add: Add Business Skills consumption (catalog at connect)

[suyask-msft]

Summary

Adds consumption support for Dataverse Business Skills (skill entity records holding org-specific domain instructions) without taxing the majority of environments that have none.

The agent detects once per session, at connect, whether the org defines any Business Skills. If it does, it loads a lightweight catalog (skillid, name, description) into context for the session and pulls a skill's full body on demand when a request matches one by description. If the org has none — the common case — the capability stays dormant: no per-operation queries, no added context cost.

What changed

• dv-connect — a step to detect and load the Business Skills catalog once per session, plus references/business-skills.md covering the model, the probe query, and how to consult a skill. (The --validate troubleshooting block was condensed to keep the skill within its token budget.)
• dv-data / dv-query — a short, conditional reminder to consult the already-loaded catalog when an operation is governed by a Business Skill. This is explicitly not a per-operation re-query.
• Version bump 1.6.0 → 1.6.1.

Scope

Consumption only. Authoring and management of Business Skills (create/update/delete, file resources, validation, publishing) is intentionally out of scope here.

Design notes

• Matching is done by the agent reading catalog descriptions, not by keyword search, so natural-language requests resolve to the right skill even when names differ.
• Loading the full catalog is cheap into the low hundreds of skills; beyond that, discovery should move to a server-side retrieval channel (future enhancement).

TODO — evaluations (follow-up)

Behavioral evals are required before this is considered validated; static_checks only covers structure. Planned as a business-skills.biceval.json ablation suite plus a regression pass:

• Efficacy (org with skills):
  • A governed request (e.g. "import refunds following our process") loads the catalog at connect and consults the matching skill's body.
  • A natural-language request resolves to the correct skill by description, not keyword.
  • An unrelated request injects no skill body (precision).
  • The catalog probe fires once per session, not per operation.
• Safety (org without skills, and existing scenarios):
  • No-skill orgs perform data/query operations with no skill-entity query and no added context — behavior identical to baseline (guards against re-introducing a per-op tax via the dv-data/dv-query reminders).
  • Re-run existing dv-data / dv-query suites on this branch vs baseline; no score regression on any test.
• Fixtures: requires a test environment seeded with sample Business Skills and a clean one with none (or mocked skill-table responses).
• Post-ship: confirm via telemetry that skill-entity read volume rises only in orgs that have skills, and that dv-data / dv-query latency is unchanged for orgs without.

🤖 Generated with Claude Code

[harishku]

Thanks for taking care of it

[suyask-msft]

Adding a comment on Jeff's behalf: Add a SHORT explanation here of how to get the header info both from MCP and CLI so the LLM doesn't have to load the full business-skills.md unless it wants to know more.