the same service form, including advanced timeout settings and a JSON view; an optional host-client compatibility mode is available. A suite's tools register as mcp__<suite>__<server>__<tool>; a server you declare yourself in ~/.agents/mcp.json mounts under its own key, mcp__<server>__<tool>.
LSP. Self-provisioned: installing the plugin is the whole setup, and the lsp tool mounts only while a language server is wanted. The server executable itself must be on PATH. Upgrading from a release that asked you to expose LSP from your profile? A profile that still carries that hand-added layer reports a seam conflict; the LSP panel names the file and removes the layer for you, keeping a backup.
Agent personas and delegation. Role cards keep their instructions and exact provider/model/effort settings. Without Agent Teams, use subagent_role; with Agent Teams, use spawn_teammate_role to create a real Team member and manage it through native Team tools.
Project dimension. Skills, agents, commands, MCP servers and hooks are read from the project's own directories with no install step.
Your own resources. Author skills, commands and agent personas as Markdown under ~/.agents/, then edit them or disable them without deleting the files.
Background source updates. Optionally refresh every configured source on a timer; off by default.
Web workspace. Seven tabs: Market, Skills, Commands, Agent personas, MCP services, LSP servers and Hooks, each with search, filters and a grid/list toggle, plus status panels with diagnostics, credential editing, and an install confirmation that warns before executable third-party content is enabled. The settings page is the canonical layout. Every surface shares one tab row of equal columns that never scrolls sideways, and a label that does not fit ellipsizes. Every surface also shares one card and detail anatomy, and the preset manager follows it.
Bilingual interface and feedback. Workspace strings and injected prompts follow the host language. With feedback enabled, the model can file a report_market_issue report through the gh CLI or a GitHub token; with neither, it opens a prefilled GitHub issue page and hands you the complete issue text.
Translation. Descriptions and expanded documents can follow the Chinese or English interface. Names stay unchanged. Chinese defaults to on and English to off, but your saved choice takes priority. Documents offer original, translated, and paragraph-by-paragraph bilingual views, with bilingual selected by default. Translation starts on demand, tries Google Translate, Microsoft Translator, then your default model, and keeps the original when none answers. Results stay cached locally until cleared. Turning translation off keeps that cache. See translation.

Quick start

Install into your profile, replacing <name> with its name:

dsh plugin --profile <name> add dsh-agent-plugins-market

Restart DSH and open Settings → Agent Plugins.
The market already lists the first-party source; press Refresh to fetch its suites. To pull in more, add a source, for example https://github.com/anthropics/claude-plugins-official.
Open a suite, review its contents, then install it and ensure it is enabled.
For a suite with skills, check the Skills tab and type / in chat to find its user-invocable skills. For an MCP suite, check MCP services and resolve any credential or connection notice before using its tools.

Requirements, profile configuration and alternative installs: usage guide.

Everyday use

The workspace has seven tabs:

Tab	Use it to
Market	Add sources, preview suites, install / uninstall, enable / disable and refresh.
Skills	Browse skills and create or edit your own reusable instructions.
Commands	Manage prompt templates invoked as `/name`.
Agent personas	Save role instructions and model settings; delegate through `subagent_role` or create role-aware Team members with `spawn_teammate_role`.
MCP services	Add a service or configure an installed one, its credentials and authorization; inspect status and retry failures.
LSP servers	Add and configure language servers and inspect their runtime status.
Hooks	Inspect configured command hooks and supported events. Each row opens a declaration card with support, event, matcher and timeout, plus a dry run in your home directory. Installed suite hooks list read-only and follow their suite.

A source is where content comes from; a suite is an installable unit discovered there. Adding a source discovers its suites. Installing and enabling a suite controls its runtime capabilities.

Everything you author yourself lives in the shared Agent layout root: skills, commands and personas as Markdown under ~/.agents/, command hooks in ~/.agents/hooks.json (or ~/.agents/hooks/hooks.json), and the MCP and LSP services you add in the workspace in ~/.agents/mcp.json and ~/.agents/lsp.json. Commands and personas are read at any subdirectory depth, so a file another tool wrote at ~/.agents/commands/git/commit.md is callable as /git-commit. Project-native resources stay in the project. See storage and discovery for paths and precedence.

All seven tabs share a saved grid/list preference. Add and refresh actions sit at the top right; resource state rails are green when active. A hook row opens its own read-only declaration. The declaration shows the full command, the declared matcher, the declared timeout and the host support verdict, never the whole suite behind it.

Compatibility and boundaries

Supported layout dialects describe how files are organized. The shared priority table lists suite manifests and Marketplace catalogs together. All ten layout contracts in schemas/ have independent reader tests; the layout audit maps them to pinned repository fixtures.

