fix(console): LocalizationFetchProvider 重试瞬时失败 —— 一次冷环境 503 不该让区域设置整个会话失效

.changeset/localization-retry-transient.md

@@ -0,0 +1,34 @@
+---
+"@object-ui/types": minor
+"@object-ui/permissions": patch
+"@object-ui/console": patch
+---
+
+fix(console): `LocalizationFetchProvider` retries a transient `/me/localization` failure instead of degrading for the whole session
+
+`/auth/me/localization` is served by the environment kernel that owns the session
+on a multi-tenant host, and a cold one answers `503` + `Retry-After` while it
+warms (objectstack#4159). A transient failure is therefore a normal part of a
+cold start — not an exception.
+
+The provider made ONE attempt and `.catch()`-ed into silence. So a single 503
+during warm-up left currency and locale unset for the **whole session**, silently
+and permanently, long after the kernel was ready. Every money field rendered a
+plain number and nothing ever tried again.
+
+It now re-attempts a transient failure (`408`, `425`, `429`, `502`, `503`, `504`,
+or a thrown fetch), server-stated `Retry-After` first, exponential backoff
+otherwise. `401` / `403` / `404` / `500` are real answers about the caller and
+still fail on the first attempt.
+
+**It keeps its posture.** This provider is cosmetic, so it renders children
+throughout — including mid-retry — and fills the value in if and when an attempt
+succeeds. That is the opposite of `MePermissionsProvider`, which is fail-closed
+and holds its loading state across the waits. Both are pinned by tests.
+
+The retry PRIMITIVES ("is this transient", "how long to wait", `Retry-After`
+parsing) move from `@object-ui/permissions`'s internal module to
+`@object-ui/types` — the lowest package both callers can reach — and
+`PermissionsFetchError` becomes the generic `HttpFetchError`. One definition of
+transient, two policies, rather than a second copy free to drift from the first.
+No behaviour change for `MePermissionsProvider`.

