Epic: Permission sets — Studio designs every facet, Setup assigns users (ADR-0056)

[os-zhuang]

Epic: Permission sets — Studio designs every facet, Setup assigns users (ADR-0056)

ADR: objectui ADR-0056 — Permission editing belongs in Studio; Setup keeps user management and assignment (#2399 — being revised from "Option B" to the pure model below).
Relates to (framework): ADR-0090 (Permission Model v2), ADR-0092 (identity write guard), ADR-0086 (authz metadata↔config boundary), ADR-0066 (sys_capability registry), ADR-0049 (no unenforced security properties), ADR-0084 (app-builder IA).

⚠️ Direction reset (2026-07-12). This epic previously encoded Option B, which kept system_permissions editable in Setup as an exception. That exception is dropped. The authoritative model is now pure separation:

• Studio (/studio/<pkg>/access, studio.access) = where an app developer DESIGNS the permission set — every facet (object, field, system capabilities, RLS, tab visibility, admin scope) via structured editors.
• Setup (com.objectstack.setup, setup.access) = where a system admin ASSIGNS users to a set and sees each facet as a read-only summary with a "Design in Studio →" deep-link.
• Setup shows NO JSON and NO [Object]. Nothing permission-shaped is free-text-editable in Setup.

Problem

A permission set (sys_permission_set) is authored across two surfaces over the same record, and the Setup surface is actively broken UX:

• Setup renders the sys_permission_set object with its six authorization columns as raw values: object_permissions and field_permissions show as [Object], and system_permissions / row_level_security / tab_permissions / admin_scope show as raw JSON Field.textareas. An admin on .../sys_permission_set/record/<id> sees 对象权限 [Object], 字段权限 [Object], and hand-editable JSON blobs.
• Studio (studio.access) Access pillar has a structured PermissionMatrixEditor, but it edits only objects (CRUD/VAMA/lifecycle) and fields (FLS). The other four concerns are authorable only as raw JSON in Setup.

Raw/[Object] authorization metadata in Setup is: unreadable to an admin, invisible to the ADR-0090 explain engine (D6) and publish linter (D7), unvalidated at author time, duplicated (objects/fields editable in two surfaces), and a silent-incident footgun (ADR-0049).

Decision (pure model)

Design → Studio. Assign → Setup. Setup is read-only for every permission facet.

• Studio gets structured editors for all six facets (objects + fields already exist; capabilities, RLS, tab, admin_scope are new).
• Setup's sys_permission_set record page becomes: identity fields + user assignment + a read-only summary card per facet, each with a deep-link into the structured (Studio-grade) editor. No JSON. No [Object].
• system_permissions is designed in Studio like every other facet (this is the change from the old Option B — there is no longer a "capabilities stay editable in Setup" carve-out). Setup only displays the granted capabilities read-only + deep-links to Studio.

Deep-link target. Package-owned sets → the package's Studio Access pillar. Environment-authored sets (created in Setup, not owned by a package — so they have no /studio/<pkg>/access home) → the environment-scope structured editor via the metadata-admin permission route: /apps/com.objectstack.setup/metadata/permission/<setName> (packageId optional ⇒ env scope). Either target renders the structured editor, never JSON. (Reconciling these two entry points is open Q1.)

Already exists (reuse, don't rebuild): PermissionMatrixEditor (objects+fields) + its PermissionMatrixEditPage env-scope route; sys_capability registry object + seeded platform caps (setup.access/studio.access/manage_*); AdminScopeSchema (ADR-0090 D12); RowLevelSecurityPolicySchema; AccessExplainPanel (D6); AssignedUsersSection (add/remove sys_user_permission_set, shows sys_position_permission_set).

