@@ -0,0 +1,12 @@
+---
+'@object-ui/app-shell': patch
+---
+
+fix(app-shell): a server error at an RLS clause's path holds the permission-set autosave
+
+Clause-②: no
+Metadata-protocol: none
+
+In Studio's permission-set editor, an RLS `using` or `check` clause that the publish will refuse (for example a `using` that reads a formula field, refused as `rls-predicate-unenforceable`) showed the server's message on the field but kept autosaving: the message was displayed, and the save gate never counted it. Now the server's errors at the clause's own path hold the autosave, the same way a CEL syntax error does, while the clause still shows the text that verdict judged.
+
+The verdict describes the stored draft and nothing else, and no door judges a text without saving it. So an edit that changes the clause's text is not held: it saves, and that save's re-read brings the next verdict, which holds or releases from there. The editor decides nothing itself: it takes the server's errors as served and restates no publish rule.
@@ -54,7 +54,7 @@ import {
   type CelSuggestionKind,
   type CelValueType,
 } from './celAuthoring.js';
-import { useServerErrorsAt } from './draftVerdict.js';
+import { useServerErrorsAt, useServerHoldAt, type ServerDiagnosticEntry } from './draftVerdict.js';
 
 const LINT_DEBOUNCE_MS = 250;
 
@@ -124,11 +124,21 @@ export interface CelPredicateFieldProps {
    * the page holds (`rowLevelSecurity.0.using`). When the page's served verdict
    * (`DraftVerdictContext`) has an error at exactly that path — what publish
    * will refuse — the editor shows the server's message and withholds "Valid
-   * CEL". The lint, its findings and `onLintChange` are unaffected, as with
-   * `scopeIssue`. Omitted: no server verdict is read.
+   * CEL". The lint and its findings are unaffected, as with `scopeIssue`.
+   * Omitted: no server verdict is read.
+   *
+   * objectstack#27482 — and while the text on screen is the text that verdict
+   * judged (`DraftVerdictBodyContext`), each of those errors is also reported
+   * through `onLintChange` as a blocking error, so the host's Save gate holds
+   * on what publish will refuse exactly as on a parse fault. Once the author
+   * edits the text away the verdict no longer describes it and holds nothing:
+   * the save goes out, and its re-read brings the next verdict.
    */
   verdictPath?: string;
-  /** Reports the current lint issues up so the editor can gate Save on errors. */
+  /**
+   * Reports the current lint issues up so the editor can gate Save on errors —
+   * plus the server's errors at {@link verdictPath} on the judged text.
+   */
   onLintChange?: (issues: CelLintIssue[]) => void;
   /**
    * `role="value"` only: reports the inferred result type after each
@@ -206,14 +216,42 @@ export function CelPredicateField({
   const fieldsKey = fieldNames.join(',');
   const rootsKey = roots ? roots.join(',') : '';
 
+  // objectstack#27482 — the server's errors at this editor's path while the
+  // text on screen is the text they judged; they ride the lint report up.
+  const held = useServerHoldAt(verdictPath, value);
+  const heldKey = held.map((e) => e.message).join('\n');
+
   // Keep the reporters out of the lint effect deps (parent may not memoize them).
   const onLintChangeRef = React.useRef(onLintChange);
   const onInferredTypeChangeRef = React.useRef(onInferredTypeChange);
+  const heldRef = React.useRef<readonly ServerDiagnosticEntry[]>(held);
   React.useEffect(() => {
     onLintChangeRef.current = onLintChange;
     onInferredTypeChangeRef.current = onInferredTypeChange;
+    heldRef.current = held;
   });
 
+  // The last lint result, so a verdict that arrives or lapses between two
+  // lint passes is reported against it.
+  const lintRef = React.useRef<CelLintIssue[]>([]);
+  const report = React.useCallback((lint: CelLintIssue[]) => {
+    lintRef.current = lint;
+    const server = heldRef.current;
+    onLintChangeRef.current?.(
+      server.length === 0
+        ? lint
+        : [...lint, ...server.map((e): CelLintIssue => ({ severity: 'error', message: e.message }))],
+    );
+  }, []);
+  // Keyed on the messages, not on the array (rebuilt every render): report
+  // when the hold starts, changes or lapses, never on mount with none.
+  const reportedHeldRef = React.useRef('');
+  React.useEffect(() => {
+    if (heldKey === reportedHeldRef.current) return;
+    reportedHeldRef.current = heldKey;
+    report(lintRef.current);
+  }, [heldKey, report]);
+
   /* Debounced lint (and, for role="value", result-type inference). */
   React.useEffect(() => {
     let cancelled = false;
@@ -223,7 +261,7 @@ export function CelPredicateField({
         if (cancelled) return;
         setIssues(res);
         setLinted(true);
-        onLintChangeRef.current?.(res);
+        report(res);
       });
       if (role === 'value') {
         inferCelValueType(value, hint).then((res) => {
@@ -246,7 +284,7 @@ export function CelPredicateField({
       setIssues([]);
       setLinted(false);
       setInferredType(null);
-      onLintChangeRef.current?.([]);
+      report([]);
       if (role === 'value') onInferredTypeChangeRef.current?.(null);
     }
     // eslint-disable-next-line react-hooks/exhaustive-deps
 
