Governing the Fabric estate
The OneLake catalog's Govern tab reports a tenant's governance
posture across three areas — Manage your data estate, Protect, secure
& comply, and Discover, trust, and reuse. It also hosts governance
experiences, Policies (preview) among them (§3, Policy sets),
and administrative ones
— capacities, workspaces, tenant settings, Azure connections, workloads
and more — each shown as far as the viewer's admin role allows (Learn,
2026-10-06). Govern itself went generally available in September 2026,
rolling out by region, with the Admin portal standing in where it has
not arrived; What's New keeps its February 2025 preview row too (Learn,
2026-10-13). This skill does the same work through the REST APIs, where
it can be counted, scheduled and fixed in bulk.
The output is findings, not a rewrite. Do not write to a tenant
unless asked separately, and then only through §6.
Four neighbours own the adjacent ground:
- Catalog search — finding an item across workspaces — is the
Explore tab. That is
fab find and POST /v1/catalog/search, both
rest-api's.
- Access — workspace roles, item permissions, OneLake security, and
the shape of the
sensitivityLabel field including resolving its GUID
through Graph — is security. This skill measures label
coverage; that one explains what a label does.
- REST mechanics — long-running operations,
continuationToken
pagination, Retry-After and 429 handling — is rest-api.
- Curating item metadata, and Microsoft's own take on this ground,
is the Govern Skills the Fabric Updates Blog announced on 2026-10-28,
in microsoft/skills-for-fabric
(MIT). There
onelake-catalog-cli (read at v0.3.20, 2026-11-02) runs
the same tier-by-action grid as this skill, and adds a curate mode:
an owner's item descriptions, display names and tags, and table and
column descriptions through the OneLake table API. This skill keeps
its own ground and vendors none of it.
Everything below was drilled from Microsoft Learn on 2026-09-11, bar
the Policy sets material (2026-11-02), and
none of it has been exercised against a tenant from this machine.
Treat it as first-party documentation, not as verified practice.
Past it, its references/, then Learn, by microsoft-learn-mcp
(microsoft_docs_search, then microsoft_docs_fetch), or with no server, curl -sL -H 'Accept: text/markdown' <page URL> (undocumented, read 2026-11-07). A live page
that contradicts a line here wins: say so, act on the page, and route
the fix by capture-learning, bar a dated measured line that names the
page it overrules and the call that measured it, which a page changed since
its date sends back to re-measure. An observed claim is cited or marked
pending a doc check, and stands, marked observed, only where no page covers
it.
1. Establish the caller's tier before anything else
Tier is the API surface the caller can reach, not their job title.
Getting this wrong wastes a whole audit: the endpoints that answer
"across the tenant" refuse everyone who is not a Fabric administrator,
and no amount of retrying changes that.
| Tier |
Surface |
Scope |
| Fabric administrator |
/v1/admin/*, Power BI admin (scanner, activity events) |
The whole tenant |
| Data owner |
Core API (/v1/workspaces, /v1/items, /v1/domains), Power BI per-item |
Only objects the caller holds a role on |
A domain admin, capacity admin or workspace admin who is not also a
Fabric administrator is a data owner here. The admin endpoints say so
in as many words — "The caller must be a Fabric administrator" — and
answer InsufficientPrivileges, "The caller doesn't have permissions to
call the API", to everyone else.
A 403 is the answer, not an obstacle. Report the tier boundary and
route the finding to someone who can act (§7). Never go around it by
trying the same read through another surface.
The domain-admin gap — say this up front
A domain admin who is not a Fabric administrator has no API that
returns their domain's posture. GET /v1/admin/domains/{id}/workspaces
requires the Fabric administrator role, and the Core API has no
equivalent: GET /v1/domains lists domains with only id,
displayName, description and parentDomainId — no membership, and
not even the defaultLabelId the admin version carries.
So a data-owner audit covers workspaces the caller can reach, never
"my whole domain". State that limitation before running it, rather than
producing a number that quietly means something narrower than asked.
Service principals are a separate gate
- Admin reads and most admin writes accept a service principal or
managed identity, but only once a tenant admin enables it. There are
two independent switches — one for read-only admin APIs, one for
the admin APIs used for updates — each limited to a named security
group.
- Bulk label writes reject them outright.
bulkSetLabels and
bulkRemoveLabels support User only; a service principal cannot set
a sensitivity label at all (§6).
- Item ownership takeover likewise "doesn't cover ownership takeover as
a service principal".
fab reaches the admin tier, not the data-owner one
The Fabric CLI's domain verbs — ls, get, mkdir, rm, set,
assign, unassign against .domains — and every fab label
command require tenant-level Fabric Administrator privileges, per the
CLI's own docs. fab label also reads label names from a local JSON
file registered with config set local_definition_labels.
A data owner assigning their own workspace to a domain therefore cannot
use fab assign .domains/..., even though the Core API would allow it.
Use the REST passthrough instead:
fab api -X post "workspaces/$WS_ID/assignToDomain" -i '{"domainId":"<id>"}'
See rest-api for fab syntax and token audiences: its
references/cli.md and references/auth.md.
