Merge pull request #47 from kristianwind/offline-blip-resilience

Offline: ride out short connection drops instead of losing the save

web/e2e/smoke.spec.js

@@ -565,8 +565,13 @@ test('en opgave kan trækkes ned i en tom sektion, og en fejlet indlæsning sige
 	await expect(section.getByText('brænde kaffe')).toBeVisible();
 	await expect(page.locator('.toast')).toHaveCount(0);
 
-	// A load that fails says so, and offers to try again.
-	await page.route('**/api/v1/projects/*', (r) => r.abort());
+	// A load that fails with a server error says so at once and offers to try again.
+	// A 500 is the server answering, not a dropped connection, so it is not a blip to
+	// wait out — it is not retried, and the message is immediate. (A dropped
+	// connection now retries for a few seconds first; that is the self-heal test.)
+	await page.route('**/api/v1/projects/*', (r) =>
+		r.fulfill({ status: 500, contentType: 'application/json', body: '{"code":"internal_error"}' })
+	);
 	await page.reload();
 	await expect(page.getByText('Serveren svarede ikke')).toBeVisible();
 	await expect(page.getByRole('button', { name: 'Prøv igen' })).toBeVisible();
@@ -577,6 +582,52 @@ test('en opgave kan trækkes ned i en tom sektion, og en fejlet indlæsning sige
 	expect(trouble.filter((t) => !t.includes('projects'))).toEqual([]);
 });
 
+/**
+ * A save made during a short drop is not lost to it.
+ *
+ * This is the whole of what "offline" means here: not a cache to work in with no
+ * server, but a save that rides out a few seconds of no connection and goes through
+ * when it comes back — with a quiet "reconnecting" while it waits, and no error and
+ * no reverted change once it heals. A drop that outlasts the retry window is a real
+ * outage and takes the normal error path; that is not this test.
+ */
+test('en kort netaudsætning under en handling genoprettes af sig selv', async ({ page }) => {
+	const trouble = watchForTrouble(page);
+	await page.goto('/');
+
+	// Wait for the app to finish loading before cutting the line — `goto` resolves at
+	// the load event, but the data load runs after hydration, and aborting that would
+	// test a cold start, not a blip mid-use.
+	const box = page.getByLabel('Ny opgave');
+	await expect(box).toBeVisible();
+
+	// The app is loaded and online. Now the connection drops — every call fails, the
+	// way it does in a tunnel, rather than one endpoint erroring.
+	let down = true;
+	await page.route('**/api/v1/**', (route) => (down ? route.abort() : route.continue()));
+
+	// A task captured during the drop. Its save fails on the network and starts
+	// retrying rather than giving up.
+	await box.fill('under et udfald');
+	await box.press('Enter');
+
+	// The quiet pill, once the retry has lasted long enough to be worth showing.
+	await expect(page.locator('.netstatus')).toBeVisible();
+
+	// The drop ends. The retry goes through on its own and the pill leaves — no
+	// error, nothing toasted as failed.
+	down = false;
+	await expect(page.locator('.netstatus')).toHaveCount(0, { timeout: 10000 });
+	await expect(page.locator('.toast')).toHaveCount(0);
+
+	// And the save landed: an undated quick capture goes to the Inbox, and there it
+	// is, saved across the drop rather than lost to it.
+	await page.getByRole('link', { name: /Indbakke/ }).click();
+	await expect(page.getByText('under et udfald')).toBeVisible({ timeout: 10000 });
+
+	expect(trouble).toEqual([]);
+});
+
 /**
  * A task can be dragged onto another day, and onto another project.
  *

web/src/lib/api.js

@@ -23,6 +23,57 @@ export class ApiError extends Error {
 }
 
 import { t } from './i18n.svelte.js';
+import { connectivity, sleepOrOnline } from './connectivity.svelte.js';
+
+/**
+ * How long a request keeps retrying a network failure before giving up.
+ *
+ * This is the whole of the "offline" feature and its whole limit: it rides out a
+ * blip of a few seconds so a save is not lost to it, and then — because the app is
+ * a live view of a shared list, not a local database — it stops and says so rather
+ * than holding an edit for a network that is not coming back. Fifteen seconds is
+ * long enough for a handover or a tunnel and short enough that a real outage fails
+ * while the user is still looking at what they did.
+ */
+const RETRY_WINDOW_MS = 15000;
+
+/** Backoff between attempts: 0.5s, 1s, 2s, then 3s, capped. A waiting sleep also
+ *  wakes the instant the OS reports a link again, so most blips end on the first. */
+const backoff = (attempt) => Math.min(500 * 2 ** attempt, 3000);
+
+/**
+ * Runs `send` (a fetch), retrying only network failures for a short window.
+ *
+ * A network failure is fetch rejecting — no HTTP status ever arrived. An HTTP
+ * error (4xx/5xx) is the server answering, so it is returned, not retried: retrying
+ * a 400 just asks the same wrong question again. An abort is the caller changing
+ * its mind (a superseded search), so it is left to propagate.
+ *
+ * Retrying is not free of risk: a create whose request reached the server but whose
+ * response was lost would be made twice. For a drop of seconds that is rare, and the
+ * alternative — losing the save every time the network hiccups — is the thing this
+ * exists to stop. A server-side idempotency key would remove even that; it is a
+ * larger change noted for later.
+ */
+async function withRetry(send, noRetry = false) {
+	const deadline = Date.now() + (noRetry ? 0 : RETRY_WINDOW_MS);
+	let attempt = 0;
+	for (;;) {
+		try {
+			const res = await send();
+			connectivity.reachable();
+			return res;
+		} catch (e) {
+			if (e?.name === 'AbortError') throw e;
+			if (Date.now() >= deadline) {
+				connectivity.lost();
+				throw new ApiError(0, 'offline', 'ingen forbindelse');
+			}
+			connectivity.retrying();
+			await sleepOrOnline(backoff(attempt++));
+		}
+	}
+}
 
