Skip to content

DSH-PLUGIN.md

Latest commit

 

History

History
71 lines (54 loc) · 3.58 KB

DSH-PLUGIN.md

File metadata and controls

71 lines (54 loc) · 3.58 KB
Raw

The DeepSeek Harness bundle

DeepSeek Harness (dsh) is a coding agent built on "everything is a plugin". It is the fourth shop window for the same skill, and the only one that can ship the connector and the CLI path in a single zero-config install.

Install

dsh plugin --profile <name> add github:superdesigndev/treg
dsh --profile <name>

Why there is no manifest

dsh reads no plugin.json. It installs an npm package whose package.json declares dsh.bundle, pointing at a cordis.patch.yml config layer, and appends that bundle to the profile's dsh.profile.bundles list. Layers compose in list order, so a user can override any row we ship from their own profile patch without touching this package (publish.md).

package.json                  the bundle manifest — `dsh.bundle.patch`
dsh/cordis.patch.yml          the layer: the skill row + the MCP row
dsh/index.js                  a ctx.skills provider over the tree below
dsh/skills/treg/SKILL.md      GENERATED — never edit by hand

The MCP row is disabled until there is a token

The Claude Code plugin declares no connector at all, because anything declared in a manifest is registered at install time, before any human has signed in: five always-on tools that 401 on every call (CLAUDE-PLUGIN.md). dsh can express the better version of that trade, because a row carries a disabled expression evaluated at boot:

- id: treg-mcp
  name: '@deepseek-ai/dsh-mcp-client'
  disabled: !!js !process.env.TREG_TOKEN

So a user with TREG_TOKEN in their environment gets mcp__treg__catalog_search and friends at boot, and everyone else gets a clean install and a skill that walks them through treg login. The row is evaluated at boot, so a token exported afterwards needs a dsh restart — the bootstrap says so, because a human waiting for tools that will not appear this session is the failure mode.

treg mcp install is deliberately not part of that path: it writes Claude Code, Cursor and opencode configs (mcp_install.py), and a dsh profile is none of those. treg mcp install prints that instruction when it detects ~/.dsh, and the dsh bootstrap steers away from it.

Two constraints on dsh/index.js

No build step. dsh plugin add github:... fetches sources, not build output. A TypeScript package would need a prepare script, and pnpm ≥10 refuses to run one from a git dependency until the user adds an allowBuilds entry to their profile's pnpm-workspace.yaml — i.e. the documented install command fails on first run for everyone. Plain hand-written ESM sidesteps it.

No @deepseek-ai/* dependency. The obvious implementation mounts the in-box @deepseek-ai/dsh-skill-filesystem with customSkillDirs, but an out-of-tree bundle depending on an in-box package installs a second copy that drifts from the host's. The provider's list/get contract is the stable seam, so index.js implements it directly — the shape dsh-skill-badge uses for its own bundled skill. tests/test_plugin.py pins both constraints.

The skill's description is parsed out of SKILL.md at runtime rather than restated in JS, so there is no second copy to rot.

Discovery

Plugins are found through the dsh-plugin GitHub topic, which is what the ecosystem catalogs (awesome-dsh-plugin, dsh.tools) sync from. The topic is set on this repo; a listing PR to those catalogs is the manual half.