2. Audit before you remediate
Establish the current state first, in the caller's own tier. An audit
mode does not write, even when asked mid-run: if the answer to a finding
is a write, say so, switch deliberately, and apply §6's gate. Do not
improvise a write inside an audit.
3. What to measure, and where it actually comes from
The trap in this whole skill, stated once:
GET /v1/admin/items carries no sensitivity label, no endorsement and
no domain. Its item object is id, type, name, description,
state, lastUpdatedDate, workspaceId, capacityId,
creatorPrincipal, defaultIdentity, tags and folderId — nothing
else. An "unlabeled items" count taken from it is not a finding; it is
the API's silence. Labels and endorsement come from the Power BI
scanner API, or per item from Core sensitivityLabel.
| Area |
Check |
Fabric admin source |
Data owner source |
| Estate |
Workspaces with no domain |
Admin List Workspaces — domainId omitted when unassigned |
Core List Workspaces (?roles=Admin) — domainId is in the schema |
| Estate |
Domains without admins; contributor scope |
Admin List Domains + role assignments |
none — see the gap in §1 |
| Estate |
Capacity assignment, state, region |
Admin capacities; capacityId per workspace |
Core List Capacities — only capacities the caller administers or contributes to |
| Protect |
Unlabeled items |
Scanner sensitivityLabel.labelId (GUID only) |
Core items' sensitivityLabel (see security) |
| Protect |
DLP coverage |
Govern tab report only — no API drilled |
none |
| Protect |
Active policy set per scope (preview) |
none found; the Core read, given a workspace role |
Core List Policy Sets per workspace — see Policy sets below |
| Access |
Item-level sharing: who holds Reshare, Write or ReadAll on an item |
Admin List Item Access Details, one call per item (Preview) |
not drilled |
| Trust |
Missing descriptions |
Admin List Items description |
Core List Items |
| Trust |
Endorsement coverage |
Scanner endorsementDetails — absent on dashboards, which cannot be endorsed |
not drilled |
| Trust |
Tag coverage |
Admin List Items tags |
Core List Tags + per-item tags |
| Trust |
Staleness |
lastUpdatedDate; activity events |
Job instances; refresh history |
| Trust |
Ownership |
creatorPrincipal, documented as "the item's owner" |
item-level only |
Per-endpoint permissions, scopes, limits and response fields are in
references/endpoints.md. Read it before
writing any call.
Policy sets (preview)
A Fabric policy set allows or blocks user actions at tenant or
capacity scope, and only an active one does anything. It is a
PolicySet item in a workspace on a Fabric capacity, one set is active
per scope at most, and domain and workspace scopes are not in the
preview (Learn, 2026-11-02). The audit asks which set is active for
each scope and what its rules allow:
- Scope and status come from List Policy Sets, per workspace:
properties.scope (Tenant, or Capacity with its id) and
properties.status (Active or Inactive). It needs a Viewer role
there, and it is the Core API, so a Fabric administrator's finding is
bounded by the workspaces they hold a role in, too.
- Rules come from Get Policy Set Definition, whose
policySet.json
lists policyRules, each a policy, its conditions and an Allow
effect. It needs read and write on the set, and its one format,
Beta, is "not recommended for production use".
- Report the rules as an allow list. Once a policy has rules, an
action no rule matches is denied, so one narrow allow rule blocks
everything else. Learn gives a policy with no rules its default, but
the portal also lets an admin set one to Blocked, and the
definition schema shows no field for that: call a rule-less policy's
state unconfirmed, not the default.
- Enforcement reaches only the users the Policies (preview) tenant
setting applies to, the entire organization or named security groups,
less any exceptions. A change takes up to 15 minutes, and the Fabric
audit logs record each decision.
Learn's policies REST page also names Activate and Deactivate Policy
Set, Read Active Policy Sets and a Policy Evaluation API, but no
reference page was found for any of them (2026-11-02): say so rather
than guess a route. Activation is §6's to gate.
4. Reporting conventions
A governance statistic with no denominator is not evidence. Every
number this skill reports carries four things.
The denominator, and what it excludes. Admin List Items returns
active items only. Bulk domain assignment "excludes My
workspaces", so personal workspaces are outside any domain finding.
The Govern tab excludes subitems such as tables, cross-tenant and guest
scenarios, and third-party workload items from its charts, and is
unavailable when Private Link is on.
Pagination completeness. Follow continuationToken until it is
absent or null. Admin List Items returns at most 10,000 records per page
in a fixed type order — Fabric items, then Datamarts, Reports,
Dashboards, SemanticModels, Apps, Dataflows — so a walk that stops early
does not lose a random sample, it loses the Power BI types at the end of
that order. Say how many pages were read.
Freshness, and which clock. API reads are live. The Govern tab's
admin insights come from Admin Monitoring Storage, which "refreshes
automatically every day", so "there could be gaps between the data
reflected and the actual state". A data owner's insights refresh every
time the tab opens. An API count that disagrees with the portal by less
than a day is the expected behaviour, not a bug.
