platform-policy-rule-generate

Authoring Policy Rule Definitions

Gating: requires the EnforceOMatic and PolicyRuleMDAPI org permissions. Min API version: 64.0 (66.0 for conditions using PolicyJsonExpression).

This skill covers the on-disk metadata XML format for authoring policies. Use it whenever a task asks to write a *.policyRuleDefinition or *.policyRuleDefinitionSet file, or ship a metadata package containing them. The runtime side (RuleProvider, hooks) is out of scope.

Eval coverage: This skill is exercised by the team's ADK eval framework, not by tests/evals/ under the skill directory. Five datasets covering the ACCESS / GOVERNANCE / RECORD / TRANSFORM variants live in packages/adk-eval/eval/domains/platform-policy-rule-generate/datasets/.

1. Package Layout

A deployable package always contains:

<fixture>/
  package.xml
  policyRuleDefinitionSets/<setName>.policyRuleDefinitionSet
  policyRuleDefinitions/<ruleName>.policyRuleDefinition

package.xml template (use <version>[ftest]</version> for ftests, 64.0 or higher for real orgs):

<?xml version="1.0" encoding="UTF-8"

2. PolicyRuleDefinitionSet Schema

3. PolicyRuleDefinition — Core Fields

4. Category Decision Tree

Category names a domain (where in the platform's enforcement layers the rule applies). Effect names the action (allow / deny / transform). They are independent except for TRANSFORM.

The only Category × Effect rule the platform validates:

• effect=Transform ⇔ category=TRANSFORM_POLICY_RULE_DEFINITION (bidirectional; mismatched throws INVALIDFORCATEGORY).
• All other categories (ACCESS, GOVERNANCE, RECORD, IDENTIFIED_RECORD) accept either Permit or Forbid.

Note on the platform's auto-fill default: When <category> is omitted from authored XML, the server fills it in from effect: Permit→ACCESS, Forbid→GOVERNANCE, Transform→TRANSFORM. This is a default-fill, not a validation. If you author an explicit category that contradicts this default, it is accepted and persisted as-is.

Picking the category

ACCESS vs GOVERNANCE — how to choose

Both legally accept Permit and Forbid. Pick by which enforcement layer should record/audit the rule and what the prompt literally asks for:

Important — honor the explicit category in the prompt. If the prompt says "ACCESS policy rule that denies …" or "GOVERNANCE policy rule that permits …", emit exactly that category. Do not silently swap to the auto-fill default just because effect is Forbid (or Permit). The platform accepts both. The agent must not override the user's stated intent.

Scope × category compatibility — any combination outside this matrix throws INVALIDFORCATEGORY:

5. Condition Patterns (Quick Reference)

Every <conditions> block needs all four: <clause>, <operator>, one path element, and the value.

<conjunctionExpression> is 1-indexed prefix notation: 1, (AND 1 2), (OR 1 2), (AND (OR 1 2) (AND 3)). A bare top-level index like 1 is valid for deploy but crashes the Data Governance Policy Builder UI — see §7 if UI editability matters.

For full path enums (RulePrincipalPathType, RuleResourcePathType, RuleContextPathType), operators, and JSON expressions (PROJECTION / ARGLIST / SOQLTARGETLISTEXPR), see references/policy-schema-full.md. For copy-paste templates for all policy variants, see references/templates.md.

6. Validation Guardrails

1. Every <conditions> needs an <operator>. Missing operator → reject.
2. Every <conditions> needs at least one path element (<resourcePath>, <principalPath>, <contextPath>, or <valueDomain>).
3. <contextPath> is exclusively SESSION_DATASPACE. Never put a resource path value there.
4. Scope × category must be in the §4 matrix. Common offenders: ACCESS/GOVERNANCE + RECORD scope; RECORD + ANY/FIELD scope; TRANSFORM + ANY/RECORD scope. (Effect is independent of category except for TRANSFORM — see §4.)
5. <resourceTransform> and effect=Transform are coupled. Transform effect needs a resourceTransform. Permit/Forbid must not have one.
6. <policyRuleResourceDomains> is required for RECORD (RLS) and FIELD-scope TRANSFORM; forbidden on ACCESS/GOVERNANCE.
7. <conjunctionExpression> indices must match actual <conditions> count. Off-by-one → reject.
8. <clause> inside <conditions> must match the wrapper (WHEN inside <when…>, UNLESS inside <unless…>).
9. JSON literals in <valueString> must escape " to ". Wrong escaping silently corrupts the literal.
10. Reference targets (<valueReference>, <resourceDomain>) must exist in the target org at deploy time.
11. SCALAR_ATTRIBUTE / PLURAL_ATTRIBUTE are not in RulePrincipalPathType — not in MDAPI contract. Use a runtime RuleProvider for those shapes.
12. IDENTIFIED_RECORD is not authorable via MDAPI — SESSION_CONSUMER_ID not in RuleContextPathType.
13. Standard tag/classification dev names are fully-qualified dotted paths (e.g. DataGovernanceTags.ExternalData.Visibility.Public). Retrieve an existing rule to get the exact string before authoring.
14. Min API versions: PolicyRuleDefinition = 64.0; PolicyJsonExpression conditions = 66.0; <replicated> = 260+; <builderCompatible> / <builderValidated> = 262+.

