@yogsoth-ai/dare-dsh

DeepSeek Harness plugin for the De-Anthropocentric Research Engine — 920 research skills on ctx.skills, with an opt-in MCP server fleet.

Install

dsh is in developer preview and its latest npm tag lags behind next. Install the peers explicitly:

npm i @yogsoth-ai/dare-dsh \
      @deepseek-ai/dsh-skill@next \
      @deepseek-ai/dsh-skill-filesystem@next

The MCP fleet is optional. Add it only if you want the research servers:

npm i @deepseek-ai/dsh-mcp-client@next

Use

Copy cordis.example.yml and run:

npx @deepseek-ai/dsh web --patch ./cordis.example.yml

The skill row needs no credentials. Drop the dare-mcp row to run skills-only.

What you get

Skills. 920 research skills registered as one provider named dare, at bundled rank (600) — a project or user skill of the same name always wins, so this library never shadows your own. Discovery, SKILL.md parsing, and relative-resource resolution are delegated to @deepseek-ai/dsh-skill-filesystem; 385 skills reference a sibling prompt.md through the provider's directory resource base.

MCP fleet. Seven research servers, one dsh-mcp-client instance each. Tools arrive as mcp__<serverName>__<rawName>, matching the names four skills reference literally (mcp__semantic-scholar__relevanceSearch, mcp__wiki-vault__vault_search, and two more), so those skills work unmodified.

| Server | Credential | Notes | |---|---|---| | alphaxiv | — | HTTP | | keenable | — | HTTP | | semantic-scholar | SS_API_KEY | | | brave-search | BRAVE_API_KEY | | | tavily-search | TAVILY_API_KEY | | | wiki-vault | VAULT_ROOT | path to your vault | | apify | APIFY_TOKEN | |

A server missing its credential is skipped with a warning at load, not started empty. dsh's own default (failOnStartupError: false) would activate it with zero tools, leaving dependent skills silently inert with no visible cause.

Restrict the fleet with only or exclude:

- id: dare-mcp
  name: '@yogsoth-ai/dare-dsh/mcp'
  config:
    only: ['alphaxiv', 'semantic-scholar']

Catalog weight

dsh-tool-skill injects every model-invocable skill's name and description into a durable session catalog. All 920 are exposed, costing ~135 KB per request while registered. Budget for it, or mount a narrower skill root of your own.

One description exceeds the consumer's 500-character cap and loses its tail in the catalog (mechanism-gap-hunting); its body is unaffected.

Verification status

npm test runs the skill library against the real ctx.skills registry, the real filesystem provider, and a real Cordis context — no mocks. It asserts all 920 skills are discovered, the snapshot is complete, every skill is attributed to this provider at bundled rank, every body loads non-empty, and prompt.md resolves through resourceBase.

Not verified: behavior inside a live harness driving a model. That needs a DeepSeek API key, which this package's tests do not use. The registry contract is proven; the end-to-end model experience is not.

License

Apache-2.0