Data Operations

Data Operations

This skill owns Dataverse row create, read, update, upsert, and delete, plus query design and execution. It is the data branch of the powerplatform-core orchestrator — start there to discover repo context and confirm the target environment and solution. Keep data work separate from schema work: a solution import is not the way to sync environment-specific config data, and creating/changing tables or columns belongs in the dataverse-schema skill.

Default safety posture: create, read, update, and upsert are in scope when the user asks for them. Read is non-mutating and needs no preflight. Delete is supported but is the highest-risk write: only delete when the user explicitly asks, and only after the live-mutation preflight and explicit approval.

When to use this

• Create a single row, or create-or-update by primary key or alternate key.
• Read a row back by primary key or alternate key (verification, debugging, or "what's the current value").
• List or query rows with a FetchXML/spec query, with bounded paging and a count of how many rows match.
• Update only the columns that changed on a known record.
• Delete a row by primary key or alternate key (explicit request + preflight).
• Seed or sync reference/configuration data into an environment.
• Design a query and get the OData path, FetchXML, and Power Automate "List rows" parameters for a structured spec.

For table/column/lookup/form/view/ribbon/icon changes, use the dataverse-schema skill instead.

Helpers

The Python helpers live in the plugin's scripts/ directory at the plugin root, not inside this skill folder. Resolve the plugin root and invoke a helper like this (use $CODEX_PLUGIN_ROOT, also exposed as $PLUGIN_ROOT, when running under Codex; if installed standalone, the scripts/ folder sits beside this bundle):

• Claude Code: python "$CLAUDE_PLUGIN_ROOT/scripts/upsert_data.py" --table contoso_account --data ./row.json --key ./key.json --auth-dialog --verify
• Codex: python "$CODEX_PLUGIN_ROOT/scripts/upsert_data.py" --table contoso_account --data ./row.json --key ./key.json --auth-dialog --verify

• design_dataverse_query.py — offline query designer. Takes --spec (a JSON object or path describing tableLogicalName, entitySetName, select, filters, orderBy, top, expand) plus optional --repo-root / --output. Emits a Web API OData path, FetchXML, Power Automate "List rows" parameters, and warnings about missing filters, row limits, select lists, or entity-set names. No live call.
• upsert_data.py — live row write. Key flags: --table (logical name, required), --data (JSON object or path, required), --mode {create,update,upsert} (default upsert), --id (primary-key GUID for update/upsert), --key (JSON object or path with alternate-key values), --verify (retrieve the changed columns after the write), plus the shared auth flags (--environment-url/--target-url, --username, --tenant-id, --auth-dialog, --auto-validate, --auth-flow {auto,devicecode,interactive,clientsecret,certificate}, --app-id, --certificate-path, --force-prompt, --verbose) and --repo-root (resolves project-profile deployment defaults, including typed row-write coercion). --dry-run reads the current row and prints a per-column diff of the would-be write (changed/added/unchanged) without mutating anything; it needs --id or --key for update/upsert (create mode reports every column as new).
• read_data.py — live, read-only row read (no preflight). --mode retrieve reads one row by --id or --key (optional --columns / --all-columns); --mode list runs a query from --fetchxml or --spec (built with the offline designer) with --max-rows (default 100), --page-size, and --exact-total. The response reports returnedCount, moreRecords, and totalRecordCount (Dataverse caps that count at 5000; totalRecordCountLimitExceeded flags when the true total is higher). Shares the same auth flags as upsert_data.py.
• delete_data.py — live row delete by --id or --key (--table required). Mutating and irreversible: run the live-mutation preflight and get explicit approval first; prefer --auth-dialog to confirm the target. Shares the same auth flags.

If the repo already has a proven seed-data or data-sync mechanism, or server-side SDK code, prefer that repo-owned path over reconstructing one. The Web API is the simpler direct/scriptable path when no SDK code exists.

Unattended (service-principal) auth. For CI/CD, scheduled agents, or headless verification, read_data.py/upsert_data.py/delete_data.py accept --auth-flow clientsecret (with --app-id and --tenant-id, client secret read from the DATAVERSE_CLIENT_SECRET environment variable) or --auth-flow certificate (--app-id, --tenant-id, --certificate-path, password from DATAVERSE_CERTIFICATE_PASSWORD). Secrets come from environment variables only — never pass a secret or certificate password as a CLI argument. Service-principal mode skips the auth dialog and PAC-profile defaulting, so pass --environment-url explicitly. The application user must exist in the target environment (see the platform-administration skill for pac admin create-service-principal).

