Recipes

Cross-feature compositions for @mongez/http — patterns that come up once you’ve moved past one-off requests.

Auth interceptor with token refresh on 401

A single before() interceptor injects the access token. A single after() interceptor catches a 401, refreshes the token, and calls context.replay() to re-fire the original request. The replay() helper is guarded against infinite loops — inside an already-replayed request it short-circuits to the current result.

import { http } from "@mongez/http";

let refreshing: Promise<string | null> | null = null;

http.before(req => {
  const token = getAccessToken();
  if (token) {
    return { ...req, headers: { ...req.headers, Authorization: `Bearer ${token}` } };
  }
});

http.after(async (result, { replay }) => {
  if (!result.error?.isUnauthorized) return;

  refreshing ??= refreshAccessToken().finally(() => { refreshing = null; });
  const token = await refreshing;
  if (!token) return;

  return replay();
});

The refreshing lock collapses concurrent 401s into one refresh call instead of stampeding the refresh endpoint.

Retry with exponential backoff

For transient failures (5xx, network), retry up to N times with widening delay. @mongez/http has built-in retry — configure once on the Http instance and every request inherits it, or pass retry per-request.

import { Http } from "@mongez/http";

const http = new Http({
  baseURL: "https://api.example.com",
  retry: {
    attempts: 5,
    delay: 200,
    backoff: true,            // delay * 2^attempt — default true
    jitter: true,             // multiply by random [0.5, 1.0) to spread thundering herds
    retryOn: [429, 500, 502, 503, 504],   // default
    onRetry: (attempt, error, delay) => {
      console.warn(`retry ${attempt} after ${delay}ms — ${error.message}`);
    },
  },
});

// Per-request override — disable retry for one call
await http.get("/no-retry", { retry: false });

// Or tweak it per-request
await http.get("/flaky", { retry: { attempts: 2, delay: 100 } });

Network errors always retry while attempts remain. Aborts and timeouts never retry. When the server returns Retry-After (typical on 429), the next delay is never shorter than the value the server requested.

Cancel a request on component unmount

React’s effect cleanup is the right hook. Every request returns a CancellablePromise with .cancel() — no external AbortController plumbing required.

import { useEffect, useState } from "react";
import { http } from "@mongez/http";

function UserPanel({ id }: { id: string }) {
  const [user, setUser] = useState<User | null>(null);

  useEffect(() => {
    const req = http.get<User>(`/users/${id}`);
    req.then(({ data, error }) => {
      if (error?.isAborted) return;        // cleanup ran — drop the result
      if (data) setUser(data);
    });
    return () => req.cancel("unmounted");
  }, [id]);

  return user ? <h1>{user.name}</h1> : <Spinner />;
}

When id changes, the previous fetch aborts before the new one fires — no stale-data race. If you already have an external AbortSignal (e.g. from React Query), pass it via options.signal and the request aborts when either signal fires.

Multipart file upload with abort

FormData is the standard. http.post takes the body as a positional argument and infers the content-type — never set Content-Type for FormData yourself or you’ll strip the multipart boundary.

async function uploadAvatar(file: File, signal: AbortSignal) {
  const form = new FormData();
  form.append("avatar", file);
  form.append("crop", "square");

  const { data, error } = await http.post<{ url: string }>(
    "/users/me/avatar",
    form,
    { signal },
  );

  if (error) throw error;
  return data.url;
}

For progress reporting on string or ArrayBuffer bodies, pass onUploadProgress — the body is wrapped in a ReadableStream and chunked at 64 KiB. FormData bodies do not support progress because the browser serializes them internally without exposing the size. Requires duplex: "half" fetch support (Chrome 105+, Node 18+, Safari 17.4+).

Typed CRUD via a Resource subclass

When a backend resource (users, products, orders) has the standard CRUD endpoints, subclass Resource once and get typed methods for the lot.

import { Resource } from "@mongez/http";

class UsersResource extends Resource {
  route = "/users";
}

export const usersApi = new UsersResource();

// Now everywhere:
const { data: list }    = await usersApi.list<User[]>({ page: 1 });
const { data: user }    = await usersApi.get<User>(userId);
const { data: created } = await usersApi.create<User>({ name: "Ada" });
const { data: updated } = await usersApi.update<User>(userId, { name: "Ada L." });
await usersApi.delete(userId);

// Bonus — domain actions inherit the same {data, error} shape
await usersApi.action(userId, "ban");                  // POST /users/:id/ban
await usersApi.bulkDelete({ ids: [1, 2, 3] });         // DELETE /users  { ids: [...] }

One file declares the resource; the call sites stay free of URL-string sprawl. Subclasses can add domain methods alongside the inherited CRUD. Resource resolves its Http lazily via getCurrentHttp() — call setCurrentHttp(http) once at boot, or override per-resource with .useHttp(instance).

Dedupe identical concurrent requests

Two components mount in the same tick and both ask for /products/42. By default, concurrent GET calls to the same URL + serialised params already share one underlying fetch — each caller still gets its own CancellablePromise, and the shared request only aborts when every caller has cancelled.

Need to tweak how the dedup key is computed (e.g. ignore params, or coalesce paginated calls)? Set dedupeKey on the Http config — a function (url, params) => string.

import { Http } from "@mongez/http";

// Default — dedup by full URL + serialised params (different params = different fetches)
const http = new Http({ baseURL: "https://api.example.com" });

// Aggressive — dedup by URL only, so page=1 and page=2 share one in-flight fetch
const aggressive = new Http({
  baseURL: "https://api.example.com",
  dedupeKey: (url) => url,
});

Useful in React trees where multiple components legitimately need the same data on first render. The dedup window is the lifetime of the in-flight request — once it resolves, the cache layer (configured separately) takes over for subsequent calls.

Response cache by URL with a custom TTL

Read-heavy endpoints (settings, feature flags, public catalogues) can cache their successful responses. Configure once on the Http instance — caching applies to GET requests only.

import { Http } from "@mongez/http";

// Minimal in-memory driver — any object matching the CacheDriver shape works
const store = new Map<string, unknown>();
const memoryDriver = {
  get: async <T>(k: string) => (store.get(k) as T) ?? null,
  set: async (k: string, v: unknown) => { store.set(k, v); },
  remove: async (k: string) => { store.delete(k); },
  clear: async () => { store.clear(); },
};

const http = new Http({
  baseURL: "https://api.example.com",
  cache: { driver: memoryDriver, ttl: 60 },   // seconds
});

const { data } = await http.get<Settings>("/settings");
// Same call within 60 seconds returns instantly from cache — no network round-trip.

For browser-survivable caching, swap the in-memory Map for a @mongez/cache driver (localStorage / sessionStorage). To invalidate after a mutation:

await http.invalidate("http:https://api.example.com/settings");   // a single key
await http.invalidateAll();                                       // everything (needs driver.clear())

Default keys are http:<url>:<serialised-params>. Override with options.cacheKey per request, or supply generateKey(url, params) on the cache config.

Pair with @mongez/atomic-query for React state

@mongez/http is the request layer. @mongez/atomic-query is the React cache layer. Keep them separate and let each do its job:

const { data, isLoading, error } = useQuery({
  queryKey: ["product", id],
  queryFn: async () => {
    const { data, error } = await http.get<Product>(`/products/${id}`);
    if (error) throw error;
    return data;
  },
});

http’s {data, error} shape maps cleanly: throw on error so atomic-query’s error slot activates, return on success so its data slot populates. No exception bubbling, no manual loading state.