dsh-dolphindb

DolphinDB integration for DeepSeek Harness as an installable profile bundle. One package ships the WebSocket executor, the model-facing query/write tools, and the 23 bundled DolphinDB skills.

What you get

Component	Surface
dolphindb_query	Read-only DolphinDB scripts/SQL (result tables, vectors, matrices, scalars; optional chart; optional server override)
dolphindb_execute	Write/mutation scripts, gated behind the approval seam with a code preview (optional server override)
dolphindb_switch	Switch the active server environment; persisted in the dolphindb settings section and hot-applied
23 DolphinDB skills	Bundled SKILL.md trees (syntax, SQL, vectorization, functional programming, data import, factors, ops, Orca DStream, plotting, backtest, FICC pricing, Tushare/CSMAR, …) exposed through the skill tool

Install

dsh plugin --profile <name> add github:tradercjz/dsh-dolphindb

or add @tradercjz/dsh-dolphindb to a profile's dsh.profile.bundles list after @deepseek-ai/dsh-base.

Servers and environments

The executor runs a named server registry. The bundle ships one local server (127.0.0.1:8848 / admin, credential ref DOLPHINDB_PASSWORD) as the deployment default. Real deployments live in the dolphindb settings section — $DSH_HOME/settings.yaml, hot-reloaded, and editable from web Settings → Plugins once @tradercjz/dsh-client-dolphindb is installed:

dolphindb:
  active: prod
  servers:
    prod:
      host: 183.134.101.139
      port: 8892
      username: admin
      passwordRef: DOLPHINDB_PROD_PASSWORD
    test:
      host: 183.134.101.139
      port: 30006
      username: admin
      passwordRef: DOLPHINDB_TEST_PASSWORD

Each password is resolved per connection through the credentials seam: set the referenced name in the process environment / managed credential store, never the value in YAML. dolphindb_switch changes active (persisted); the server argument on dolphindb_query / dolphindb_execute overrides per call. A profile patch may instead replace the executor row's whole config for a fixed deployment.

Verify the layer

dsh --profile <name> --dump-config

The output shows a # == @tradercjz/dsh-dolphindb layer mounting three rows — dolphindb (executor), tool-dolphindb (tools), dolphin-skills (skills). Starting the profile registers dolphindb_query / dolphindb_execute / dolphindb_switch and exposes the 23 dolphindb-* skills through the skill tool.

Development

The repo type-checks and builds against a sibling deepseek-harness checkout (see tsconfig.json):

pnpm install
pnpm run typecheck   # against ../deepseek-harness
pnpm run build       # tsc -b + tsdown → lib/

dsh plugin add github:… installs source, so pnpm runs prepare (a self-contained tsdown transpile with no harness checkout and no type check). The first install fails until the profile's pnpm-workspace.yaml authorizes the build scripts (the dsh error prints the exact commit-pinned key — prefer that over the loose form below):

allowBuilds:
  "@tradercjz/dsh-dolphindb@https://codeload.github.com/tradercjz/dsh-dolphindb/tar.gz/<commit-sha>": true
  # Native file-watcher pulled transitively by the `dolphindb` client
  # (dolphindb → xshell → sass → @parcel/watcher); set `false` to skip the
  # native build and keep the JS fallback.
  "@parcel/watcher": true

Skills provenance

The assets/skills/ trees are vendored from the official dolphindb/DolphinX_Skill repository. One source SKILL.md (dolphindb-daily-factor) had an unquoted description that breaks YAML parsing; the vendored copy quotes it.

License

MIT