dv-overview

Skill: Overview — What to Use and When

Load this skill first for any Dataverse work — it holds the cross-cutting context every task needs: scope, the tool-capability reference, the hard rules, and the change lifecycle. It does not route; the agent auto-selects specialist skills via their own WHEN/DO NOT USE WHEN frontmatter triggers. Users describe what they want in plain English; the agent chains skills automatically and never asks the user to name a skill or command.

What This Plugin Covers

Dataverse / Power Platform work for every persona — builders and agent devs, data scientists, environment admins, and business users — delivered by specialist skills. The agent loads and routes to these automatically via their frontmatter triggers — you never invoke them by name.

Model-driven apps: the building blocks (tables, forms, views) are covered by dv-metadata; composing the app shell itself — site map and navigation — is not yet a first-class skill.

Out of scope:

• Canvas apps — a different technology; use pac canvas or the maker portal
• Power Automate flows — use the maker portal or the Power Automate Management API
• Azure infrastructure beyond what's needed for service-principal setup
• Business Central or other Dynamics products

Hard Rules — Read These First

The safety rules (init state, auth, environment confirmation, no bespoke MSAL) are non-negotiable. The tool-selection guidance (Rules 1, 2, 4) is capability-based — strong defaults, not rigid mandates.

0. Check Init State Before Anything Else

Before writing ANY code or creating ANY files, check if the workspace is initialized:

ls .env scripts/auth.py 2>/dev/null

• If BOTH exist: workspace is initialized. Proceed to the relevant task.
• If EITHER is missing: Automatically run the connect flow (see the dv-connect skill). Do NOT ask the user whether to initialize — just do it. Do not create your own .env, requirements.txt, .env.example, or auth scripts. The dv-connect skill handles all of this.

Do NOT create requirements.txt, .env.example, or scaffold files manually. The connect flow produces the correct file structure. Skipping it is the #1 cause of broken setups.

1. Python for scripting; the CLIs and MCP are first-class

Python is the language for automation logic (transformation, control flow, retry, CSV). The toolchain (scripts/auth.py, the SDK, skill examples) is Python-based. But MCP tools, the Dataverse CLI (dataverse), the Python SDK, and the PAC CLI (pac) are all first-class tool invocations — use whichever fits. The Dataverse CLI has the same standing as pac, which is invoked freely across the solution and metadata skills.

NEVER:

• Write automation logic in JavaScript/TypeScript/Node.js (npm, yarn, pnpm, package.json, node_modules/)
• Use @azure/msal-node, @azure/identity, or any Node.js Azure SDK
• Implement a bespoke MSAL / device-code flow — auth is scripts/auth.py, pac auth, and the Dataverse CLI

ALWAYS:

• Use pip install and the Python SDK (PowerPlatform-Dataverse-Client) for data and schema logic
• Use scripts/auth.py for tokens/credentials; azure-identity (Python) for Azure credential flows
• Treat the Dataverse CLI (dataverse) and pac as allowed first-party CLIs

2. Pick the surface that fits — capability awareness, not a fixed order

No mandated tool order. Each surface has a capability profile; pick what fits the job and the surface you are already in — soft defaults, not a required sequence. The full matrix is in Tool Capabilities below; the principles:

• Prefer a managed surface (MCP, the Dataverse CLI, or the SDK) over hand-rolled raw OData — they carry auth, paging, retry, and geo routing that raw HTTP re-implements. When MCP can't handle it (bulk >25 records, large reads, advanced schema like forms/views/N:N relationships/global option sets/alternate keys, multi-step workflows, analytics, or MCP isn't available), the Python SDK is the default.
• Raw Web API is the last-resort escape hatch for surfaces with no managed path (unbound actions like PublishXml, global option sets, anything without a first-class SDK/CLI command) — and even then prefer dataverse api (managed auth, exit codes) over hand-rolled urllib/get_token. Forms/views are not raw-only (SDK records.create/update on systemform/savedquery; only PublishXml needs dataverse api). Aggregation/N:N joins aren't raw-only either: client.query.fetchxml() (aggregates + link-entity), or the CLI's data associate for N:N writes.
• If an SDK method fails or a PAC command seems missing, check the relevant skill before hand-rolling raw HTTP.