Which population. On the Govern tab a data owner means items you
own; a data-owner API audit is usually scoped by workspace role.
Those are different sets. Name the one you used.
Flag Preview surfaces in the report: admin List Items, Get Item,
List Item Access Details and List Workspace Access Details, and Catalog
Search are each documented as Preview and "not recommended for
production use".
Per finding: the evidence (endpoint and count), the consequence, the
recommended action, which tier can perform it, and the effort.
5. Throttling shapes the plan
Limits are per endpoint and differ by an order of magnitude, so a sweep
has to be planned rather than fired:
| Endpoint |
Limit |
| Admin List Items, Get Item, List Item Access Details / List Workspaces / activity events / capacities |
200 per hour |
| Admin List Domains, tenant settings, domain CRUD, role assignments, tag apply |
25 per minute |
| Domain workspace assignment (by IDs, capacities, principals) |
10 per minute |
bulkSetLabels / bulkRemoveLabels |
25 per hour, 2,000 items each |
Scanner postWorkspaceInfo |
500 per hour, 16 concurrent, 1–100 workspaces per call |
List Item Access Details takes one call per item, so 2,000 items
take ten hours at that limit: scope an access sweep to the workspaces or
item types at risk, never the whole tenant.
On top of each, a per-identity unified quota in a fixed 60-second
window; spend it all at the start of a window and nothing more
succeeds until the next one. A 429 has two distinct causes —
RequestBlocked (rate) and CapacityLimitExceeded (tenant capacity
overload) — which need different responses; honour Retry-After either
way (rest-api).
Scanner results expire after 24 hours, and getModifiedWorkspaces
accepts a modifiedSince between 30 minutes and 30 days ago. Neither
window is the activity-event one, which is a single UTC day per call
within the last 28 days.
6. Gates before irreversible writes
These sit here, in the always-loaded body, because a gate buried in a
reference file gets read and skipped. Every row's gate must pass
before the call.
| Write |
Tier |
Gate that fires first |
| Delete a domain |
admin |
List subdomains (parentDomainId) and the assigned workspaces of the domain and each subdomain; name them and get explicit confirmation. The API documents no cascade block — only DomainNotFound and UnknownError — and Learn does not say what becomes of the workspaces. Do not tell the user they will simply be unassigned. |
| Bulk-assign workspaces to a domain |
admin |
Read every target's current domainId and present the reassignments before writing: assignment "will be overridden unless bulk reassignment is blocked by domain management tenant settings", and that setting is enabled by default. By-IDs is synchronous; by-capacities and by-principals are 202 + LRO. |
| Assign or unassign one workspace |
data owner |
A move is a reassignment: name the domain being left. Needs domain contributor or domain admin and workspace Admin; InsufficientPermissionsToDomain versus InsufficientWorkspaceRole tells you which is missing. Unassigning needs workspace Admin only. |
| Remove a domain role assignment |
admin |
Confirm the domain is not left without an admin, and name the principal. EntireTenant can only ever be a Contributor. Syncing role assignments to subdomains works for Contributors only. |
| Delete a tag |
admin |
Count the items carrying it first — deletion "is removed from all items where it was previously applied", with no undo. |
| Bulk sensitivity-label change |
admin |
Dry-run the exact item list with before and after labels, then confirm. User only — no service principal or managed identity. 25 requests/hour, 2,000 items each; the label must be in the caller's (or the delegated user's) label policy. The response is per item, not all-or-nothing: read every status, and treat InsufficientUsageRights and FailedToGetUsageRights as protected-label refusals rather than transient errors. The label also lands on linked autogenerated items, whose IDs are not returned. |
| Write a tenant setting |
admin |
Read the current value and show it beside the proposed one — the blast radius is the tenant. Enabling domain-admin delegation for certification lets domain admins override any tenant-level certification setting, not only enable/disable as the checkbox implies. |
| Assign a workspace to a capacity |
data owner |
Needs workspace Admin plus capacity contributor or admin. Returns 202 — do not report success from the 202; poll to a terminal state. Non-Power BI Fabric items cannot migrate across regions, and the target must be a Fabric, Fabric trial or Power BI Premium capacity. |
| Associate an item identity (beta) |
data owner |
This is not an ownership change. It assigns the identity to the caller and nothing else, needs Write on the item and all its children, and stops at the first child failure with details in errorInfo. Learn lists only Lakehouse and Eventstream as supported. |
| Activate, deactivate or replace a policy set (preview) |
scope admin: Fabric administrator for Tenant, capacity admin for Capacity |
Read the set's rules and the set active for that scope now, and show both: activating replaces the active set, and once it propagates, up to 15 minutes, a policy with rules denies every action no rule allows. Deactivating returns the scope to the system defaults. Git cannot activate. No route was found in the reference (§3, Policy sets), so this is the portal's unless one is. |
| Create or update a domain |
admin |
?preview=false is required on create and update. displayName is capped at 40 characters and description at 256; a duplicate name returns EntityConflict. Update accepts only displayName, description and defaultLabelId. |
assignmentMethod on a label write defaults to Standard, which Learn
defines as "set by an automated process"; Priviledged (Learn's
spelling) means manual. Choose it deliberately — the two are treated
differently by default-label policies.
