Canonical skill and responsibilities

The complete canonical source is docs/skills/multidimensional-cube-harness/SKILL.md. Discover the release published by your server through its advertised MCP resources or call read_openepm_published_skill with part: manifest. Use the returned uri and manifestUri, and verify the complete payload against the manifest’s UTF-8 byte size and SHA-256 digest. The published skill reference describes resource and tool retrieval; this page does not create a second copy.

One harness, four responsibilities

Modeler and interviewer

Discover grain, meaning, decisions, ambiguity and missing evidence. Adapt questions to what the source and user have already established.

Target expert and implementer

Discover the live MCP contract, distinguish existing behavior from proposals, and submit only authorized typed operations.

Verifier

Reconcile source coverage, exact values, lineage, rules, permissions and durable outcomes. A successful sample is evidence for its stated slice, not whole-application acceptance.

Improvement steward

Capture evaluated lessons and product feedback with attribution, privacy and approval boundaries. A lesson does not silently change the active application or global skill.

These roles are phases of one client harness, not independent agents with competing state. They share the adopted skill version, approved baseline, source-to-cube contracts and recovery checkpoint.