apps/console/src/LocalizationFetchProvider.test.tsx

@@ -0,0 +1,120 @@
+/**
+ * ObjectUI
+ * Copyright (c) 2024-present ObjectStack Inc.
+ *
+ * `/auth/me/localization` answers `503` + `Retry-After` while a cold environment
+ * kernel warms (objectstack#4159), so a transient failure is a normal part of a
+ * cold start on a multi-tenant host. This provider used to make ONE attempt and
+ * `.catch()` into silence, which meant a single 503 during warm-up degraded
+ * currency and locale for the whole session — silently, permanently, long after
+ * the kernel was ready.
+ *
+ * These pin the recovery AND the posture it must not lose: unlike the permission
+ * layer, this one is cosmetic, so it keeps rendering children throughout and
+ * never gates the app on the answer.
+ */
+
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import { render, screen, waitFor } from '@testing-library/react';
+import { LocalizationFetchProvider } from './LocalizationFetchProvider';
+
+const ENDPOINT = '/api/v1/auth/me/localization';
+
+const ok = (body: unknown) => ({
+  ok: true,
+  status: 200,
+  headers: new Headers(),
+  json: async () => body,
+});
+const fail = (status: number, headers: Record<string, string> = {}) => ({
+  ok: false,
+  status,
+  headers: new Headers(headers),
+  json: async () => ({}),
+});
+
+function renderProvider() {
+  return render(
+    <LocalizationFetchProvider endpoint={ENDPOINT}>
+      <span data-testid="child">child</span>
+    </LocalizationFetchProvider>,
+  );
+}
+
+describe('LocalizationFetchProvider', () => {
+  beforeEach(() => { vi.spyOn(globalThis, 'fetch' as never); });
+  afterEach(() => { vi.restoreAllMocks(); });
+
+  it('recovers from a warming 503 instead of degrading for the session', async () => {
+    (globalThis.fetch as never as ReturnType<typeof vi.fn>)
+      // `Retry-After: 0` keeps the test timer-free while still exercising the
+      // server-stated-delay path.
+      .mockResolvedValueOnce(fail(503, { 'Retry-After': '0' }))
+      .mockResolvedValueOnce(ok({ authenticated: true, currency: 'CNY', locale: 'zh-CN' }));
+
+    renderProvider();
+
+    await waitFor(() =>
+      expect((globalThis.fetch as never as ReturnType<typeof vi.fn>)).toHaveBeenCalledTimes(2),
+    );
+    // The point of the retry: the tenant default actually arrives.
+    expect(screen.getByTestId('child')).toBeTruthy();
+  });
+
+  it('retries a thrown fetch (offline / DNS / aborted) too', async () => {
+    (globalThis.fetch as never as ReturnType<typeof vi.fn>)
+      .mockRejectedValueOnce(new TypeError('Failed to fetch'))
+      .mockResolvedValueOnce(ok({ authenticated: true, currency: 'EUR' }));
+
+    renderProvider();
+
+    // A thrown fetch carries no `Retry-After`, so this one waits out the real
+    // exponential backoff (BASE_DELAY_MS) rather than a server-stated 0.
+    await waitFor(
+      () => expect((globalThis.fetch as never as ReturnType<typeof vi.fn>)).toHaveBeenCalledTimes(2),
+      { timeout: 4000 },
+    );
+  }, 10_000);
+
+  it.each([401, 403, 404, 500])(
+    'does NOT retry %i — a real answer about this caller will not change',
+    async (status) => {
+      (globalThis.fetch as never as ReturnType<typeof vi.fn>).mockResolvedValue(fail(status));
+
+      renderProvider();
+
+      // Give any (wrongly scheduled) retry a chance to fire before asserting.
+      await new Promise((r) => setTimeout(r, 50));
+      expect((globalThis.fetch as never as ReturnType<typeof vi.fn>)).toHaveBeenCalledTimes(1);
+    },
+  );
+
+  it('never blocks: children render while the fetch is still failing', async () => {
+    (globalThis.fetch as never as ReturnType<typeof vi.fn>).mockResolvedValue(
+      fail(503, { 'Retry-After': '0' }),
+    );
+
+    renderProvider();
+
+    // The posture that separates this provider from MePermissionsProvider — it
+    // is cosmetic, so nothing waits on the answer, not even mid-retry.
+    expect(screen.getByTestId('child')).toBeTruthy();
+    await new Promise((r) => setTimeout(r, 30));
+    expect(screen.getByTestId('child')).toBeTruthy();
+  });
+
+  it('stops after the retry budget rather than hammering forever', async () => {
+    (globalThis.fetch as never as ReturnType<typeof vi.fn>).mockResolvedValue(
+      fail(503, { 'Retry-After': '0' }),
+    );
+
+    renderProvider();
+
+    // 1 attempt + MAX_RETRIES(4); waitFor settles once the count stops moving.
+    await waitFor(() =>
+      expect((globalThis.fetch as never as ReturnType<typeof vi.fn>)).toHaveBeenCalledTimes(5),
+    );
+    await new Promise((r) => setTimeout(r, 50));
+    expect((globalThis.fetch as never as ReturnType<typeof vi.fn>)).toHaveBeenCalledTimes(5);
+  });
+});

apps/console/src/LocalizationFetchProvider.tsx

@@ -7,9 +7,28 @@
  * Cosmetic, NOT fail-closed: while loading or on error it renders children with
  * an empty value (no tenant default → renderers show a plain number), so a slow
  * or missing endpoint never blocks the app.
+ *
+ * That posture is why it retries QUIETLY. On a multi-tenant host this endpoint
+ * is served by the environment kernel that owns the session, and a cold one
+ * answers `503` + `Retry-After` while it warms (objectstack#4159) — so a
+ * transient failure is a normal part of a cold start, not an exception. A single
+ * attempt meant one 503 during warm-up degraded currency and locale for the
+ * WHOLE session, silently and permanently, long after the kernel was ready.
+ *
+ * It shares the "is this transient / how long to wait" primitives with
+ * `MePermissionsProvider` but not its policy: that one is fail-closed and holds
+ * its loading state across the waits, this one keeps rendering throughout and
+ * simply fills the value in if and when an attempt succeeds.
  */
 import { useEffect, useState } from 'react';
 import { LocalizationProvider, type LocalizationValue } from '@object-ui/i18n';
+import {
+  HttpFetchError,
+  backoffMs,
+  isTransientFailure,
+  retryAfterFrom,
+  sleep,
+} from '@object-ui/types';
 
 interface MeLocalizationResponse {
   authenticated?: boolean;
@@ -18,6 +37,16 @@ interface MeLocalizationResponse {
   timezone?: string | null;
 }
 
