discover-shennong-data

Files (2)

• SKILL.md
• shennong.skill.yaml

SKILL.md (first lines)

# Discover Shennong data

ShennongData is preinstalled in the Runtime image together with its bounded
read-only MCP server. In Agent conversation, the governed OS `db.*` tools remain
the authority; in reproducible R scripts use the equivalent `ShennongData::`
API and keep materialization explicit.

1. Call `db.discover_resources` with broad disease, cohort, assay, or modality
   terms. An empty narrow search is not proof that the biological entity is
   absent.
2. Call `db.inspect_resource` for each plausible Resource. Confirm organism,
   assay, build, annotation release, measurement/transformation, sparse and
   implicit-zero semantics, identifiers, cohort axes, declared operations,
   version, license, and analysis readiness.
3. Use `db.query_resource` only with a declared operation, exact context labels,
   and the smallest useful row limit. Resolve stable identifiers before joining
   Resources; do not use gene symbols as cross-Resource keys.
4. In R, use `ShennongData::sn_connect()` and
   `ShennongData::sn_load_data(resource = ...)`, select the declared assay/layer,
   inspect the lazy query and provenance, then call `ShennongData::collect()`
   only when bounded materialization is required. Pass the resulting matrix,
   path, or Seurat object to `Shennong`.
5. The package MCP sequence is compatibility, discovery, inspection, identifier
   resolution, query planning, then bounded fetch. Its tools are discovery-only
   and cannot grant credentials, mutate data, or execute R code.
6. Call `db.get_provenance` for every Resource used in a result. Cite only
   EvidenceRef identifiers returned by governed tools and state truncation,
   missing annotations, and permission boundaries.

Do not treat an empty catalog search as proof that the biological feature is absent.

From zerostwo/shennong-os, Apache-2.0.

More from zerostwo/shennong-os