 /**
  * The key for each code the API can return.
@@ -110,14 +161,10 @@ async function request(method, path, body, options = {}) {
 		init.body = JSON.stringify(body);
 	}
 
-	let res;
-	try {
-		res = await fetch(`/api/v1${path}`, init);
-	} catch (e) {
-		// A network failure is not an HTTP status; give it one the app can handle
-		// the same way as everything else.
-		throw new ApiError(0, 'offline', 'ingen forbindelse');
-	}
+	// Retries a network failure for a few seconds so a short drop does not lose the
+	// call; `withRetry` turns an exhausted window into the same `offline` error the
+	// app already handles. `options.noRetry` opts a call out — used by nothing yet.
+	const res = await withRetry(() => fetch(`/api/v1${path}`, init), options.noRetry);
 
 	if (res.status === 204) return null;
 
@@ -157,12 +204,9 @@ async function upload(path, file, field = 'file') {
 	const form = new FormData();
 	form.append(field, file);
 
-	let res;
-	try {
-		res = await fetch(`/api/v1${path}`, { method: 'POST', body: form, credentials: 'same-origin' });
-	} catch {
-		throw new ApiError(0, 'offline', 'ingen forbindelse');
-	}
+	const res = await withRetry(() =>
+		fetch(`/api/v1${path}`, { method: 'POST', body: form, credentials: 'same-origin' })
+	);
 
 	const text = await res.text();
 	let payload = null;

web/src/lib/connectivity.svelte.js

@@ -0,0 +1,70 @@
+/**
+ * Connectivity, as the app experiences it — not as the browser reports it.
+ *
+ * This is for short drops: a few seconds of lost Wi-Fi, a tunnel, a base station
+ * handing over. It is deliberately NOT offline support. The app is already loaded
+ * and its data is already in memory, so there is nothing to cache and nothing to
+ * open from cold; what a blip needs is for the save you just made to survive it,
+ * not for the whole app to run without a server.
+ *
+ * The request layer drives this: it retries a failed request for a short window
+ * (see api.js) and tells this store while it is trying, so the interface can say
+ * "reconnecting" rather than reverting the change under the user's hands. If the
+ * window runs out it is a real outage, and then the normal error path takes over.
+ *
+ * `navigator.onLine` is only a hint — it goes false when the machine has no route,
+ * but it stays true for a dead gateway or a server that is down, so it cannot be
+ * the source of truth. It is used here only to wake a waiting retry the moment the
+ * OS sees a link again, which is the one thing it is reliable for.
+ */
+import { browser } from '$app/environment';
+
+class Connectivity {
+	/** 'online' — every recent request reached the server.
+	 *  'retrying' — a request failed on the network and is being retried.
+	 *  'lost' — a request gave up after the retry window; a real outage. */
+	status = $state('online');
+
+	/** A request reached the server. The network is up, whatever it was doing. */
+	reachable() {
+		this.status = 'online';
+	}
+
+	/** A request failed on the network and is about to be tried again. */
+	retrying() {
+		if (this.status !== 'lost') this.status = 'retrying';
+	}
+
+	/** A request used up its retry window without reaching the server. */
+	lost() {
+		this.status = 'lost';
+	}
+}
+
+export const connectivity = new Connectivity();
+
+if (browser) {
+	// The OS lost its link: say so at once rather than waiting for the next request
+	// to fail. It does not, on its own, mean the server is unreachable — but with no
+	// route there is nothing a save can do but wait, and the honest thing is to show
+	// it waiting.
+	window.addEventListener('offline', () => connectivity.retrying());
+}
+
+/**
+ * Sleeps for `ms`, but wakes early the moment the OS reports a link again — so a
+ * retry that is biding its time does not sit out the last two seconds of a blip
+ * that has already ended.
+ */
+export function sleepOrOnline(ms) {
+	return new Promise((resolve) => {
+		let timer;
+		const done = () => {
+			clearTimeout(timer);
+			if (browser) window.removeEventListener('online', done);
+			resolve();
+		};
+		timer = setTimeout(done, ms);
+		if (browser) window.addEventListener('online', done, { once: true });
+	});
+}

web/src/lib/locales/da.js

@@ -130,6 +130,8 @@ export const da = {
 	'nav.toggleTheme': 'Skift mellem lyst og mørkt tema',
 	'nav.offline': 'Offline',
 	'nav.offlineHint': 'Ændringer fra andre vises ikke lige nu',
+	'net.reconnecting': 'Genopretter forbindelse…',
+	'net.offline': 'Ingen forbindelse',
 	'nav.doItToday': 'Gør det i dag',
 