@@ -0,0 +1,174 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * objectstack#27482 — a server error at an RLS clause's own path holds the
+ * package door's autosave, the way a CEL parse fault does.
+ *
+ * objectstack#26858 made the clause SHOW the draft read's verdict; it never
+ * reached `celErrorCount`, so an expression the publish refuses (a USING that
+ * reads a formula field) kept autosaving with the server's message on screen.
+ * The verdict judges the stored draft and nothing else, so it holds while the
+ * clause still shows the text it judged, and lets the edit that changes that
+ * text go out: only that save brings the next verdict. ⛔ The fake door below
+ * is the only place the refusal is decided; the editor restates no rule.
+ */
+
+import '@testing-library/jest-dom/vitest';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import { cleanup, fireEvent, render, screen, waitFor } from '@testing-library/react';
+import { MemoryRouter } from 'react-router-dom';
+
+const REFUSED_USING = 'f_total > 5';
+const ACCEPTED_USING = 'owner_id == current_user.id';
+const REFUSAL = "rowLevelSecurity 'hot_deals' using reads 'f_total', a formula field the database cannot filter on";
+
+const SET = {
+  name: 'sales_perms',
+  label: 'Sales',
+  objects: { a_account: { allowRead: true } },
+  fields: {},
+};
+
+interface Server {
+  draft: Record<string, unknown> | null;
+  draftReads: number;
+  /** Every body a save sent. */
+  sent: Array<Record<string, unknown>>;
+}
+
+/** The door's publish verdict on a stored draft: it refuses the one USING clause. */
+function verdictOf(body: Record<string, unknown>) {
+  const rows = (body.rowLevelSecurity ?? []) as Array<{ using?: string }>;
+  const errors = rows.flatMap((r, i) =>
+    r.using === REFUSED_USING
+      ? [{ path: `rowLevelSecurity.${i}.using`, message: REFUSAL, code: 'rls-predicate-unenforceable' }]
+      : [],
+  );
+  return errors.length ? { valid: false, errors } : { valid: true };
+}
+
+// `withPreviewDrafts` answers the client itself, not a promise (objectstack#26840).
+type FakeClient = Record<string, (...args: never[]) => unknown>;
+
+function makeClient(server: Server): FakeClient {
+  return {
+    layered: async () => ({ effective: SET, code: null, overlay: null, overlayScope: null }),
+    getDraft: async () => {
+      server.draftReads += 1;
+      return server.draft
+        ? { type: 'permission', name: 'sales_perms', item: { ...server.draft, _draft: true, _diagnostics: verdictOf(server.draft) } }
+        : null;
+    },
+    withPreviewDrafts() { return this; },
+    list: async (type: string) => (type === 'object' ? [{ item: { name: 'a_account' } }] : []),
+    get: async () => null,
+    save: async (_type: string, _name: string, payload: Record<string, unknown>) => {
+      const body = JSON.parse(JSON.stringify(payload)) as Record<string, unknown>;
+      server.sent.push(body);
+      server.draft = body;
+      return payload;
+    },
+  };
+}
+
+let clientImpl: FakeClient;
+
+vi.mock('./useMetadata', () => ({
+  useMetadataClient: () => clientImpl,
+  useMetadataTypes: () => ({
+    loading: false,
+    error: null,
+    entries: [{ type: 'permission', label: 'Permission', allowEnvironmentOverlay: true }],
+  }),
+}));
+
+vi.mock('./AssignedUsersSection', () => ({ AssignedUsersSection: () => null }));
+
+import { PermissionMatrixEditPage } from './PermissionMatrixEditor';
+
+afterEach(cleanup);
+
+const USING_PLACEHOLDER = 'organization_id == current_user.organization_id';
+const CHECK_PLACEHOLDER = 'optional — defaults to USING';
+/** The shared autosave fires 1.5s after the last edit. */
