feat: add dv-skill for Dataverse Business Skills management#64
Conversation
|
@harishku please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.
Contributor License AgreementContribution License AgreementThis Contribution License Agreement (“Agreement”) is agreed to by the party signing below (“You”),
|
|
Can we please see the eval results as part of the PR ? |
aadharshkannan
left a comment
There was a problem hiding this comment.
It doesn't seem like we have any evals run for these changes. Can we please get the evals ?
|
|
||
| ## Consult Business Skills — Mandatory | ||
|
|
||
| **Before generating any Python SDK code, consult Dataverse Business Skills.** Extract keywords from the user's request, query the skill entity, and incorporate any matching skill instructions as context. If no skills match, proceed normally. See **dv-skill** § Consult Helper for the full pattern. |
There was a problem hiding this comment.
Can we have evals for these edits?
| ### List All Skills | ||
|
|
||
| ```python | ||
| skills = client.records.get("skill", select="name,uniquename,description,body") |
There was a problem hiding this comment.
What happens in a non-python interface ? If I use CLI as an example
There was a problem hiding this comment.
Genuinely useful capability — org-specific grounding for data ops is valuable. My feedback is about how it loads (it adds a mandatory per-op step to the hottest skills) and packaging — not the feature itself. Two things, details inline:
- (design) Should the consult be mandatory per-op, and does this need to be its own skill? — the main one. (Load it once per session, relevance-filtered: that's better for every org — whether they have 0 Business Skills or 500 — not just the ones without.)
- (footgun)
.gitignoreis shadowing the plugin's ownscripts/.
Also: the dv-overview refactor (+8/-33, new tool-capabilities.md) and the dv_data eval rewrite (+142/-31) read as separate changes bundled in — easier to review split out, and worth confirming the dv_data eval isn't being loosened just to accept the new mandatory consult.
|
|
||
| **If MCP tools are available** (`create_record`, `update_record`) and the task is ≤10 records, **use MCP directly — no script needed.** Only write a Python script when the task requires: bulk operations (10+ records), data transformation, retry logic, CSV import, or operations the SDK supports that MCP cannot (upsert, file uploads). Sequential MCP tool calls are not "multi-step logic" — use MCP for those. | ||
|
|
||
| ## Consult Business Skills — Mandatory |
There was a problem hiding this comment.
Two concerns, and these are the main ones:
(a) This shouldn't be mandatory per-op. Very few orgs have Business Skills configured, so a mandatory skill-entity query before every dv-data (and dv-query) op taxes the 99% for the 1% — an extra round-trip + context on the hottest paths, usually returning nothing. (It also cuts against the de-steering direction in the design repo.)
Better: gate it on a cheap, once-per-session existence check — at dv-connect, detect whether the org has any active skill records and cache it; if none, this stays dormant (zero per-op cost). Better long-term: surface org guidance through the MCP server's instructions channel (the initialize handshake already returns per-org instructions) — loaded once, server-side, opt-in, no skill-prose coupling.
(b) Does this need to be a separate skill? The management half (CRUD on skill/skillresource + file upload) is just dv-data on one more entity — it could fold into dv-data as a short reference rather than a new always-loaded skill (which adds to Level-1 frontmatter for everyone). The "consult" half is a cross-cutting behavior, not a skill.
(Aside on naming: dv-skill is itself confusing — it collides with the plugin's own Agent Skills and this file's own "Business Skills" title; dv-business-skills would read better. But that's moot if we don't make it a separate skill.)
Future-proof either way: build the consult relevance-filtered + once-per-session now and it scales for free — if adoption surges it just starts returning results, no load-path change. And "separate skill" is a feature-depth call, not an adoption-count one — revisit a dedicated dv-business-skills skill only if Business Skills grow richer than CRUD.
Net: keep the capability, change the load model — orgs that use Business Skills get the grounding; everyone else pays nothing.
| # Local-only artifacts from running evals against a personal Dataverse org | ||
| .mcp.json | ||
| .env | ||
| scripts/ |
There was a problem hiding this comment.
scripts/ is unanchored, so it matches .github/plugins/dataverse/scripts/ — which holds committed plugin code (auth.py, enable-mcp-client.py, mcp_proxy.py; this PR even edits auth.py). Already-tracked files stay tracked, but new files under any scripts/ would be silently ignored and hidden from git status — a footgun. Same risk with evals/*.py. Suggest anchoring to the specific local path (e.g. a root-level /scripts/ or the personal eval folder), not a blanket scripts/.
Adds a new dv-skill skill that teaches the agent how to manage Dataverse Business Skills (the skill entity) - list, create, update, delete skills and skill resources via the Python SDK.
Changes