frontend-optimistic-mutations – skill Thiết kế & UI cho agent AI

Frontend Optimistic Mutations (the write path)

When to Use

Use this skill when you need a portable, framework-agnostic discipline for the write path of any React or React Native app using a query/cache layer. Codifies the optimistic-update lifecycle (cancel in-flight queries → snapshot every affected cache → patch instantly → roll back verbatim on error → invalidate on...

> Portable skill — readable by Claude Code, OpenCode, Codex, Cursor, Windsurf, and others.

> This skill describes the discipline of the write path — optimistic updates, rollback,

> idempotency, cache coherence — not a UI library or a styling system. It builds directly on the

> frontend-data-contracts skill (writes go through the typed client) and the

> frontend-architecture skill (mutations live in modules/{feature}/hooks/, keyed by a factory).

The goal: a write feels instant (the UI reflects it before the server confirms), is safe

(a failure reconciles with authoritative state, and retries require server-enforced deduplication), and leaves the cache

coherent (the detail view and every list page agree). All three at once — that's the craft.

---

0. The five core ideas

1. The optimistic lifecycle is fixed. cancel → snapshot → patch → (error: roll back) → (settle: invalidate). Every optimistic mutation follows the same five beats.

2. Roll back verbatim. On failure, restore the exact snapshot taken before the patch — not a "best guess" re-derivation. Keep the snapshot in mutation context.

3. Idempotency is generated once, not per attempt. The key is created at form init (or first intent), so a network retry replays the original server response instead of performing the action twice.

4. Caches move in lock-step. A status change patches the detail cache and every list page that contains the entity, so badges never disagree across surfaces.

5. Server state never enters the client store. Optimistic state lives in the query cache, not Zustand/Redux. The cache is the single source of truth for server data (per frontend-architecture §4).

---

1. When to be optimistic (and when not)

| Situation | Strategy |

| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ |

| High-confidence, low-conflict write (toggle nonfinancial status, like, reorder) | Optimistic — patch immediately, roll back on error. |

| Create that returns a server-generated id/number/total | Pending state, then setQueryData from the server response. A temporary optimistic row is optional; reconcile on success. |

| Destructive or hard-to-reverse write (delete with cascade, send money) | Confirm first, then optimistic _or_ pending — never silent-optimistic. |

| Write whose result the user can't see yet (background job) | Pending + toast, invalidate when done. No optimistic patch. |

Optimism is a UX tool for writes you're confident will succeed. If failure is common or expensive to

undo, prefer a pending state.

---

2. The optimistic lifecycle (TanStack Query)

The canonical shape. Each beat has a job; skipping one breaks correctness.

// modules/invoice/hooks/useInvoiceMutations.ts
interface MarkPaidContext {
  previousInvoice: Invoice | undefined; // detail snapshot
  previousLists: Array<[readonly unknown[], InvoiceListResponse]>; // every list page snapshot
}

export function useMarkInvoicePaid() {
  const queryClient = useQueryClient();
  const notifyError = useApiErrorToast();

  return useMutation<Invoice, ApiError, { id: InvoiceId }, MarkPaidContext>({
    mutationKey: invoiceKeys.all,
    retry: false, // a network error alone does not make this POST safe to retry
    mutationFn: ({ id }) => apiClient.post<Invoice>(INVOICE_API.markPaid(id)),

    // 1 + 2 + 3: cancel in-flight reads, snapshot, patch
    onMutate: async ({ id }) => {
      await queryClient.cancelQueries({ queryKey: invoiceKeys.all }); // (1) no late refetch clobber

      const detailKey = invoiceKeys.detail(id);
      const previousInvoice = queryClient.getQueryData<Invoice>(detailKey); // (2) snapshot detail
      if (previousInvoice) {
        queryClient.setQueryData<Invoice>(detailKey, {
          // (3) patch detail
          ...previousInvoice,
          status: InvoiceStatus.Paid,
        });
      }