Dataverse MCP interop. If Microsoft's Dataverse MCP is connected, prefer it for quick interactive "what does this data look like" lookups (search/describe/read_query). Use these helpers whenever the read must be paged (read_data.py --mode list returns returnedCount/moreRecords/totalRecordCount), keyed and verified, or followed by a write — and route every live mutation through the helpers so it passes the preflight gate, never through the MCP. When an MCP read_query result is truncated by its row cap or its SQL subset rejects a query, fall back to design_dataverse_query.py + read_data.py. See the orchestrator skill for the full routing rules.

Workflow

1. Confirm targeting before writing. Resolve the target environment, the table logical name (and entity set name when querying), whether the operation is create-only, update-only, or create-or-update, the matching key (record ID, alternate key, or agreed lookup), the exact columns to set, and duplicate-handling expectations. If ambiguous, ask only for the missing high-risk parts.
2. Choose the mode deliberately. Use create to let Dataverse assign the key; update to send only changed columns by primary or alternate key; upsert for create-or-update when the row might already exist. For Web API update-only behavior, If-Match: * prevents a silent upsert.
3. Key choice for rerunnable work. For integration/migration/rerunnable writes, prefer alternate keys over user-editable business fields, and prefer immutable keys. When the source identifier comes from another system, do not rely on a bare reused ID unless it is globally unique in the tenant — prefer a source-system-qualified key (source_system + source_id).
4. Resolve lookups and choices safely. Bind lookups using supported references, not display-name guessing; use logical names and valid choice values; reuse early-bound classes, enums, or constants when the repo already has them; prefer reference tables or global choices over repeated free text for controlled value sets.
5. Plan before bulk/config apply. For config-data sync, imports, migrations, or bulk corrections, run a dry-run/diff/plan phase first that names rows to be created, updated, skipped, invalid, and any target-only rows present in Dataverse but missing from the source extract (that last category drives whether deactivation, deletion, or an explicit no-op policy is needed). Do not apply config data to TST/TEST until that plan exists.
6. Apply and verify. Validate the smallest possible payload, use --verify when confirming the write matters, and capture whether the action created or updated the row.
7. Preserve provenance. When syncing from another source, keep enough provenance to identify origin, retain source-system-qualified keys when multiple upstreams may reuse an external ID, and avoid casually overwriting provenance columns used for reconciliation or audit.

Key safety and decision rules

• Run the live-mutation preflight defined in the powerplatform-core orchestrator before any live row write or delete. Reads (read_data.py) are non-mutating and skip the preflight. Stop if any required field is missing.
• Prefer the auth dialog (--auth-dialog) so the user confirms the target URL and selects the working solution; warn when the requested target does not match the active PAC profile. Treat DEV as the working environment and TEST as validation. Do not touch production.
• Capability boundary: upsert_data.py performs live row writes and optional post-write verification, plus a keyed --dry-run that diffs a single would-be update/upsert against the current row (no mutation); read_data.py reads (retrieve/list) and is non-mutating; delete_data.py deletes a single row. There is no bulk dry-run/plan mode across many rows, and read_data.py list returns at most --max-rows rows per call (default 100) with a moreRecords flag rather than streaming an entire table. For multi-row plan/apply (config sync, migrations), keep that planning phase in the workflow or repo-specific tooling layer; do not imply the shared helpers already batch it.
• Do not delete business data unless the user explicitly asked for deletion and separately approved it. Prefer safe create, keyed update, or upsert over any delete path.
• Separate solution customization from configuration rows. A solution import is not the default way to sync environment-specific config data.
• If no write executed because a required key or environment value was missing, say so explicitly. When a write is performed, report the table targeted, whether it created/updated/upserted, the key or alternate key used, which columns changed, and what safety checks were applied.

Query design notes

• Prefer the simplest query that matches the requirement: use logical names (not display names), select only needed columns, set row limits and pagination intentionally, and guard against nulls and empty arrays.
• Use OData ($select, then $filter only for necessary predicates, $expand sparingly) for straightforward server-side filtering. Dataverse Web API does not support every OData option; if the query needs complex joins, aggregates, or view definitions, prefer FetchXML.
• If the entity set name is unknown, provide it explicitly in the helper input rather than guessing a pluralized logical name.
• For Power Automate expressions, prefer clear guards (coalesce, if, empty) over nested fragile expressions. Validate in filters against the specific connector action, falling back to FetchXML if rejected.

References

Read the shared references at the plugin's references/ directory for depth:

• references/data-operations.md — create/update/upsert paths, keying and provenance, dry-run/apply planning, and validation/reporting.
• references/queries-and-xml.md — OData and FetchXML guidance, view FetchXml/LayoutXml, and direct solution XML rules.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: satriotsubasa
• Source: satriotsubasa/PowerPlatform-Core
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.