+/**
+ * Attempts for a transient failure, and the base for the exponential backoff.
+ * Deliberately more patient than the permission layer's: nothing is waiting on
+ * this, so it can afford to still be trying when a slow environment finishes
+ * warming, and a user who never sees a currency symbol is better served by a
+ * late answer than by no answer.
+ */
+const MAX_RETRIES = 4;
+const BASE_DELAY_MS = 1_000;
+
 export function LocalizationFetchProvider({
   endpoint,
   children,
@@ -29,15 +58,33 @@ export function LocalizationFetchProvider({
 
   useEffect(() => {
     let cancelled = false;
-    fetch(endpoint, { credentials: 'include', headers: { Accept: 'application/json' } })
-      .then((r) => (r.ok ? (r.json() as Promise<MeLocalizationResponse>) : null))
-      .then((json) => {
-        if (cancelled || !json) return;
-        setValue({ currency: json.currency ?? undefined, locale: json.locale ?? undefined });
-      })
-      .catch(() => {
-        /* cosmetic — leave the empty value, renderers show plain numbers */
-      });
+
+    void (async () => {
+      for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+        try {
+          const res = await fetch(endpoint, {
+            credentials: 'include',
+            headers: { Accept: 'application/json' },
+          });
+          if (!res.ok) throw new HttpFetchError(res.status, retryAfterFrom(res));
+
+          const json = (await res.json()) as MeLocalizationResponse;
+          if (cancelled) return;
+          setValue({ currency: json.currency ?? undefined, locale: json.locale ?? undefined });
+          return;
+        } catch (err) {
+          // A real answer about this caller (401/403/404/500) will not change on
+          // a retry, and neither will the last attempt — either way, stop and
+          // leave the empty value. Cosmetic: renderers show plain numbers.
+          if (!isTransientFailure(err) || attempt === MAX_RETRIES) return;
+
+          const stated = err instanceof HttpFetchError ? err.retryAfterMs : undefined;
+          await sleep(backoffMs(attempt, BASE_DELAY_MS, stated));
+          if (cancelled) return;
+        }
+      }
+    })();
+
     return () => {
       cancelled = true;
     };

packages/permissions/src/MePermissionsProvider.tsx

@@ -8,12 +8,12 @@
 
 import React, { useEffect, useMemo, useState, useCallback } from 'react';
 import {
-  PermissionsFetchError,
+  HttpFetchError,
   backoffMs,
   isTransientFailure,
-  parseRetryAfterMs,
+  retryAfterFrom,
   sleep,
-} from './retry';
+} from '@object-ui/types';
 import type {
   PermissionAction,
   PermissionCheckResult,
@@ -74,7 +74,7 @@ export interface MePermissionsProviderProps {
   errorFallback?: (err: Error, retry: () => void) => React.ReactNode;
   /**
    * How many times to re-attempt a TRANSIENT failure before giving up — a
-   * response that says "not now" (`TRANSIENT_STATUS` in `./retry`) or a network
+   * response that says "not now" (`TRANSIENT_STATUS` in `@object-ui/types`) or a network
    * error. `0` disables retrying.
    *
    * This exists because "not now" is a real answer from this endpoint. On a
@@ -148,9 +148,10 @@ export function MePermissionsProvider({
             headers: { Accept: 'application/json' },
           });
           if (!res.ok) {
-            throw new PermissionsFetchError(
+            throw new HttpFetchError(
               res.status,
-              parseRetryAfterMs(res.headers?.get?.('Retry-After'), Date.now()),
+              retryAfterFrom(res),
+              `Permissions endpoint returned ${res.status}`,
             );
           }
           const json = (await res.json()) as MePermissionsResponse;
@@ -168,7 +169,7 @@ export function MePermissionsProvider({
           }
           // `loading` stays true across the wait, so the fail-closed loading
           // state holds and the recovery is invisible to consumers.
-          const stated = err instanceof PermissionsFetchError ? err.retryAfterMs : undefined;
+          const stated = err instanceof HttpFetchError ? err.retryAfterMs : undefined;
           await sleep(backoffMs(attempt, retryBaseDelayMs, stated));
           if (!live()) return;
         }

packages/permissions/src/__tests__/MePermissionsProvider.retry.test.tsx

@@ -22,7 +22,7 @@ import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
 import { render, screen, waitFor } from '@testing-library/react';
 import React from 'react';
 import { MePermissionsProvider } from '../MePermissionsProvider';
-import { parseRetryAfterMs } from '../retry';
+import { parseRetryAfterMs } from '@object-ui/types';
 import { usePermissions } from '../usePermissions';
 
