DeepSeek Harness (dsh) connector / plugin
connector/dsh/ is the DeepSeek Harness (dsh) flavor of this package, providing
an automated Cordis plugin and connector bundle for DeepSeek Harness
(CLI, TUI, and Web UI).
Bundle contents
A single biomcp stdio MCP server (pinned biomcp@1.6.0, 120 s connection
timeout, automatic China npm mirror detection via Intl.DateTimeFormat or env)
plus four bundled skills and the bioresearcher-dr-worker subagent prompt:
| Bundled | Not bundled |
|---|
| bioresearcher-deep-research | bioresearcher-onboard |
| bioresearcher-plot-making | |
| bioresearcher-pubmed-weekly | |
| bioresearcher-python-setup-uv | |
bioresearcher-onboard is excluded on purpose: its purpose — installing and
registering the biomcp server in harness configs — is performed automatically by
the plugin's apply hook on startup.
The bundled skill list is defined in connector/dsh/skill-bundle.json.
Architecture & runtime behavior
The plugin entry point (index.js) is an ESM module exporting a Cordis plugin:
export const name = "bioresearcher";
export const inject = ["tools", "skills"];
export async function apply(ctx, config) { ... }
When booted in a dsh profile (e.g. web, headless, tui), the plugin performs:
- Automatic MCP Server Registration:
Dynamically mounts
@deepseek-ai/dsh-mcp-client with serverName: "biomcp",
launching npx -y -p biomcp@1.6.0 biomcp over stdio with timeout 120000 ms.
DeepSeek Harness automatically registers the tools under the mcp__biomcp__<tool>
namespace (e.g. mcp__biomcp__article_search).
- Dynamic Skills Discovery:
Reads the packaged
skills/ directory and registers bundled skills on
ctx.skills.register(...) with source: "bioresearcher" and resourceBase
pointing to the local skill directory so relative paths in instructions
resolve cleanly.
- Subagent Worker Provisioning:
If an agent roster service (
ctx.agents) is present, registers bioresearcher-dr-worker
with the specialized prompt (with ${CLAUDE_PLUGIN_ROOT} replaced by the
installed plugin root). If absent, bioresearcher-deep-research transparently
uses generic subagent delegation (Tier B) or sequential execution (Tier C).
Build
node scripts/ci/build-connector-dsh.mjs # dist/
node scripts/ci/build-connector-dsh.mjs --out DIR
Stages dist/bioresearcher/ (root directory inside the tarball) and writes a
reproducible dist/bioresearcher-connector_dsh-v<VERSION>.tar.gz (GNU tar
--sort=name --mtime=@0 --owner=0 --group=0 --numeric-owner piped through
gzip -n -9). CI runs this build script as a gate (.github/workflows/ci.yml),
and the release workflow attaches the tarball to every GitHub release.
Version policy
connector-meta.json and package.json version must equal the repo VERSION
(Series 1, manifest-governed via scripts/ci/version-coupling.json and enforced by
scripts/ci/check-drift.mjs). Release PRs bump them in unison with VERSION.
Installation & usage
Users can install the release archive through either of two methods:
Method A: Profile plugin installation (Recommended)
Install the package bundle directly into the target profile (e.g. web, headless, or custom profile):
dsh plugin --profile web add /path/to/bioresearcher-connector_dsh-v<VERSION>.tar.gz
dsh installs the package into the profile environment and automatically mounts the bundled cordis.patch.yml.
Method B: Standalone overlay patch
Extract the archive and load it directly via an overlay patch specifying the local package root:
tar -xzf bioresearcher-connector_dsh-v<VERSION>.tar.gz
dsh --profile web --patch <(echo "- insert: [{ id: bioresearcher, name: $(pwd)/bioresearcher/index.js }]")
Or add the entry to $DSH_HOME/cordis.patch.yml (~/.dsh/cordis.patch.yml)
to enable it across all profiles on the machine.