7. When there is no write API, route it to a person
Do not promise a remediation that has no API, and do not invent one.
Produce a named route instead.
- Endorsement — promoted, certified, master data — has no REST
write.
UpdateItemRequest accepts only displayName and
description. Anyone with write permission on an item can promote it
in the portal; only users a Fabric admin has specified can certify,
and that field "accepts security groups only. You can't enter named
users". Certification and master data must be enabled tenant-wide
first, and enablement can be delegated per domain. So the route is:
the certifier security group for that domain.
- DLP policies are defined in the Microsoft Purview portal. Route to
the Purview compliance owner.
- Item ownership — "Currently, there's no API support for changing
ownership of Fabric items." Takeover is a portal action by someone
with read and write permission, it cascades to child items and cannot
be done on a child directly. It leaves existing schedules with their
old owner and authentication context, and a successful on-demand run
does not show a schedule will run, so review schedules separately
(Schedule ownership and item
takeover).
Mirrored databases cannot change ownership at all — a broken one must
be recreated. Semantic models, reports, dataflows and warehouses keep
their own separate mechanisms. Route to the item's
creatorPrincipal,
or to a workspace admin.
- A domain's default sensitivity label can be set by a domain admin
only through the portal, behind the "Domain admins can set default
sensitivity labels for their domains (preview)" tenant setting; the
defaultLabelId write on the domain object needs a Fabric admin. And
it is not retroactive: it applies when a new item is created and
saved, or when an existing unlabeled item is updated and saved. It
never overrides a manually applied label, and it does not work with
deployment pipelines or Git integration. A default label therefore
stops the backlog growing; it does not clear it.
8. Constraints and false negatives
- Admin List Items' silence is not evidence (§3). This is the most
likely way this skill produces a confidently wrong number.
- A sensitivity label comes back as a GUID only. Resolve names
separately —
security.
- Refresh history needs Write on the semantic model, so a Viewer's
403 means "cannot see", not "never refreshed". OneDrive refresh
history is never returned. Shared capacity allows eight scheduled or
API-triggered refreshes a day; manual portal refreshes don't count
toward it.
- Job history caps at 100 recently completed runs per item, so
"inactive" derived from it is bounded by that ceiling.
- A newly applied tag takes hours to appear in the icon and in
search, so an immediate read-back that shows nothing is not a failed
write. Tags cap at 10 per item and 10 per workspace, counted
independently, and 10,000 per tenant. Moving a workspace to another
domain keeps its domain-level tags applied, but they may not be
re-appliable there once removed.
- Catalog Search returns only
id, type, displayName,
description, catalogEntryType and hierarchy.workspace — no
label, endorsement or refresh state — and mixes Workspace entries
in with items. It cannot be the basis of a coverage statistic. Its
reference no longer lists excluded item types, and its ItemType enum
lists Dashboard and Dataflow, but no call here has shown a search
returning either (Catalog Search reference, read 2026-11-14).
What's New (2026-10-02) adds table discovery (preview): tables from
semantic models, lakehouses and mirrored databases, found by table
name, description or exact column name. The Search API reference still
says search covers only display name, workspace display name and
description, so the extension is undocumented there (2026-10-06).
- Domain assignment is not access. Assigning a workspace to a domain
"doesn't affect item visibility or accessibility"; access is workspace
role plus item permissions. All tenant users can see all domains.
- Nothing here was measured against a tenant (2026-09-11). Where a
behaviour matters and Learn is silent — what a deleted domain does to
its workspaces, whether Associate Identity answers 200 or 202 — say
it is undocumented rather than predicting it.