Target end-state (facet by facet)

sys_permission_set column	Structured editor (design)	Design home	Setup shows
object_permissions	PermissionMatrixEditor (exists)	Studio	read-only summary + deep-link
field_permissions	PermissionMatrixEditor (exists)	Studio	read-only summary + deep-link
system_permissions	sys_capability multi-select (new — from #2400)	Studio	read-only capability chips + deep-link
admin_scope	Delegated-admin editor (new)	Studio	read-only summary + deep-link
row_level_security	Structured RLS editor (new)	Studio	read-only summary + deep-link
tab_permissions	Tab-visibility editor (new)	Studio	read-only summary + deep-link
(assignment)	AssignedUsersSection (exists)	Setup	the assignment UI itself

Phases (supersedes the prior P1–P5 scoping)

• P1 — Setup: kill [Object]/JSON; render read-only summary + deep-link for all six facets; surface user assignment. (repo: objectui)
  On Setup's sys_permission_set record page, replace the raw display of all six permission columns with a read-only summary card per facet (e.g. "5 objects granted", capability chips, "3 RLS policies") + a "Design in Studio →" deep-link to the structured editor for that set; and surface AssignedUsersSection (add/remove users) on the same page. Reuses the feat(fields,plugin-form,plugin-detail): Setup system_permissions → capability multi-select (ADR-0056 P2) #2400 groundwork (single getObjectSchema widget-override choke point + form/detail widget-honoring), but points the fields at a summary+deep-link renderer instead of an editable widget.
  Acceptance: Setup shows no [Object] and no editable JSON for any permission facet; each facet has a read-only summary + a deep-link that opens the correct set in the structured editor at the right scope and returns to the Setup row; admins can add/remove users from the set in Setup; package-managed sets render read-only per ADR-0086; ADR-0054 browser proof.

• P2 — Studio: System Capabilities editor. (repo: objectui)
  Add a System Capabilities editor to Studio's Access pillar (and the env-scope PermissionMatrixEditPage) using the CapabilityMultiSelectField already built on the #2400 branch (feat/perm-setup-capability-picker): the live sys_capability registry (active = true) as scope-grouped, labelled chips with description tooltips; edits systemPermissions in the metadata draft; round-trips byte-equivalent to the stored string[] of capability names; unknown/legacy names preserved.
  Acceptance: granting/revoking studio.access and other capabilities is done in Studio via the picker, never JSON; the value persists byte-equivalent; Setup only reflects it read-only (P1); ADR-0054 proof.

• P3 — Studio: structured editors for RLS, tab visibility, and admin scope. (repo: objectui)

  • RLS: per-policy rows (object · operation · USING · CHECK) with a predicate field that round-trips CEL (simple-rule builder + raw-CEL escape hatch — open Q2); validated against the RLS predicate surface.
  • Tab visibility: per-app/tab selector → visible | hidden | default_on | default_off, enumerating installed apps/tabs.
  • Admin scope: BU-subtree picker (businessUnit + includeSubtree), manageAssignments/manageBindings/authorEnvironmentSets toggles, assignablePermissionSets[] allowlist (ADR-0090 D12).
    Acceptance: row_level_security, tab_permissions, and admin_scope are authorable structurally in Studio; editors read the canonical spec shapes (RowLevelSecurityPolicySchema, tabPermissions enum, AdminScopeSchema); malformed legacy JSON degrades to a guarded raw view rather than data loss; Studio-only (behind studio.access); ADR-0054 proofs.

• P4 — Move assignment fully into Setup; retire residual JSON; explain/lint alignment. (repo: objectui; possible framework companion)
  Move AssignedUsersSection out of Studio's PermissionMatrixEditor (currently embedded there) so assignment lives only in Setup (per the pure model); confirm every facet is now homed in a structured editor + summarized read-only in Setup, and remove any remaining raw-JSON rendering from Setup; verify each facet is legible to the ADR-0090 explain engine + publish linter.
  Acceptance: no permission concern is editable as free-text JSON anywhere in Setup; assignment UI appears in Setup, not Studio; Setup permission-set page = identity + assignment + read-only summaries with deep-links; legacy rows with irrecoverable JSON keep a guarded read-only escape (open Q5); if any field-widget narrowing is done framework-side, a companion plugin-security changeset cites ADR-0056; ADR-0054 proof.

