Console perf: opening a record refetches the list 10×, repeats field-access ×4 and security/explain ×2, and the inbox bell polls every 10 s — dedupe in-flight requests, pin the list's query key, push instead of poll #27626

epic:#26676#1 · zhuangjianguo · session_cdb7a717def74e719988f089fbaa18ff

Filed from the record-page epic #26676 as a sub-issue after the maintainer's 「别人的界面很丝滑，但是我们的界面总觉得卡顿，云端更卡」 and the seat's measurement (2026-10-29, console main 273137d227 in Vite dev mode against HotCRM on localhost:4001, Chrome, 1918 px). Ruling: 「立 A（ui 缺陷：列表不重取、同 URL 去重、消息中心停轮询）…两张卡，p1」. This card is the client-side waste; the server-side shape is #27627.

Measured (Resource Timing + PerformanceObserver longtask, buffer raised to 20 000)

Opening one HotCRM contact from the contacts list (?recordId=):

Pure waste inside those:

• The contacts list query (GET /api/v1/data/crm_contact?…top=50…) is refetched 10 times between 309 and 517 ms of the cold open — one per arriving child-object metadata (/meta/object/crm_event, crm_event_attendee, crm_lead, crm_opportunity, crm_task, crm_case, crm_contract, crm_campaign_member, crm_quote — 9) plus one. Each metadata arrival re-renders the list host and its data hook refetches. Twenty hit the same defect (Stop side panel URL changes from re-rendering the record index twentyhq/twenty#27438 cut row renders per open from 244 to 4).
• GET /api/v1/data/crm_contact/:id/field-access ×4, identical URL, in the same 50 ms, all 404 (consumers: ui/packages/plugin-detail/src/recordFieldAccess.tsx, plugin-form/src/recordFieldAccess.tsx, permissions/src/RecordFieldAccessScope.tsx, data-objectstack/src/index.ts).
• POST /api/v1/security/explain ×2 at the same instant (plugin-detail/src/useRecordEditable.ts, plugin-detail/src/queueClaim.ts).
• The inbox bell polls sys_inbox_message every 10 s and approvals/requests every 30 s (ui/packages/app-shell/src/hooks/sharedUserFeeds.ts: INBOX_POLL_MS = 10_000, APPROVALS_POLL_MS = 30_000, HIDDEN_POLL_MS = 60_000) while the page already holds two WebSocket connections; the measuring tab had made more than 110 000 requests over one working day. On the cloud deployment the same poll took 2.4–7.5 s per call under load, competing with the user's own navigation.

Why it matters more on the cloud: with ~0.5 s per request on huayan.objectos.app (see #27627 for the server side), every redundant request is half a second the user waits.

Target

1. The list's query key is independent of ?recordId= and of child metadata arrivals. Opening, stepping and closing a record must not refetch the list (a render-count / request-count test: 0 list requests on open). The list host re-renders at most once per record change.
2. In-flight request dedupe at the data layer: identical GETs (same URL) and identical POST security/explain bodies share one in-flight promise and one short-lived cache entry (per page life for metadata, per record for field-access / explain); a 404 field-access is cached as "no scope" for the record's life, not retried four times.
3. Child metadata loads once: the first open of a record pulls each related object's metadata once per session (today: once per open on the cold path), cached in the same data layer.
4. No polling while a push channel is open: the inbox bell and the approvals badge subscribe to the existing WebSocket feed; polling remains only as the fallback when the socket is down, at ≥ 60 s, and pauses in hidden tabs (today's HIDDEN_POLL_MS logic stays for the fallback).
5. No metadata key; ⛔ packages/spec/**.

File surface

ui/packages/app-shell/src/views/objectViewListRenderer.tsx + the list data hook it uses, ui/packages/data-objectstack/src/index.ts (dedupe / cache), ui/packages/plugin-detail/src/recordFieldAccess.tsx, plugin-form/src/recordFieldAccess.tsx, permissions/src/RecordFieldAccessScope.tsx, plugin-detail/src/useRecordEditable.ts, plugin-detail/src/queueClaim.ts, ui/packages/app-shell/src/hooks/sharedUserFeeds.ts + useInboxBell.ts (push subscription), tests beside, changeset.

Acceptance

1. Tests: opening a record from a list fires 0 list requests; the same URL in flight is fetched once; security/explain for the same record and user is posted once; child metadata is fetched once per session; with a socket open the bell makes no polling requests in 60 s, with the socket closed it polls at ≥ 60 s.
2. A before / after request log on the same HotCRM contact (cold and warm), in the PR body, from the measurement recipe below.
3. The epic seat re-measures in the console preview before ACCEPT (maintainer: 「车队开发完后你审核效果」).

performance.setResourceTimingBufferSize(20000); performance.clearResourceTimings();
const t0 = performance.now(); document.querySelector('a[data-testid="primary-field-link"]').click();
await new Promise(r => setTimeout(r, 7000));
const g = {}; for (const r of performance.getEntriesByType('resource').filter(r => r.initiatorType !== 'script')) {
  const k = r.name.replace(/^https?:\/\/[^/]+/, '').replace(/\?.*$/, '') + (r.name.includes('top=50') ? ' [LIST]' : '');
  (g[k] = g[k] || { n: 0, first: 1e9 }).n++; g[k].first = Math.min(g[k].first, Math.round(r.startTime - t0)); }
console.table(Object.entries(g).map(([u, v]) => ({ u, ...v })));