1---2name: catalog-governance3description: "Audit and remediate Microsoft Fabric governance posture through the REST APIs — what the OneLake catalog Govern tab reports, done by API. Covers domains and subdomains, workspace-to-domain and workspace-to-capacity assignment, domain roles, sensitivity-label coverage and bulkSetLabels, tags, description and endorsement coverage, item ownership, and which Fabric policy set is active and what it allows (preview), split by caller tier: Fabric administrator (/v1/admin/*, scanner API) versus data owner (Core API plus workspace, domain or capacity roles). Use when asked to audit governance, find unlabeled, undescribed or unendorsed items or workspaces with no domain, bulk-assign workspaces to a domain, change labels in bulk, or say who can fix a finding. Gates every irreversible write and names the fixes that have no API. For catalog search use rest-api (fab find); for workspace roles and OneLake security use security."4# model: inherit # any model: value blocks Copilot slash invocation5# effort: max # unset = inherit session effort; there is no 'effort: inherit'6disable-model-invocation: false7metadata:8 source: https://skillmd.com/skills/wardawgmalvicious/catalog-governance9---1011# Governing the Fabric estate1213The OneLake catalog's **Govern** tab reports a tenant's governance14posture across three areas — *Manage your data estate*, *Protect, secure15& comply*, and *Discover, trust, and reuse*. It also hosts governance16experiences, **Policies** (preview) among them (§3, *Policy sets*),17and administrative ones18— capacities, workspaces, tenant settings, Azure connections, workloads19and more — each shown as far as the viewer's admin role allows (Learn,202026-10-06). Govern itself went generally available in September 2026,21rolling out by region, with the Admin portal standing in where it has22not arrived; What's New keeps its February 2025 preview row too (Learn,232026-10-13). This skill does the same work through the REST APIs, where24it can be counted, scheduled and fixed in bulk.2526The output is **findings**, not a rewrite. Do not write to a tenant27unless asked separately, and then only through §6.2829Four neighbours own the adjacent ground:3031- **Catalog search** — finding an item across workspaces — is the32 Explore tab. That is `fab find` and `POST /v1/catalog/search`, both33 `rest-api`'s.34- **Access** — workspace roles, item permissions, OneLake security, and35 the shape of the `sensitivityLabel` field including resolving its GUID36 through Graph — is `security`. This skill measures label37 *coverage*; that one explains what a label does.38- **REST mechanics** — long-running operations, `continuationToken`39 pagination, `Retry-After` and 429 handling — is `rest-api`.40- **Curating item metadata**, and Microsoft's own take on this ground,41 is the Govern Skills the Fabric Updates Blog announced on 2026-10-28,42 in [microsoft/skills-for-fabric](https://github.com/microsoft/skills-for-fabric)43 (MIT). There `onelake-catalog-cli` (read at v0.3.20, 2026-11-02) runs44 the same tier-by-action grid as this skill, and adds a curate mode:45 an owner's item descriptions, display names and tags, and table and46 column descriptions through the OneLake table API. This skill keeps47 its own ground and vendors none of it.4849Everything below was drilled from Microsoft Learn on 2026-09-11, bar50the *Policy sets* material (2026-11-02), and51**none of it has been exercised against a tenant from this machine**.52Treat it as first-party documentation, not as verified practice.53Past it, its `references/`, then Learn, by `microsoft-learn-mcp`54(`microsoft_docs_search`, then `microsoft_docs_fetch`), or with no server, `curl -sL -H 'Accept: text/markdown' <page URL>` (undocumented, read 2026-11-07). A live page55that contradicts a line here wins: say so, act on the page, and route56the fix by `capture-learning`, bar a dated measured line that names the57page it overrules and the call that measured it, which a page changed since58its date sends back to re-measure. An observed claim is cited or marked59pending a doc check, and stands, marked observed, only where no page covers60it.6162## 1. Establish the caller's tier before anything else6364**Tier is the API surface the caller can reach, not their job title.**65Getting this wrong wastes a whole audit: the endpoints that answer66"across the tenant" refuse everyone who is not a Fabric administrator,67and no amount of retrying changes that.6869| Tier | Surface | Scope |70| --- | --- | --- |71| **Fabric administrator** | `/v1/admin/*`, Power BI admin (scanner, activity events) | The whole tenant |72| **Data owner** | Core API (`/v1/workspaces`, `/v1/items`, `/v1/domains`), Power BI per-item | Only objects the caller holds a role on |7374A **domain admin, capacity admin or workspace admin who is not also a75Fabric administrator is a data owner here.** The admin endpoints say so76in as many words — "The caller must be a Fabric administrator" — and77answer `InsufficientPrivileges`, "The caller doesn't have permissions to78call the API", to everyone else.7980**A 403 is the answer, not an obstacle.** Report the tier boundary and81route the finding to someone who can act (§7). Never go around it by82trying the same read through another surface.8384### The domain-admin gap — say this up front8586A domain admin who is not a Fabric administrator **has no API that87returns their domain's posture.** `GET /v1/admin/domains/{id}/workspaces`88requires the Fabric administrator role, and the Core API has no89equivalent: `GET /v1/domains` lists domains with only `id`,90`displayName`, `description` and `parentDomainId` — no membership, and91not even the `defaultLabelId` the admin version carries.9293So a data-owner audit covers **workspaces the caller can reach**, never94"my whole domain". State that limitation before running it, rather than95producing a number that quietly means something narrower than asked.9697### Service principals are a separate gate9899- Admin reads and most admin writes accept a service principal or100 managed identity, but only once a tenant admin enables it. There are101 **two independent switches** — one for read-only admin APIs, one for102 the admin APIs used for updates — each limited to a named security103 group.104- **Bulk label writes reject them outright.