 function Probe() {

packages/permissions/src/retry.ts → packages/types/src/http-retry.ts

@@ -7,14 +7,23 @@
  */
 
 /**
- * Retry primitives for {@link MePermissionsProvider}'s `/me/permissions` fetch.
+ * Primitives for retrying a TRANSIENT HTTP failure.
  *
- * Their own module rather than the component's: they are plain functions with
- * plain tests, and a component file that also exports helpers breaks fast
- * refresh.
+ * These answer two questions only — "is this worth re-attempting?" and "how long
+ * until the next attempt?" — and nothing about what to do while waiting. That is
+ * deliberate: the two providers that use them have opposite postures.
+ * `MePermissionsProvider` is FAIL-CLOSED and holds its loading state across the
+ * waits, because rendering a permissive default early is a real hazard;
+ * `LocalizationFetchProvider` is COSMETIC and never blocks, because a missing
+ * currency just renders a plain number. One definition of "transient", two
+ * policies — rather than one policy forced on both, or two copies of the
+ * definition drifting apart.
+ *
+ * They live in this package because it is the lowest one both callers can reach
+ * (`@object-ui/types` has no `@object-ui` dependencies of its own).
  */
 
-/** Upper bound on any single backoff wait, so a bad `Retry-After` cannot park the UI. */
+/** Upper bound on any single backoff wait, so a bad `Retry-After` cannot park a caller. */
 export const MAX_RETRY_DELAY_MS = 30_000;
 
 /**
@@ -24,10 +33,10 @@ export const MAX_RETRY_DELAY_MS = 30_000;
  *  - `408` / `425` — the request itself did not land.
  *  - `429` — rate limited; `Retry-After` usually accompanies it.
  *  - `502` / `503` / `504` — the upstream is absent, warming or timing out.
- *    `503` is the load-bearing one here: on a multi-tenant host this endpoint is
- *    served by the environment kernel that owns the session, and the framework
- *    answers `503` + `Retry-After` while a cold one is still being built
- *    (objectstack#4159).
+ *    `503` is the load-bearing one: on a multi-tenant host the `/auth/me/*`
+ *    endpoints are served by the environment kernel that owns the session, and
+ *    the framework answers `503` + `Retry-After` while a cold one is still being
+ *    built (objectstack#4159). Both providers see it, so both must survive it.
  *
  * Deliberately NOT retried: `401` / `403` (a real answer about this caller),
  * `404` (no such endpoint — retrying cannot conjure one) and `500` (a genuine
@@ -36,10 +45,14 @@ export const MAX_RETRY_DELAY_MS = 30_000;
 export const TRANSIENT_STATUS: ReadonlySet<number> = new Set([408, 425, 429, 502, 503, 504]);
 
 /** An `Error` that also carries the HTTP status it came from. */
-export class PermissionsFetchError extends Error {
-  constructor(readonly status: number, readonly retryAfterMs?: number) {
-    super(`Permissions endpoint returned ${status}`);
-    this.name = 'PermissionsFetchError';
+export class HttpFetchError extends Error {
+  constructor(
+    readonly status: number,
+    readonly retryAfterMs?: number,
+    message?: string,
+  ) {
+    super(message ?? `Endpoint returned ${status}`);
+    this.name = 'HttpFetchError';
   }
 }
 
@@ -49,15 +62,18 @@ export class PermissionsFetchError extends Error {
  * never got an answer at all.
  */
 export function isTransientFailure(err: unknown): boolean {
-  return err instanceof PermissionsFetchError ? TRANSIENT_STATUS.has(err.status) : true;
+  return err instanceof HttpFetchError ? TRANSIENT_STATUS.has(err.status) : true;
 }
 
 /**
  * `Retry-After` in ms, or `undefined` when absent/unparseable. Accepts both wire
  * forms (delta-seconds and an HTTP-date) and clamps to {@link MAX_RETRY_DELAY_MS}
- * so a hostile or buggy value cannot park the UI on its loading state for hours.
+ * so a hostile or buggy value cannot park a caller for hours.
  */
-export function parseRetryAfterMs(value: string | null | undefined, nowMs: number): number | undefined {
+export function parseRetryAfterMs(
+  value: string | null | undefined,
+  nowMs: number,
+): number | undefined {
   if (!value) return undefined;
   const trimmed = value.trim();
   if (!trimmed) return undefined;
@@ -81,3 +97,8 @@ export function backoffMs(attempt: number, baseDelayMs: number, statedMs?: numbe
 
 export const sleep = (ms: number): Promise<void> =>
   new Promise<void>((resolve) => setTimeout(resolve, ms));
+
+/** Read `Retry-After` off a response without assuming `headers` exists (test doubles often omit it). */
+export function retryAfterFrom(res: { headers?: { get?(name: string): string | null } }): number | undefined {
+  return parseRetryAfterMs(res.headers?.get?.('Retry-After'), Date.now());
+}