Supported runtime surfaces describe what DSH can use:

Surface	Runtime support and conditions
Skills	Host skill catalog and user-invocable slash entries; supported root placeholders are expanded.
Commands	Slash commands through the host command service.
Agents	Dynamic role catalog; `subagent_role` without Agent Teams, `spawn_teammate_role` with it. Team mode also requires session query and system-prompt services.
MCP	Built-in bridge by default: stdio, Streamable HTTP with OAuth, and legacy SSE. Optional host-client compatibility mode is also available; it enforces the tool-call timeout but not tool filters or a startup timeout.
Hooks	The command-hook subset mapped by `dsh-hooks-claude-code`.
LSP	Self-provisioned: installing the plugin is the whole setup, and `lsp` mounts only while a language server is wanted; the server executable must be on `PATH`.

Role discovery remains available in either mode. spawn_teammate_role(agent, name, description, prompt) applies role instructions and model settings from the first request, returns a native Team target, and retains the creation snapshot for resume. Only the Lead creates members, after the user explicitly requests Team work. Native spawn_teammate stays available; roles do not add a second management API. A separate Team coordination briefing keeps the Lead on user intent and integration while guiding members through scoped execution and reporting, even without a role catalog. tools and disallowedTools are preserved but not enforced. See agent roles for routing, prerequisites and limits.

Layout detection precedence

When multiple manifests exist in the same suite directory, the first existing file in this order selects the layout:

Priority	Layout	Suite manifest	Marketplace catalog
1	agent-plugins / root compatibility	`plugin.json`	No dedicated catalog
2	Universal compatibility	`.plugin/plugin.json`	`.plugin/marketplace.json`
3	Claude Code	`.claude-plugin/plugin.json`	`.claude-plugin/marketplace.json`
4	Cursor	`.cursor-plugin/plugin.json`	`.cursor-plugin/marketplace.json`
5	Kimi Code	`kimi.plugin.json`, then `.kimi-plugin/plugin.json`	`.kimi-plugin/marketplace.json`
6	Codex	`.codex-plugin/plugin.json`	`.agents/plugins/marketplace.json`, then `.agents/plugins/api_marketplace.json`
7	ZCode	`.zcode-plugin/plugin.json`	No dedicated catalog
8	Qoder CLI	`.qoder-plugin/plugin.json`	`.qoder-plugin/marketplace.json`
9	GitHub Copilot CLI	`.github/plugin/plugin.json`	`.github/plugin/marketplace.json`
Fallback	Skill collection / shared catalog	Discover skills when no recognized manifest exists	Root `marketplace.json`

Manifests are tried in order: a manifest that cannot be read or validated is reported and the next one is tried, down to the fallback. If every candidate fails, the suite is diagnosed instead of loading half of it.
Component fallback: for a root plugin.json that does not declare a recognized agent-plugins $schema, missing component declarations can come from .claude-plugin/plugin.json; explicit root declarations win, and marketplace entry declarations fill remaining gaps.
Marketplace catalogs follow the same order: the first catalog that produces suites wins, and invalid or empty catalogs allow the next candidate to be tried.

The order is defined in packages/market-contracts/src/model/layouts.ts; selection and root-manifest fallback are implemented in packages/market-catalog/src/scanning/manifests.ts.

Layout support matrix

The table says per layout whether this plugin reads a given surface at all. Yes means the layout's own files are read and injected; Partial means only part of the formats is understood, or the upstream layout has no such definition — the notes below say which. Evidence: the compatibility report and layout audit.

Layout	Skills	Agents	Commands	MCP	Hooks	LSP
Claude Code	Yes	Yes	Yes	Yes	Partial	Yes
Codex	Yes	Yes	Yes	Partial	Partial	Yes
Cursor	Yes	Partial	Partial	Partial	No	Yes
Kimi Code	Yes	Yes	Yes	Yes	Partial	Partial
ZCode `.zcode-plugin/`	Yes	Yes	Yes	Yes	Partial	Partial
Qoder CLI `.qoder-plugin/`	Yes	Yes	Yes	Partial	Partial	Partial
GitHub Copilot CLI	Yes	Yes	Yes	Yes	Partial	Yes
Universal compatibility layout `.plugin/`	Yes	Yes	Yes	Yes	Partial	Yes
agent-plugins	Yes	Yes	Yes	Yes	Yes	Yes
Manifest-less skill collection	Yes	Yes	Yes	Yes	Partial	Partial
Project-native directories	Yes	Yes	Yes	Yes	Partial	No