** `bulkSetLabels` and105 `bulkRemoveLabels` support `User` only; a service principal cannot set106 a sensitivity label at all (§6).107- Item ownership takeover likewise "doesn't cover ownership takeover as108 a service principal".109110### `fab` reaches the admin tier, not the data-owner one111112The Fabric CLI's domain verbs — `ls`, `get`, `mkdir`, `rm`, `set`,113`assign`, `unassign` against `.domains` — and **every** `fab label`114command require tenant-level Fabric Administrator privileges, per the115CLI's own docs. `fab label` also reads label names from a local JSON116file registered with `config set local_definition_labels`.117118A data owner assigning their own workspace to a domain therefore cannot119use `fab assign .domains/...`, even though the Core API would allow it.120Use the REST passthrough instead:121122```bash123fab api -X post "workspaces/$WS_ID/assignToDomain" -i '{"domainId":"<id>"}'124```125126See `rest-api` for `fab` syntax and token audiences: its127`references/cli.md` and `references/auth.md`.128129## 2. Audit before you remediate130131Establish the current state first, in the caller's own tier. An audit132mode does not write, even when asked mid-run: if the answer to a finding133is a write, say so, switch deliberately, and apply §6's gate. Do not134improvise a write inside an audit.135136## 3. What to measure, and where it actually comes from137138The trap in this whole skill, stated once:139140**`GET /v1/admin/items` carries no sensitivity label, no endorsement and141no domain.** Its item object is `id`, `type`, `name`, `description`,142`state`, `lastUpdatedDate`, `workspaceId`, `capacityId`,143`creatorPrincipal`, `defaultIdentity`, `tags` and `folderId` — nothing144else. An "unlabeled items" count taken from it is not a finding; it is145the API's silence. Labels and endorsement come from the **Power BI146scanner API**, or per item from Core `sensitivityLabel`.147148| Area | Check | Fabric admin source | Data owner source |149| --- | --- | --- | --- |150| Estate | Workspaces with no domain | Admin List Workspaces — `domainId` omitted when unassigned | Core List Workspaces (`?roles=Admin`) — `domainId` is in the schema |151| Estate | Domains without admins; contributor scope | Admin List Domains + role assignments | none — see the gap in §1 |152| Estate | Capacity assignment, state, region | Admin capacities; `capacityId` per workspace | Core List Capacities — only capacities the caller administers or contributes to |153| Protect | Unlabeled items | Scanner `sensitivityLabel.labelId` (GUID only) | Core items' `sensitivityLabel` (see `security`) |154| Protect | DLP coverage | Govern tab report only — no API drilled | none |155| Protect | Active policy set per scope (preview) | none found; the Core read, given a workspace role | Core List Policy Sets per workspace — see *Policy sets* below |156| Access | Item-level sharing: who holds `Reshare`, `Write` or `ReadAll` on an item | Admin List Item Access Details, one call per item (Preview) | not drilled |157| Trust | Missing descriptions | Admin List Items `description` | Core List Items |158| Trust | Endorsement coverage | Scanner `endorsementDetails` — absent on dashboards, which cannot be endorsed | not drilled |159| Trust | Tag coverage | Admin List Items `tags` | Core List Tags + per-item tags |160| Trust | Staleness | `lastUpdatedDate`; activity events | Job instances; refresh history |161| Trust | Ownership | `creatorPrincipal`, documented as "the item's owner" | item-level only |162163Per-endpoint permissions, scopes, limits and response fields are in164[references/endpoints.md](references/endpoints.md). Read it before165writing any call.166167### Policy sets (preview)168169**A Fabric policy set allows or blocks user actions at tenant or170capacity scope, and only an active one does anything.** It is a171`PolicySet` item in a workspace on a Fabric capacity, one set is active172per scope at most, and domain and workspace scopes are not in the173preview (Learn, 2026-11-02). The audit asks which set is active for174each scope and what its rules allow:175176- **Scope and status** come from List Policy Sets, per workspace:177 `properties.scope` (`Tenant`, or `Capacity` with its `id`) and178 `properties.status` (`Active` or `Inactive`). It needs a Viewer role179 there, and it is the Core API, so a Fabric administrator's finding is180 bounded by the workspaces they hold a role in, too.181- **Rules** come from Get Policy Set Definition, whose `policySet.json`182 lists `policyRules`, each a `policy`, its `conditions` and an `Allow`183 effect. It needs read *and write* on the set, and its one format,184 `Beta`, is "not recommended for production use".185- **Report the rules as an allow list.** Once a policy has rules, an186 action no rule matches is **denied**, so one narrow allow rule blocks187 everything else. Learn gives a policy with no rules its default, but188 the portal also lets an admin set one to *Blocked*, and the189 definition schema shows no field for that: call a rule-less policy's190 state unconfirmed, not the default.191- **Enforcement reaches only** the users the *Policies (preview)* tenant192 setting applies to, the entire organization or named security groups,193 less any exceptions. A change takes up to 15 minutes, and the Fabric194 audit logs record each decision.195196Learn's policies REST page also names Activate and Deactivate Policy197Set, Read Active Policy Sets and a Policy Evaluation API, but no198reference page was found for any of them (2026-11-02): say so rather199than guess a route. Activation is §6's to gate.200201## 4. Reporting conventions202203A governance statistic with no denominator is not evidence. Every204number this skill reports carries four things.205206**The denominator, and what it excludes.