Field casing: $select/$filter use lowercase logical names (new_name). $expand and @odata.bind use Navigation Property Names that are case-sensitive and must match $metadata (e.g., new_AccountId). Getting this wrong causes 400 errors. SDK record payloads: provide the correct SchemaName casing on @odata.bind keys (e.g., new_AccountId@odata.bind); the SDK does not auto-correct wrong casing. Raw Web API calls (forms, views, metadata): casing is entirely manual — a lowercase new_accountid@odata.bind will 400.

Publisher prefix: Never hardcode a prefix (especially new); query existing publishers and ask the user. The prefix is permanent. See the solution skill's publisher discovery flow.

3. Use Documented Auth Patterns

Three entry points, one shared sign-in:

• dataverse auth create (Dataverse CLI) writes a shared MSAL token cache under the DataverseCLI app registration. That single sign-in serves the CLI, the @microsoft/dataverse MCP proxy, and scripts/auth.py — which silently reuses the same cache via msal-extensions (the sanctioned MSAL API, not raw-file parsing), so Python scripts and the SDK don't prompt again.
• scripts/auth.py is the auth entry point for all Python/SDK code. Its order: service principal (CLIENT_ID + CLIENT_SECRET in .env) → shared Dataverse CLI cache → device-code fallback. Use get_client(skill) (SDK) or get_plugin_headers(skill, get_token()) (raw Web API) — both stamp the skill attribution; bare get_token() does not.
• pac auth create (PAC CLI) authenticates pac for dv-solution and dv-admin.

Telemetry attribution (keep it deterministic): every request carries a closed-schema app=dataverse-skills/<ver>;skill=<skill>;agent=<agent> context so the server sees which skill routed each OData call. It is baked in — get_client(skill) and get_plugin_headers(skill, ...) stamp it on the SDK and raw-HTTP paths; the Dataverse CLI auto-stamps DataverseCli/<ver> + the command, and you add the skill with --context "app=dataverse-skills/<ver>;skill=<skill>;agent=<agent>" (the CLI wraps it in parentheses itself — do not pre-wrap). Never modify, omit, or free-form this context — it is a closed schema (allowlisted skill/agent, no PII).

NEVER:

• Read or parse raw token cache files (e.g., tokencache_msalv3.dat) — reuse the cache only through scripts/auth.py / msal-extensions
• Implement your own MSAL device-code flow
• Hard-code tokens or credentials in scripts
• Invent a new auth mechanism

If auth is expired or missing, re-run dataverse auth create (or pac auth create), or check .env. See the dv-connect skill.

4. Be honest about gaps — don't hallucinate

Each skill documents a tested sequence — follow it when it fits. The skills are the source of truth for the supported, non-deprecated API. If a call fails with AttributeError, the installed SDK version may not have it — check the skill's version note and use the documented alternative.

The honesty guard: if you hit a gap the skills don't cover, say so and suggest a workaround. Do not hallucinate an unsupported path — do not invent a method, parameter, or endpoint that isn't documented. If unsure, say so.

Connectivity is not auth. A login.microsoftonline.com token can succeed while the org's data-plane domain is unreachable (restricted-egress hosts like ChatGPT Work Mode). Never report a count or result you didn't get from a real call that returned — verify with python scripts/auth.py --check; if it fails, say "unreachable," never a fabricated number. On a constrained host, lead with the SDK, not CLI/MCP. See dv-connect/references/headless-hosts.md.

Tool Capabilities — Which Tool for Which Job

Understanding the real limits of each tool prevents hallucinated paths. This is the one piece of context no individual skill owns.

Routing: the table shows what each surface does; the how to choose principle (soft defaults, not a fixed order) is Hard Rule 2. MCP tools not in your list? Load dv-connect.

Volume guidance: MCP for up to ~25 records per call or simple filters; the SDK's CreateMultiple for larger bulk writes (chunk large sets starting ~1,000 — see dv-data) and dv-query for bulk reads; Web API for $apply aggregation.

SDK method cheat-sheet (anti-hallucination, not a preference signal): SDK method names are the least discoverable surface, so agents invent them. This maps common ops to the exact call. Each op is equally reachable via MCP/CLI per Hard Rule 2; see the noted skill for the full pattern.

MCP Availability Check

If the user's request involves MCP — either explicitly ("connect via MCP", "use MCP", "query via MCP") or implicitly (conversational data queries where MCP would be the natural tool) — check whether Dataverse MCP tools are available in your current tool list (e.g., search, describe, read_query, create_record).