Skills are read from the paths a manifest declares and from the conventional skills/ directory, including flat <name>.md files. A path that carries its own SKILL.md is one skill: the documents beside it belong to that skill and are not read as skills. A conformant agent-plugins package is the exception: skills come from one level of skills/ subdirectories, exactly as its specification requires.
Agents and commands are read as Markdown (agents/*.md, commands/*.md). Cursor plugin commands accept .md, .mdc, .markdown and .txt. Codex and Kimi native agent/command formats (TOML, YAML) have no adapter yet.
MCP covers declared files, inline tables and arrays. Cursor's schema-less mcp.json and agent-plugins' strict mcp.json both work; Kimi Code is inline-only. Codex app connectors stay outside this adapter.
Hooks map the command-style events DSH has an equivalent for; events without one (for example afterFileEdit) are reported instead of simulated. Cursor's native events are not read.
LSP accepts declared files, arrays and inline tables plus the conventional .lsp.json / lsp.json locations. Declarations inside a project are reported but not mounted: the host LSP registry is global. Some layouts only expose LSP directories for preview.
agent-plugins suites read the portable core (skills/, mcp.json) per the specification, plus this plugin's com.deepseek.harness extension namespace for commands, agents, hooks, LSP and per-server MCP policy (OAuth, tool lists, timeouts). Root-level commands/, agents/, hooks/ and .mcp.json files belong to other layouts and are reported as unread for this dialect; inline manifest component keys are reported and ignored per §5.2. Both recognized releases (1.0.0, 1.1.0) validate against their own vendored schemas.
Universal is a compatibility-layout label used by this plugin; the OpenHands SDK documents the same .plugin/plugin.json location and a Vercel repository uses it, but no cross-vendor specification exists.

Sources such as jeecgboot/skills group undeclared, skill-only directories into one plugin card. Marketplace entries and roots with plugin manifests or other runtime surfaces retain their separate identities. The aggregate requires its own installation and does not inherit individual skill installs. See skill collection rules.

Reading a layout does not guarantee every behavior of its original platform. Invalid declarations are diagnosed and skipped.

Project layout switch

The plugin settings card has Scan project Agent layouts (dsh-agent-plugins-market.scanProjectLayouts, default off). It controls one thing: whether the project you are working in contributes skills, commands, agent roles, MCP servers and hooks from its own directories (.claude, .agents, .codex, .cursor, .kimi, .zcode, .qoder, .github). Turning it on adds those candidates; turning it off removes them on the next discovery pass. The card applies a switch when you save it, and each setting offers Use default while it carries your own value. Configured sources and installed suites are unaffected.

Project layouts lists the directories and files read per layout and how they are mounted.

Verified samples

The README repositories have offline snapshots in tests/fixtures/real-layouts/ with commit IDs, hashes and licenses, and each layout has an isolated reader test. The compatibility report records the sampled repositories, their schema verdicts and the scanner output; the audit records independent resource checks. These are documentation, source and sample checks — not end-to-end certification for every platform.

Review third-party suites before enabling them: enabled services and hooks can execute programs. See the runtime and security details.

FAQ

Why is an installed skill or tool missing?

Check that the suite and the relevant capability are enabled. Skills may restrict manual invocation. MCP / LSP panels show user-service failures. Project resources additionally require the project-scan switch; unsupported native fields and project LSP declarations appear in scan diagnostics.

Where do I configure MCP tokens?

Open the service in MCP services. Missing environment references show needs-credentials. Host-managed credentials are write-only; launch-environment credentials require changing the environment and restarting DSH.

How do I add a service that no suite declares?

Use Add in MCP services or LSP servers. The declaration is validated, stored in ~/.agents/mcp.json or ~/.agents/lsp.json, and mounted through the same lifecycle as suite services. Host-observed services remain read-only.

Do sources refresh automatically?

Only when Background source updates is on: every configured source is then refreshed every 6 hours, starting one interval after you enable it. The switch is off by default, and the refresh button always works.

Where do the suites that appear without adding a source come from?

The plugin presets one source record pointing at its own suite collection, so the market lists that repository on first run; press Refresh to fetch its suites. From there it behaves like any other source — including removal, which the next activation undoes just as it does for a source seeded through configuration.

What if a source download fails?

Use a local directory, adopt a manual checkout, or configure a proxy / mirror. See source configuration.

When do local edits become visible?

There is no file watcher. Local-source discovery caches results for up to 30 seconds; refresh the source to invalidate them immediately. Project discovery has a separate five-second cache. A refresh does not happen automatically on an already-open page.

Does removing a source delete its files?

Only when you tick also delete the managed market directory in the confirmation. That removes the source's checkout under ~/.dsh/agent-plugins/.sources/<id> — including one you cloned yourself and adopted. A local-directory source pointing outside .sources/ is never deleted.

Repository layout

The published package stays dsh-agent-plugins-market and ships one artifact. The code lives in eight private workspace packages under packages/: market-bundle (composition root and HTTP surface), market-catalog (source scanning and catalog use cases), market-contracts (stateless shared records), market-runtime (session selection and harness effects), market-mcp, market-lsp, market-translation, and market-ui (the browser bundle). Ownership, dependency direction and the migration gates are recorded in the domain workspace refactor; AGENTS.md holds the layout map maintainers edit.

Community

Scan the QR code to join the dsh-agent-plugins-market WeChat group, where we answer questions and take feature requests.

WeChat group QR code for dsh-agent-plugins-market

dsh-agent-plugins-market

安装

源码

标签

作者

说明文档