** Admin List Items returns207**active items only**. Bulk domain assignment "excludes *My208workspaces*", so personal workspaces are outside any domain finding.209The Govern tab excludes subitems such as tables, cross-tenant and guest210scenarios, and third-party workload items from its charts, and is211unavailable when Private Link is on.212213**Pagination completeness.** Follow `continuationToken` until it is214absent or null. Admin List Items returns at most 10,000 records per page215**in a fixed type order** — Fabric items, then Datamarts, Reports,216Dashboards, SemanticModels, Apps, Dataflows — so a walk that stops early217does not lose a random sample, it loses the Power BI types at the end of218that order. Say how many pages were read.219220**Freshness, and which clock.** API reads are live. The Govern tab's221admin insights come from Admin Monitoring Storage, which "refreshes222automatically every day", so "there could be gaps between the data223reflected and the actual state". A data owner's insights refresh every224time the tab opens. An API count that disagrees with the portal by less225than a day is the expected behaviour, not a bug.226227**Which population.** On the Govern tab a *data owner* means **items you228own**; a data-owner API audit is usually scoped by **workspace role**.229Those are different sets. Name the one you used.230231Flag **Preview** surfaces in the report: admin List Items, Get Item,232List Item Access Details and List Workspace Access Details, and Catalog233Search are each documented as Preview and "not recommended for234production use".235236Per finding: the evidence (endpoint and count), the consequence, the237recommended action, **which tier can perform it**, and the effort.238239## 5. Throttling shapes the plan240241Limits are per endpoint and differ by an order of magnitude, so a sweep242has to be planned rather than fired:243244| Endpoint | Limit |245| --- | --- |246| Admin List Items, Get Item, List Item Access Details / List Workspaces / activity events / capacities | 200 per hour |247| Admin List Domains, tenant settings, domain CRUD, role assignments, tag apply | 25 per minute |248| Domain workspace assignment (by IDs, capacities, principals) | 10 per minute |249| `bulkSetLabels` / `bulkRemoveLabels` | 25 per hour, 2,000 items each |250| Scanner `postWorkspaceInfo` | 500 per hour, 16 concurrent, 1–100 workspaces per call |251252**List Item Access Details takes one call per item**, so 2,000 items253take ten hours at that limit: scope an access sweep to the workspaces or254item types at risk, never the whole tenant.255256On top of each, a **per-identity unified quota in a fixed 60-second257window**; spend it all at the start of a window and nothing more258succeeds until the next one. A 429 has two distinct causes —259`RequestBlocked` (rate) and `CapacityLimitExceeded` (tenant capacity260overload) — which need different responses; honour `Retry-After` either261way (`rest-api`).262263Scanner results **expire after 24 hours**, and `getModifiedWorkspaces`264accepts a `modifiedSince` between 30 minutes and 30 days ago. Neither265window is the activity-event one, which is a single UTC day per call266within the last 28 days.267268## 6. Gates before irreversible writes269270These sit here, in the always-loaded body, because a gate buried in a271reference file gets read and skipped. **Every row's gate must pass272before the call.**273274| Write | Tier | Gate that fires first |275| --- | --- | --- |276| **Delete a domain** | admin | List subdomains (`parentDomainId`) and the assigned workspaces of the domain **and each subdomain**; name them and get explicit confirmation. The API documents no cascade block — only `DomainNotFound` and `UnknownError` — and Learn does not say what becomes of the workspaces. Do not tell the user they will simply be unassigned. |277| **Bulk-assign workspaces to a domain** | admin | Read every target's current `domainId` and present the reassignments before writing: assignment "will be overridden unless bulk reassignment is blocked by domain management tenant settings", and that setting is **enabled by default**. By-IDs is synchronous; by-capacities and by-principals are 202 + LRO. |278| **Assign or unassign one workspace** | data owner | A move is a reassignment: name the domain being left. Needs domain contributor **or** domain admin **and** workspace Admin; `InsufficientPermissionsToDomain` versus `InsufficientWorkspaceRole` tells you which is missing. Unassigning needs workspace Admin only. |279| **Remove a domain role assignment** | admin | Confirm the domain is not left without an admin, and name the principal. `EntireTenant` can only ever be a Contributor. Syncing role assignments to subdomains works for Contributors only. |280| **Delete a tag** | admin | Count the items carrying it first — deletion "is removed from all items where it was previously applied", with no undo. |281| **Bulk sensitivity-label change** | admin | Dry-run the exact item list with before and after labels, then confirm. **User only — no service principal or managed identity.** 25 requests/hour, 2,000 items each; the label must be in the caller's (or the delegated user's) label policy. The response is **per item**, not all-or-nothing: read every status, and treat `InsufficientUsageRights` and `FailedToGetUsageRights` as protected-label refusals rather than transient errors. The label also lands on linked autogenerated items, whose IDs are not returned. |282| **Write a tenant setting** | admin | Read the current value and show it beside the proposed one — the blast radius is the tenant. Enabling domain-admin delegation for certification lets domain admins override *any* tenant-level certification setting, not only enable/disable as the checkbox implies. |283| **Assign a workspace to a capacity** | data owner | Needs workspace Admin plus capacity contributor or admin. Returns **202** — do not report success from the 202; poll to a terminal state. Non-Power BI Fabric items cannot migrate across regions, and the target must be a Fabric, Fabric