If MCP tools are NOT available and the user explicitly asked for MCP (e.g., "use MCP to query", "why isn't MCP working"):

1. Do NOT silently fall back to the Python SDK or Web API
2. Tell the user: "Dataverse MCP tools aren't configured in this session yet."
3. Load the dv-connect skill to set up the MCP server
4. After MCP is configured, stop here — the session must be restarted for MCP tools to appear. Remind the user to resume the session without losing context (Claude Code: claude --continue; Cursor: reload the window with Ctrl+Shift+P → "Developer: Reload Window"; Copilot: reopen the Copilot panel). Do not proceed with SDK. Wait for the user to restart.

If MCP tools are NOT available and the user asked a data question without explicitly requesting MCP (e.g., "how many accounts with 'jeff'?", "show me open tickets"):

1. This is a SDK fallback case — use the Python SDK to answer the question. Do not block the user.
2. After answering, offer: "MCP would handle this conversationally — want me to set it up?"

The distinction matters: explicit MCP request → block and set up MCP; implicit question → answer with SDK, offer MCP setup.

If MCP tools ARE available, prefer MCP for simple reads/queries/small CRUD. Use the SDK only when a script is needed.

The Change Lifecycle — Operate Safely

For any real change, walk these three steps in order: confirm where, confirm the container, then persist the result.

Step 1 — Confirm the Environment (MANDATORY)

Dataverse work often spans multiple environments (dev, test, staging, prod) and multiple sets of credentials. Never assume the active PAC auth profile, values in .env, or anything from memory or a previous session reflects the correct target for the current task.

Before the FIRST operation that touches a specific environment — creating a table, deploying a plugin, pushing a solution, inserting data — you MUST:

1. Show the user the environment URL you intend to use
2. Ask them to confirm it is correct
3. Run pac org who to verify the active connection matches

"I'm about to make changes to <URL>. Is this the correct target environment?"

Do not proceed until the user explicitly confirms. This is the single most important safety check in the plugin. Skipping it risks making irreversible changes to the wrong environment. Once confirmed for a session, you do not need to re-confirm for every subsequent operation in the same session against the same environment.

Step 2 — Confirm the Solution (before any metadata change)

Before creating tables, columns, or other metadata, ensure a solution exists to contain the work:

1. Ask the user: "What solution should these components go into?"
2. If a solution name is in .env (SOLUTION_NAME), confirm it with the user
3. If no solution exists yet, load the dv-solution skill and follow its publisher discovery + solution creation flow. Use the SDK — never raw Web API — to create publisher and solution records:

# Quick reference — full pattern with publisher discovery is in dv-solution
publisher_id = client.records.create("publisher", {
    "uniquename": "<name>", "friendlyname": "<display>",
    "customizationprefix": "<prefix>", "description": "<desc>",
})
solution_id = client.records.create("solution", {
    "uniquename": "<Name>", "friendlyname": "<Display>",
    "version": "1.0.0.0",
    "publisherid@odata.bind": f"/publishers({publisher_id})",
})

4. Pass solution="<UniqueName>" on all SDK calls, or include "MSCRM.SolutionName": "<UniqueName>" header on raw Web API metadata calls.

Creating metadata without a solution means it exists only in the default solution and cannot be cleanly exported or deployed. Always solution-first.

Step 3 — Pull to Repo (MANDATORY)

Any time you make a metadata change (via MCP, Web API, or the maker portal), you must end the session by pulling:

pac solution export --name <SOLUTION_NAME> --path ./solutions/<SOLUTION_NAME>.zip --managed false
pac solution unpack --zipfile ./solutions/<SOLUTION_NAME>.zip --folder ./solutions/<SOLUTION_NAME>
rm ./solutions/<SOLUTION_NAME>.zip
git add ./solutions/<SOLUTION_NAME>
git commit -m "feat: <description>"
git push

The repo is always the source of truth.

Scripts

The plugin ships scripts/auth.py (Azure Identity token/credential acquisition — used by all other scripts and the SDK). Any Web API call beyond a one-off query should be a Python script committed to /scripts/, using scripts/auth.py for tokens. For writes see dv-data; queries and analytics see dv-query; post-import validation see dv-solution.

Windows Scripting

Platform-specific shell rules (ASCII in .py, no multiline python -c, PAC PowerShell wrapper, unbuffered background output) live in references/windows-scripting.md. Read it when running on Windows.