objectstack-ai
diff --git a/‎.changeset/16870-scope-beside-superuser-bit-refused.md‎
Lines changed: 81 additions & 0 deletions b/‎.changeset/16870-scope-beside-superuser-bit-refused.md‎
Lines changed: 81 additions & 0 deletions
diff --git a/‎content/docs/permissions/permission-sets.mdx‎
Lines changed: 11 additions & 0 deletions b/‎content/docs/permissions/permission-sets.mdx‎
Lines changed: 11 additions & 0 deletions
diff --git a/‎packages/spec/src/security/explain.test.ts‎
Lines changed: 46 additions & 1 deletion b/‎packages/spec/src/security/explain.test.ts‎
Lines changed: 46 additions & 1 deletion
diff --git a/‎packages/spec/src/security/permission.test.ts‎
Lines changed: 105 additions & 0 deletions b/‎packages/spec/src/security/permission.test.ts‎
Lines changed: 105 additions & 0 deletions
@@ -0,0 +1,81 @@
+---
+'@objectstack/spec': minor
+---
+
+fix(spec): an object permission that declares a depth axis beside the super-user bit which short-circuits it is now REFUSED, instead of being stored and counted as coverage (#16870)
+
+**BREAKING** — `ObjectPermissionSchema` no longer accepts a `readScope` beside
+`viewAllRecords: true`. Two sibling shapes are refused with it, read off the
+same resolver lines rather than guessed at.
+
+The pair was accepted with **zero diagnostics**, materialised into
+`sys_permission_set.object_permissions`, and counted by a capability census
+reading the deployed shape as coverage — while the read stayed org-wide.
+`PermissionEvaluator.getEffectiveScope` answers `org` on the super-user bit
+**before** it consults the depth key, and `getDeclaredScope` (the ADR-0090 D10
+delegated-path input) carries the identical short-circuit ahead of the identical
+read, so the declared narrowing was dropped from the delegation fold as well.
+
+⇒ the author declared a narrowing, the platform stored it, an audit of the
+deployed shape reported the capability as exercised, and the read was still
+org-wide. That is ADR-0049 `declared ≠ enforced` at the capability container
+itself, and the accept set is the only door that stops the declaration from
+being STORED: a diagnostic raised later fires after the shape is already there.
+
+```
+FROM  ObjectPermissionSchema.parse({ allowRead: true, viewAllRecords: true,
+                                     readScope: 'own_and_reports' })
+      -> { …, viewAllRecords: true, readScope: 'own_and_reports' }   // stored, unread
+
+TO    -> ZodError, located at ['readScope']:
+         "readScope: 'own_and_reports' is declared beside viewAllRecords: true,
+          which already grants org-wide read. … Delete readScope if the org-wide
+          read is intended, or set viewAllRecords: false if the narrowing is."
+```
+
+**Which pairs move, and the one that deliberately does not.** The refusal is the
+two short-circuits, transcribed:
+
+| declaration | resolver | verdict |
+|:--|:--|:--|
+| `readScope` + `viewAllRecords: true` | `opClass === 'read' && (viewAllRecords \|\| modifyAllRecords)` | **refused** |
+| `readScope` + `modifyAllRecords: true` | same disjunct | **refused** |
+| `writeScope` + `modifyAllRecords: true` | `opClass === 'write' && modifyAllRecords` | **refused** |
+| `writeScope` + `viewAllRecords: true` | the write short-circuit does not name `viewAllRecords` | **accepted — honoured, and refusing it would delete a real grant** |
+
+⛔ **What `viewAllRecords: true` GRANTS is untouched.** This changes which
+declarations are accepted, never what an accepted one does — a permission-
+semantics change is not in this change's remit. `viewAllRecords: true` alone,
+`viewAllRecords: false` beside a `readScope` (the ordinary, honoured shape), and
+a bare `readScope` all parse exactly as before; each is pinned as a
+cost-direction guard in `permission.test.ts`, and an ablation that widens the
+refusal one shape too far turns the `writeScope`-beside-`viewAllRecords` pin red.
+
+**The wire surface stays tolerant.** The refinement rides on the AUTHORING
+wrapper only; `EffectiveObjectPermissionSchema` extends the unrefined base, so a
+server still running an older toolchain can return a stored pair in an
+effective-permission response without crashing a client (#4001's authorable/wire
+split). `AccessMatrixEntry` likewise keeps describing the pair: it is a derived
+SNAPSHOT shape whose committed `access-matrix.json` may predate this refusal, and
+its tolerance is now stated with that reason in `explain.test.ts` rather than
+reading as evidence that the platform accepts the declaration.
+
+**Scope is one object-permission entry**, which is exactly the resolver's input —
+`resolveObjectPermission` returns a single entry (explicit, else the `'*'`
+wildcard) and never merges two. A super-user bit in one permission set widening
+past another set's `readScope` is ADR-0090's documented additive "widest wins"
+semantics, not a contradictory declaration, and is not judged here.
+
+**Nothing in the fleet moves.** Measured across shipped defaults, both seeded
+examples, two built access matrices, the built artifact fixture and every tracked
+`.ts` / `.json`: **0** object permissions carry any refused pair, with lit
+controls on every probe (130 nodes declaring `viewAllRecords`, 53 of them `true`,
+18 declaring `readScope`; 133 brace-local `viewAllRecords: true` literals).
+
+<!-- adr-0087: not-required (no-migration-prescription) No authorable key is added, renamed or
+retired — `readScope`, `writeScope`, `viewAllRecords` and `modifyAllRecords` all
+keep their spelling, position and meaning. What narrows is a COMBINATION, and it
+has no mechanical conversion: the two remedies (delete the depth key, or clear
+the super-user bit) express opposite author intents and only the author can
+choose. The refusal message names both at the located path, which is the whole
+notification channel a migration entry would have provided. -->
@@ -131,6 +131,17 @@ the anchor decides *which* unit `unit*` means.
 > require the enterprise hierarchy resolver; without it they **fail closed to
 > `own`** (never fail-open).
 