 	// --- the sidebar's own sections ------------------------------------------------

web/src/lib/locales/en.js

@@ -126,6 +126,8 @@ export const en = {
 	'nav.toggleTheme': 'Switch between light and dark',
 	'nav.offline': 'Offline',
 	'nav.offlineHint': 'Changes from other people are not showing right now',
+	'net.reconnecting': 'Reconnecting…',
+	'net.offline': 'No connection',
 	'nav.doItToday': 'Do it today',
 
 	// --- the sidebar's own sections ------------------------------------------------

web/src/routes/+layout.svelte

@@ -1,6 +1,7 @@
 <script>
 	import '../app.css';
 	import { app, theme, sidebar } from '$lib/stores.svelte.js';
+	import { connectivity } from '$lib/connectivity.svelte.js';
 	import { goto } from '$app/navigation';
 	import { page } from '$app/stores';
 	import Sidebar from '$lib/components/Sidebar.svelte';
@@ -14,6 +15,29 @@
 	let paletteOpen = $state(false);
 	let sidebarOpen = $state(false);
 
+	/**
+	 * Whether to show the "reconnecting" pill.
+	 *
+	 * Held back seven tenths of a second while merely retrying, so a blip that ends
+	 * on the first retry never flashes a pill nobody needed to see. A retry that has
+	 * already given up ('lost') is a real error and shown at once — the delay is only
+	 * for the hopeful case.
+	 */
+	let showNet = $state(false);
+	$effect(() => {
+		const s = connectivity.status;
+		if (s === 'online') {
+			showNet = false;
+			return;
+		}
+		if (s === 'lost') {
+			showNet = true;
+			return;
+		}
+		const timer = setTimeout(() => (showNet = true), 700);
+		return () => clearTimeout(timer);
+	});
+
 	/**
 	 * The two pages that arrive by email, and are for somebody who is not signed in.
 	 *
@@ -204,6 +228,16 @@
 			</div>
 		{/each}
 	</div>
+
+	<!-- A short drop, said quietly: it appears only if a save is still retrying after
+	     the better part of a second, and leaves the moment the connection is back — so
+	     a two-second blip that heals itself is felt as nothing at all. -->
+	{#if showNet}
+		<div class="netstatus" class:lost={connectivity.status === 'lost'} role="status">
+			<span class="net-dot" aria-hidden="true"></span>
+			{connectivity.status === 'lost' ? t('net.offline') : t('net.reconnecting')}
+		</div>
+	{/if}
 {/if}
 
 <style>
@@ -378,6 +412,62 @@
 		padding-bottom: env(safe-area-inset-bottom);
 	}
 
+	/* The reconnecting pill: at the top, small, centred. Not a toast — a toast is a
+	   thing that happened and is over, this is a thing still happening. */
+	.netstatus {
+		position: fixed;
+		top: calc(var(--s3) + env(safe-area-inset-top));
+		left: 50%;
+		transform: translateX(-50%);
+		z-index: 60;
+		display: flex;
+		align-items: center;
+		gap: var(--s2);
+		padding: var(--s1) var(--s3);
+		border-radius: var(--radius-full);
+		background: var(--surface-raised);
+		border: 1px solid var(--line-strong);
+		box-shadow: 0 4px 18px rgba(0, 0, 0, 0.14);
+		font-size: var(--text-xs);
+		color: var(--ink-muted);
+	}
+
+	.net-dot {
+		width: 7px;
+		height: 7px;
+		flex: none;
+		border-radius: var(--radius-full);
+		background: var(--color-amber);
+		animation: net-pulse 1.1s ease-in-out infinite;
+	}
+
+	/* Given up: an amber pulse becomes a still red, the same red a failed toast
+	   carries, so "still trying" and "gave up" do not look alike. */
+	.netstatus.lost {
+		color: var(--ink);
+	}
+
+	.netstatus.lost .net-dot {
+		background: var(--p1);
+		animation: none;
+	}
+
+	@keyframes net-pulse {
+		0%,
+		100% {
+			opacity: 0.35;
+		}
+		50% {
+			opacity: 1;
+		}
+	}
+
+	@media (prefers-reduced-motion: reduce) {
+		.net-dot {
+			animation: none;
+		}
+	}
+
 	.toast {
 		display: flex;
 		align-items: center;