Note on <conjunctionExpression> shape: A bare top-level index (e.g. <conjunctionExpression>1</conjunctionExpression>) deploys cleanly — the server-side parser accepts bare tokens at the top level. It is not a deploy-time validation error. It does, however, crash the Data Governance Policy Builder UI on load — see §7.

7. UI Compatibility — Core Rules

The Data Governance Policy Builder edits a strict subset of the MDAPI. Default goal: produce UI-compatible policies. Always confirm with the operator before producing API-only XML.

Hard blockers — any of these make the policy uneditable (and several crash the builder on load):

• Missing OR-of-ENTITYTYPE clause on ACCESS/GOVERNANCE rules → hard crash: Cannot use 'in' operator to search for 'Permit' in undefined. Required even when paired with TupleRead (where it's functionally redundant at runtime). Must include IS conditions for {"t":"Text","v":"OBJECT"} and {"t":"Text","v":"FIELD"}.
• Bare top-level condition index in <conjunctionExpression> (e.g. 1, or (AND (OR 1 2) 3)) → crash on builder load in buildCriteria. Deploy is unaffected, but the policy is uneditable in the UI. Always wrap: (AND 1) for a single condition; (AND (OR 1 2) (AND 3)) instead of (AND (OR 1 2) 3).
• category = ACCESS_POLICY_RULE_DEFINITION + effect = Forbid → the Data Governance Policy Builder UI (not MDAPI) collapses it to GOVERNANCE on save; round-trip via the builder will rewrite the category. MDAPI deploy is unaffected — the original ACCESS+Forbid combination is valid and deploys without modification. If your goal is UI round-trippability, prefer GOVERNANCE for Forbid; if the source of truth is MDAPI, ACCESS+Forbid is fine.
• Any <unlessPolicyRuleDefinitionClauseConjunction> block → silently dropped on first UI save.
• More than one <action> → only the first is kept.
• ruleConsumer ≠ DATACLOUD → UI hardcodes DATACLOUD on save.
• Top-level (OR 1 2) conjunction → triggers // ERROR: Unsupported rule! path, rule silently dropped.

UI-compatible "unless" rewrite:

For the full UI-compatibility checklist, round-trip rules, and operator support matrix, see references/ui-compatibility.md.

8. Authoring Workflow

1. Start from the closest template in references/templates.md — modify from there, don't start blank.
2. Pick category first (§4). Category fixes effect, resourceScopeType, and whether policyRuleResourceDomains/resourceTransform are required.
3. Lay out the bare rule: top-level fields only, no conditions. Match the category template in references/templates.md.
4. Add conditions one at a time, each with all four anchors: <clause>, <operator>, one path element, and the value.
5. Update <conjunctionExpression> — 1-indexed prefix notation. Bare top-level index (e.g. 1) deploys but breaks the UI; wrap as (AND 1) if UI editability matters (§7).
6. Run the UI-compatibility check (§7 / references/ui-compatibility.md). If any item trips, attempt the "unless" rewrite first; if not possible, get explicit operator confirmation before continuing.
7. Update package.xml — list each <members> for both types.
8. Set <builderCompatible> on the set to true if §7 checklist passes; false if intentionally API-only.
9. Sanity-check against §6 (guardrails) before considering done.
10. Validate with dry-run before any non-dry deploy to a persistent org. Surface errors using the error reference in references/deploy-errors.md.

Three-layer correctness check before done:

• Runtime enforcement — does the rule enforce what's intended? (action choice, condition shape)
• MDAPI deploy validity — does it deploy? (§6 guardrails, scope×category, tag dev names, org perms)
• UI editability — can the builder render and re-save it? (§7 checklist, OR-of-ENTITYTYPE requirement)

9. Output Hygiene — What the Agent Must NOT Say to the User

The agent's user-facing chat response accompanies every generated file. Customers should never see internal engine or implementation details.

• Do NOT name the internal evaluation engine in user-facing text — including but not limited to Cedar, policy engine, AuthZ engine, evaluation engine, Rego, OPA, or similar. The customer-facing surface is Data Governance, Policy Builder, Data Cloud Governance — use those terms only.
• Do NOT emit trailing "How it works at runtime" narratives that describe evaluation flow, principal-resource matching semantics, or engine-internal condition ordering. The generated file is the artifact; a one-line summary of what deployed is enough. If the user explicitly asks how enforcement works, describe it in product terms (e.g. "Data Governance denies the read when …"), never in engine terms.
• Do NOT name internal packages, source directories, Java class names, method names, or line-number references in user-facing text. Internal implementation identifiers are not customer-facing; refer to platform behavior in product terms only ("the server validates …", "Data Governance rejects …").
• Do NOT explain auto-fill / default-fill / fallback semantics unless the user asks. Ship the rule; surface constraints only when they affect what the user has to do next (e.g. "the DMO tag must exist before deploy").

If the user needs more context, they'll ask — respond then, in product language.

Reference Docs