+> **A depth axis beside the super-user bit that bypasses it is REFUSED.** The
+> read-scope resolver answers `org` on the super-user bit *before* it consults
+> the depth key, so such a declaration would be stored and never read — and a
+> capability census over the deployed shape would count it as coverage. The
+> accept set now rejects it at the offending key, naming both remedies. Note the
+> **asymmetry**, which follows the resolver exactly: `viewAllRecords` bypasses
+> READ only, so `readScope` beside `viewAllRecords: true` is refused while
+> `writeScope` beside it is **accepted and honoured**. `modifyAllRecords`
+> bypasses both, so both `readScope` and `writeScope` beside it are refused.
+> Declare the depth, or declare the bypass — not both on the same axis.
+
 ## Capabilities & required permissions (ADR-0066)
 
 - **Capabilities** are first-class `sys_capability` records — named privileges
 
@@ -28,6 +28,9 @@ import {
   ExplainRecordAttributionSchema,
   EXPLAIN_BATCH_MAX_RECORD_IDS,
 } from './explain.zod';
+// [#16870] Imported for ONE assertion: the boundary between the snapshot shape
+// this file locks and the AUTHORING accept set that now refuses the same pair.
+import { ObjectPermissionSchema } from './permission.zod';
 
 describe('ExplainOperationSchema — the operation vocabulary is fixed', () => {
   it('accepts exactly the seven CRUD + lifecycle operations', () => {
@@ -367,7 +370,32 @@ describe('ExplainDecisionSchema — the full decision report L3 consumes', () =>
 });
 