trial or Power BI Premium capacity. |284| **Associate an item identity** (beta) | data owner | This is **not** an ownership change. It assigns the identity to the **caller** and nothing else, needs Write on the item *and all its children*, and stops at the first child failure with details in `errorInfo`. Learn lists only Lakehouse and Eventstream as supported. |285| **Activate, deactivate or replace a policy set** (preview) | scope admin: Fabric administrator for `Tenant`, capacity admin for `Capacity` | Read the set's rules and the set active for that scope now, and show both: activating replaces the active set, and once it propagates, up to 15 minutes, a policy with rules denies every action no rule allows. Deactivating returns the scope to the system defaults. Git cannot activate. No route was found in the reference (§3, *Policy sets*), so this is the portal's unless one is. |286| **Create or update a domain** | admin | `?preview=false` is required on create and update. `displayName` is capped at 40 characters and `description` at 256; a duplicate name returns `EntityConflict`. Update accepts only `displayName`, `description` and `defaultLabelId`. |287288`assignmentMethod` on a label write defaults to `Standard`, which Learn289defines as "set by an automated process"; `Priviledged` (Learn's290spelling) means manual. Choose it deliberately — the two are treated291differently by default-label policies.292293## 7. When there is no write API, route it to a person294295Do not promise a remediation that has no API, and do not invent one.296Produce a named route instead.297298- **Endorsement** — promoted, certified, master data — has **no REST299 write**. `UpdateItemRequest` accepts only `displayName` and300 `description`. Anyone with write permission on an item can promote it301 in the portal; only users a Fabric admin has specified can certify,302 and that field "accepts security groups only. You can't enter named303 users". Certification and master data must be enabled tenant-wide304 first, and enablement can be delegated per domain. So the route is:305 the certifier security group for that domain.306- **DLP policies** are defined in the Microsoft Purview portal. Route to307 the Purview compliance owner.308- **Item ownership** — "Currently, there's no API support for changing309 ownership of Fabric items." Takeover is a portal action by someone310 with read and write permission, it cascades to child items and cannot311 be done on a child directly. It leaves existing schedules with their312 old owner and authentication context, and a successful on-demand run313 does not show a schedule will run, so review schedules separately314 ([Schedule ownership and item315 takeover](https://learn.microsoft.com/fabric/fundamentals/job-scheduler#schedule-ownership-and-item-takeover)).316 Mirrored databases cannot change ownership at all — a broken one must317 be recreated. Semantic models, reports, dataflows and warehouses keep318 their own separate mechanisms. Route to the item's `creatorPrincipal`,319 or to a workspace admin.320- **A domain's default sensitivity label** can be set by a domain admin321 only through the portal, behind the "Domain admins can set default322 sensitivity labels for their domains (preview)" tenant setting; the323 `defaultLabelId` write on the domain object needs a Fabric admin. And324 it is **not retroactive**: it applies when a new item is created and325 saved, or when an existing *unlabeled* item is updated and saved. It326 never overrides a manually applied label, and it does not work with327 deployment pipelines or Git integration. A default label therefore328 stops the backlog growing; it does not clear it.329330## 8. Constraints and false negatives331332- **Admin List Items' silence is not evidence** (§3). This is the most333 likely way this skill produces a confidently wrong number.334- **A sensitivity label comes back as a GUID only.** Resolve names335 separately — `security`.336- **Refresh history needs Write on the semantic model**, so a Viewer's337 403 means "cannot see", not "never refreshed". OneDrive refresh338 history is never returned. Shared capacity allows eight scheduled or339 API-triggered refreshes a day; manual portal refreshes don't count340 toward it.341- **Job history caps at 100 recently completed runs per item**, so342 "inactive" derived from it is bounded by that ceiling.343- **A newly applied tag takes hours to appear** in the icon and in344 search, so an immediate read-back that shows nothing is not a failed345 write. Tags cap at 10 per item and 10 per workspace, counted346 independently, and 10,000 per tenant. Moving a workspace to another347 domain keeps its domain-level tags applied, but they may not be348 re-appliable there once removed.349- **Catalog Search returns only `id`, `type`, `displayName`,350 `description`, `catalogEntryType` and `hierarchy.workspace`** — no351 label, endorsement or refresh state — and mixes `Workspace` entries352 in with items. It cannot be the basis of a coverage statistic. Its353 reference no longer lists excluded item types, and its `ItemType` enum354 lists `Dashboard` and `Dataflow`, but no call here has shown a search355 returning either (Catalog Search reference, read 2026-11-14).356 What's New (2026-10-02) adds table discovery (preview): tables from357 semantic models, lakehouses and mirrored databases, found by table358 name, description or exact column name. The Search API reference still359 says search covers only display name, workspace display name and360 description, so the extension is undocumented there (2026-10-06).361- **Domain assignment is not access.** Assigning a workspace to a domain362 "doesn't affect item visibility or accessibility"; access is workspace363 role plus item permissions. All tenant users can see all domains.364- **Nothing here was measured against a tenant** (2026-09-11). Where a365 behaviour matters and Learn is silent — what a deleted domain does to366 its workspaces, whether Associate Identity answers 200 or 202 — say367 it is undocumented rather than predicting it.