Reusable groundwork from #2400 (do not rebuild)

The drafted P2 PR (#2400, branch feat/perm-setup-capability-picker) built plumbing that this epic reuses, even though its original placement (capability picker in Setup) is superseded by the pure model:

• CapabilityMultiSelectField (field:capability-multiselect, @object-ui/fields) + its 9 passing unit tests → moves to Studio (P2).
• Single widget-override choke point — ObjectStackAdapter.getObjectSchema stamps a field widget, honored by both ObjectForm and DetailView/DetailSection → reused by P1 to stamp the six permission fields with a summary+deep-link renderer (instead of the capability picker).
• form.tsx DATA_SOURCE_FIELD_TYPES forwarding + DetailSection widget enrichment → reused for the P1 renderer.

Disposition: #2400 is parked (draft) as superseded-in-placement; its code is preserved on-branch and re-homed by P1/P2 above. ADR #2399 is being revised from Option B to this pure model.

Open questions

1. Deep-link target/scope (P1): package-owned sets → package Studio Access pillar; env-authored sets → env-scope PermissionMatrixEditPage (/apps/com.objectstack.setup/metadata/permission/<setName>). Confirm the two entry points, package-managed read-only presentation, and back-navigation to the Setup row.
2. RLS editor depth (P3): simple-rule builder vs raw-CEL hybrid in v1; where contextVariables (on PermissionSetSchema, not a sys_permission_set column) is edited.
3. Tab-visibility source (P3): which apps/tabs are enumerated; package vs environment scope; reconciliation when apps change.
4. Summary renderers (P1): how much per-facet summary detail (counts vs first-N chips vs a compact matrix preview) before the deep-link.
5. Backward-compat (P3/P4): parse-and-render legacy JSON; guarded read-only + raw escape for malformed/unrepresentable values; whether a one-time normalization/lint pass stamps legacy rows.
6. system_permissions visibility — RESOLVED: capabilities are designed in Studio like every other facet; Setup displays them read-only (no Setup-editable exception).

Notes

• All phases are objectui UI work. Framework touchpoints are limited to how the sys_permission_set form renders and reading spec shapes; the storage columns are unchanged.
• Sequencing: P1 is the highest-value fix (it kills the visible [Object]/JSON UX the admin sees today) and is independent of P2/P3. P2 reuses feat(fields,plugin-form,plugin-detail): Setup system_permissions → capability multi-select (ADR-0056 P2) #2400 directly. P4 depends on P1–P3 homing every facet.

[os-zhuang]

All phases implemented + browser-verified in #2403 (pure model).

Phase	What	Status
P1	Setup: 6 facets → read-only summary + "Design in Studio →" deep-link (kills [Object]/JSON)	✅
P1b	Setup: user-assignment panel on the record page	✅
P2	Studio: System Capabilities editor (sys_capability multi-select)	✅
P3	Studio: structured RLS / tab-visibility / admin-scope editors	✅
P4	Assignment moved out of the design editor → Setup only; no residual JSON	✅

Verified live (app-showcase): sales_rep facet summaries + deep-link; organization_admin 18 RLS policies as structured rows; showcase_field_ops_delegate admin-scope form; capability + admin-scope save round-trips persist.

Framework follow-up (open Q + new): env-scope metadata saves of these facets don't project onto the queryable sys_permission_set data record the Setup summary reads, so a fresh Studio edit isn't reflected in Setup until the projection refreshes (enforcement reads the authoritative metadata). This is the one remaining gap for full closed-loop coherence.

[os-zhuang]

Shipped + follow-ups filed:

• Implementation merged: feat(app-shell,plugin-detail): permission sets — Studio designs, Setup assigns (ADR-0056) #2403 ✅
• ADR-0056 revised to the pure model (Accepted): docs(adr): ADR-0056 — permission editing belongs in Studio, Setup keeps assignment #2399 (updated; docs — ready to merge)
• The metadata↔data-record projection-freshness gap (Studio edits not reflected in Setup's read-only summary until the projection refreshes) is filed framework-side: sys_permission_set: env-scope permission metadata saves don't project onto the queryable data record objectstack#2857.

The objectui UI work for the pure model is complete; the one remaining gap for full closed-loop coherence is the framework projection (#2857).