 describe('AccessMatrix schemas — the authoring-time companion', () => {
-  it('AccessMatrixEntry locks the crud bits + super-user bypass + scopes + sharingModel', () => {
+  // [#16870] ⚠️ The sample row below carries `viewAllRecords: true` BESIDE a
+  // `readScope`, and that is deliberate — but its reason changed, so read this
+  // before reading the assertion.
+  //
+  // That pair is a contradiction: the read-scope resolver answers `org` on the
+  // super-user bit before it ever consults `readScope`, so the depth is unread
+  // on every runtime path. The AUTHORING accept set (`ObjectPermissionSchema`)
+  // therefore REFUSES it now — pinned in `permission.test.ts`, and asserted
+  // from here in the next test so the two surfaces cannot drift apart silently.
+  //
+  // `AccessMatrixEntry` deliberately does NOT refuse it, for two measured
+  // reasons:
+  //
+  //  1. It is a DERIVED SNAPSHOT shape, not an accept set. `buildAccessMatrix`
+  //     constructs entries from ALREADY-PARSED metadata, and `os build` reads a
+  //     committed `access-matrix.json` back with a bare `JSON.parse` to diff it.
+  //     A snapshot written by an older toolchain may well carry the pair, and
+  //     the drift diff has to keep describing it.
+  //  2. Nothing in the tree parses through this schema outside this file —
+  //     it is a TYPE contract that cloud's L3 product reads. A refusal added
+  //     here would be a check no code path can ever run: exactly the
+  //     declared-but-unenforced shape #16870 is about, reproduced one level up.
+  //
+  // ⇒ What this test locks is unchanged — the FIELD SHAPE, every key present
+  // and typed. It is no longer evidence that the platform accepts the pair.
+  it('AccessMatrixEntry locks the crud bits + super-user bypass + scopes + sharingModel (a SNAPSHOT shape: tolerant by design)', () => {
     const entry = AccessMatrixEntrySchema.parse({
       permissionSet: 'crm_admin', object: 'crm_lead',
       create: true, read: true, edit: true, delete: false,
@@ -382,6 +410,23 @@ describe('AccessMatrix schemas — the authoring-time companion', () => {
     });
   });
 
+  it('[#16870] the AUTHORING accept set refuses the very pair this snapshot shape tolerates', () => {
+    // The boundary, asserted rather than described. If a later change makes
+    // the authoring schema accept the pair again, this fails here too — the
+    // snapshot tolerance above is only defensible while the door upstream of
+    // it is shut.
+    const authored = ObjectPermissionSchema.safeParse({
+      allowRead: true, viewAllRecords: true, readScope: 'unit_and_below',
+    });
+    expect(authored.success, 'the accept set refuses a depth the resolver never reads').toBe(false);
+    const snapshot = AccessMatrixEntrySchema.safeParse({
+      permissionSet: 'crm_admin', object: 'crm_lead',
+      create: true, read: true, edit: true, delete: false,
+      viewAllRecords: true, modifyAllRecords: false, readScope: 'unit_and_below',
+    });
+    expect(snapshot.success, 'the snapshot shape still describes artifacts written before the refusal').toBe(true);
+  });
+
   it('the crud + bypass bits are REQUIRED (a missing bit is a contract break)', () => {
     expect(() => AccessMatrixEntrySchema.parse({
       permissionSet: 'x', object: 'y', create: true, read: true, edit: true, delete: true,
 
@@ -109,6 +109,111 @@ describe('ObjectPermissionSchema', () => {
   });
 });
 