+const AUTOSAVE = { timeout: 5000 };
+/** Longer than the autosave's pause. */
+const PAST_THE_PAUSE = 2500;
+const pause = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms));
+
+const policy = (using: string) => ({
+  ...SET,
+  rowLevelSecurity: [{ name: 'hot_deals', object: 'a_account', operation: 'select', using, enabled: true }],
+});
+
+async function openPackageDoor(server: Server) {
+  clientImpl = makeClient(server);
+  render(
+    <MemoryRouter>
+      <PermissionMatrixEditPage type="permission" name="sales_perms" packageId="app.a" />
+    </MemoryRouter>,
+  );
+  await screen.findByText('a_account');
+  fireEvent.click(screen.getByRole('button', { name: /Row-Level Security/ }));
+}
+
+const edit = (placeholder: string, value: string) =>
+  fireEvent.change(screen.getByPlaceholderText(placeholder), { target: { value } });
+
+describe('a server error at a clause path holds the autosave (objectstack#27482)', () => {
+  it('(a) a stored draft the publish refuses: another edit is held, no PUT, the server message stays', async () => {
+    const server: Server = { draft: policy(REFUSED_USING), draftReads: 0, sent: [] };
+    await openPackageDoor(server);
+    await screen.findByText(REFUSAL);
+
+    edit(CHECK_PLACEHOLDER, ACCEPTED_USING);
+
+    await waitFor(() => expect(screen.getByTestId('perm-autosave-held')).toBeInTheDocument(), AUTOSAVE);
+    await pause(PAST_THE_PAUSE);
+    expect(server.sent).toHaveLength(0);
+    expect(screen.getByText(REFUSAL)).toBeInTheDocument();
+  });
+
+  it('(b) the card flow: the refused clause saves once, its re-read holds the next edit', async () => {
+    const server: Server = { draft: policy(ACCEPTED_USING), draftReads: 0, sent: [] };
+    await openPackageDoor(server);
+    await screen.findByText('Valid CEL', {}, { timeout: 3000 });
+
+    // No client copy of the rule: the lint passes it and the first save goes out.
+    edit(USING_PLACEHOLDER, REFUSED_USING);
+    await waitFor(() => expect(server.sent).toHaveLength(1), AUTOSAVE);
+    expect(await screen.findByText(REFUSAL, {}, AUTOSAVE)).toBeInTheDocument();
+
+    edit(CHECK_PLACEHOLDER, ACCEPTED_USING);
+    await waitFor(() => expect(screen.getByTestId('perm-autosave-held')).toBeInTheDocument(), AUTOSAVE);
+    await pause(PAST_THE_PAUSE);
+    expect(server.sent).toHaveLength(1);
+  });
+
+  it('(c) editing the clause away releases: the save goes out and its clean re-read clears the verdict', async () => {
+    const server: Server = { draft: policy(REFUSED_USING), draftReads: 0, sent: [] };
+    await openPackageDoor(server);
+    await screen.findByText(REFUSAL);
+
+    edit(USING_PLACEHOLDER, ACCEPTED_USING);
+
+    await waitFor(() => expect(server.sent).toHaveLength(1), AUTOSAVE);
+    await waitFor(() => expect(screen.queryByText(REFUSAL)).toBeNull(), AUTOSAVE);
+    expect(await screen.findByTestId('perm-saved-at', {}, AUTOSAVE)).toBeInTheDocument();
+    expect(screen.queryByTestId('perm-autosave-held')).toBeNull();
+  });
+
+  it('(d) typed back to the judged text before the autosave fires: held again, no PUT', async () => {
+    const server: Server = { draft: policy(REFUSED_USING), draftReads: 0, sent: [] };
+    await openPackageDoor(server);
+    await screen.findByText(REFUSAL);
+
+    edit(CHECK_PLACEHOLDER, ACCEPTED_USING);
+    edit(USING_PLACEHOLDER, 'f_total > 50');
+    edit(USING_PLACEHOLDER, REFUSED_USING);
+
+    await waitFor(() => expect(screen.getByTestId('perm-autosave-held')).toBeInTheDocument(), AUTOSAVE);
+    await pause(PAST_THE_PAUSE);
+    expect(server.sent).toHaveLength(0);
+  });
+});