      const previousLists: MarkPaidContext["previousLists"] = [];
      for (const [key, list] of queryClient.getQueriesData<InvoiceListResponse>(
        {
          queryKey: invoiceKeys.lists(),
        },
      )) {
        if (!list) continue;
        previousLists.push([key, list]); // (2) snapshot each page
        if (!list.invoices.some((i) => i.id === id)) continue;
        queryClient.setQueryData<InvoiceListResponse>(key, {
          // (3) patch matching row
          ...list,
          invoices: list.invoices.map((i) =>
            i.id === id ? { ...i, status: InvoiceStatus.Paid } : i,
          ),
        });
      }
      return { previousInvoice, previousLists };
    },

    // 4: revalidate on failure; an old whole-list snapshot can erase newer writes.
    onError: (error) => {
      void queryClient.invalidateQueries({ queryKey: invoiceKeys.all });
      notifyError(error);
    },

    // 5: invalidate so authoritative server state (paidAt, aggregates) refetches
    onSettled: (_d, _e, { id }) => {
      void queryClient.invalidateQueries({ queryKey: invoiceKeys.detail(id) });
      void queryClient.invalidateQueries({ queryKey: invoiceKeys.lists() });
    },
  });
}

Why each beat:

• cancel — without it, a query that was already in flight can resolve _after_ your patch and overwrite the optimistic state.
• snapshot — the only safe rollback source; never reconstruct prior state by hand.
• patch — the instant UX; mutate detail and lists together (§4).
• failure — revalidate and surface the typed ApiError; this concurrent example deliberately does not restore whole-cache snapshots. Use pending UI while authoritative data reloads.
• invalidate on settle — success or failure, refetch so server-computed fields (timestamps, totals) are authoritative. Settle, not just success: a failed write may still have changed server state.

---

Whole-snapshot rollback is safe only when cache writes are serialized or no newer update has occurred. The reference therefore invalidates on failure. Restoring snapshots requires serialized writes or versioned per-entity patches that prove no newer update would be overwritten. Filtered lists may change membership/sort/totals when status changes; merely editing a row is insufficient. Use pending state when those rules cannot be reproduced safely. An optimistic Paid badge does not prove a payment settled.

3. Non-optimistic writes: create with server-owned fields

A create that returns an id/number/total can't be fully optimistic. Run it as a pending mutation and

seed the cache from the response.

export function useCreateInvoice() {
  const queryClient = useQueryClient();
  const notifyError = useApiErrorToast();
  return useMutation<Invoice, ApiError, CreateInvoiceInput>({
    mutationFn: ({ document, idempotencyKey }) =>
      apiClient.post<Invoice>("/invoices", document, { idempotencyKey }),
    onSuccess: (invoice) => {
      queryClient.setQueryData(invoiceKeys.detail(invoice.id), invoice); // seed detail
      void queryClient.invalidateQueries({ queryKey: invoiceKeys.lists() }); // refresh lists
    },
    onError: (error) => notifyError(error),
  });
}

---

4. Cache coherence (lock-step detail + lists)

A single entity appears in many caches: its detail, and every filtered/paginated list page. An

optimistic patch must touch all of them or surfaces disagree. Use a hierarchical key factory

(from frontend-architecture §4.3) so you can target precisely.

export const invoiceKeys = {
  all: ["invoices"] as const,
  lists: () => [...invoiceKeys.all, "list"] as const,
  list: (p: IListParams) => [...invoiceKeys.lists(), p] as const,
  detail: (id: InvoiceId) => [...invoiceKeys.all, "detail", id] as const,
} as const;

• getQueriesData({ queryKey: invoiceKeys.lists() }) enumerates every cached list page so you can patch each.
• invalidateQueries({ queryKey: invoiceKeys.lists() }) refreshes them all on settle.
• detail(id) targets exactly one entity.

Snapshot each page you touch (keyed by its exact query key) so rollback restores every page

verba

… (nội dung đầy đủ có trong gói tải về)