+describe('[#16870] a depth axis beside the super-user bit that short-circuits it is REFUSED', () => {
+  // The defect: `PermissionEvaluator.getEffectiveScope` answers `org` on the
+  // super-user bit BEFORE it consults the depth key, and `getDeclaredScope`
+  // (the ADR-0090 D10 delegated-path input) carries the identical
+  // short-circuit ahead of the identical read. So the pair was accepted with
+  // zero diagnostics, materialised into `sys_permission_set.object_permissions`
+  // and counted by a capability census as coverage, while the read stayed
+  // org-wide. Refused at the accept set, which is the only site that stops the
+  // declaration from being STORED: `saveMetaItem` runs this parse before the
+  // runtime authoring gate, and `os build` runs it (compile.ts) before both the
+  // author-time rule registry and `buildAccessMatrix`.
+
+  // ── REFUSAL direction ────────────────────────────────────────────────
+  it('readScope beside viewAllRecords: true is refused, at the readScope path', () => {
+    const r = ObjectPermissionSchema.safeParse({
+      allowRead: true, viewAllRecords: true, readScope: 'own_and_reports',
+    });
+    expect(r.success).toBe(false);
+    const issue = r.error!.issues.find((i) => i.path[i.path.length - 1] === 'readScope')!;
+    expect(issue, 'the issue is LOCATED on the key the author must remove').toBeDefined();
+    expect(issue.message).toContain('viewAllRecords: true');
+    expect(issue.message).toContain('never enforced on any read path');
+    expect(issue.message, 'both remedies are named, not just one').toContain('set viewAllRecords: false');
+  });
+
+  it('readScope beside modifyAllRecords: true is refused too — the same disjunct short-circuits READ', () => {
+    const r = ObjectPermissionSchema.safeParse({
+      allowRead: true, modifyAllRecords: true, readScope: 'unit',
+    });
+    expect(r.success).toBe(false);
+    expect(r.error!.issues.some((i) => i.path[i.path.length - 1] === 'readScope')).toBe(true);
+  });
+
+  it('writeScope beside modifyAllRecords: true is refused, at the writeScope path', () => {
+    const r = ObjectPermissionSchema.safeParse({
+      allowEdit: true, modifyAllRecords: true, writeScope: 'unit',
+    });
+    expect(r.success).toBe(false);
+    const issue = r.error!.issues.find((i) => i.path[i.path.length - 1] === 'writeScope')!;
+    expect(issue).toBeDefined();
+    expect(issue.message).toContain('modifyAllRecords: true');
+  });
+
+  it('the refusal reaches through a whole permission set, located at objects.<name>.<key>', () => {
+    const r = PermissionSetSchema.safeParse({
+      name: 'sales_manager',
+      objects: { crm_opportunity: { allowRead: true, viewAllRecords: true, readScope: 'own_and_reports' } },
+    });
+    expect(r.success).toBe(false);
+    expect(r.error!.issues.some((i) => i.path.join('.') === 'objects.crm_opportunity.readScope')).toBe(true);
+  });
+
+  // ── COST direction: what the refusal must NOT cost ───────────────────
+  // A pin set that only proves the new refusal measures nothing about its
+  // price. These four are the shapes the resolver DOES read, and every one of
+  // them has to keep parsing.
+  it('viewAllRecords: false beside a readScope stays accepted — the ordinary, honoured shape', () => {
+    const parsed = ObjectPermissionSchema.parse({
+      allowRead: true, viewAllRecords: false, readScope: 'own_and_reports',
+    });
+    expect(parsed.readScope).toBe('own_and_reports');
+  });
+
+  it('a readScope with the super-user bits merely ABSENT stays accepted (the default is false)', () => {
+    const parsed = ObjectPermissionSchema.parse({ allowRead: true, readScope: 'unit' });
+    expect(parsed.readScope).toBe('unit');
+    expect(parsed.viewAllRecords).toBe(false);
+  });
+
+  it('viewAllRecords: true ALONE stays accepted — the refusal is about the pair, not the bit', () => {
+    const parsed = ObjectPermissionSchema.parse({ allowRead: true, viewAllRecords: true });
+    expect(parsed.viewAllRecords).toBe(true);
+  });
+
+  it('writeScope beside viewAllRecords: true stays ACCEPTED — viewAllRecords does not bypass write', () => {
+    // The asymmetry is read off the resolver, not assumed: the write
+    // short-circuit is `opClass === 'write' && op.modifyAllRecords`, which does
+    // not name `viewAllRecords`. Refusing this pair would delete a grant the
+    // platform honours.
+    const parsed = ObjectPermissionSchema.parse({
+      allowRead: true, allowEdit: true, viewAllRecords: true, writeScope: 'unit',
+    });
+    expect(parsed.writeScope).toBe('unit');
+  });
+
+  it('the WIRE surface stays tolerant — an older server may still emit a stored pair', () => {
+    // #4001's authorable/wire split: the refusal rides on the AUTHORING
+    // wrapper only. `EffectiveObjectPermissionSchema` extends the unrefined
+    // base, so a response carrying a pair stored before this refusal landed
+    // does not crash an older or newer client.
+    const parsed = EffectiveObjectPermissionSchema.parse({
+      allowRead: true, viewAllRecords: true, readScope: 'own_and_reports',
+    });
+    expect(parsed.readScope).toBe('own_and_reports');
+  });
+
+  it('the refusal does not disturb the shape read-through the residue stage publishes', () => {
+    // `acceptRetiredDefaultResidue` re-attaches a read-through `shape`; the
+    // refinement rides INSIDE it precisely so that stays true (a
+    // `.superRefine()` on the pipe would have dropped it).
+    expect(Object.keys(ObjectPermissionSchema.shape)).toContain('readScope');
+    expect(Object.keys(ObjectPermissionSchema.shape)).toContain('viewAllRecords');
+  });
+});
+
 describe('allowRestore / allowPurge are RETIRED (#12497, ADR-0049)', () => {
   // Removed by the 2026-08-26 maintainer ruling accepting #1883's
   // recommendation B: the `restore`/`purge` ObjectQL operations the bits