| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
| |
and no whole-build download on the first launch after a deploy. Navigations to / were network-first, so every launch waited on a round trip for the page, and after each deploy (several a day) the page it got named a new 1 MB entry bundle plus changed chunks (~650 KB compressed) that had to download before anything showed — seconds on a phone's data, which is where "so slow to load" (a user on the APK, 2026-11-05) comes from. Now / is cache-first from this build's own cache; the browser checks sw.js on the same navigation, a new build installs in the background (shell, entry and chunks together) and the next launch opens it. A refresh (request cache mode reload/no-cache) still goes to the network and gets the new build at once. Other paths (server-rendered note and profile pages) stay network-first, and only this build's own page is stored as its shell — before, any navigation's response (a profile page included) was put as './'. Navigation preload is on, so a network-bound navigation's request starts while the worker is still starting; cs nginx answers a launch's preload (no Cache-Control on it; a refresh sends max-age=0) with an empty 204, so a launch spends no bytes on a page it paints from the cache, and the worker then fetches normally if it does need it. Activation keeps the previous build's cache one deploy longer, so a page launched from the old shell still finds its lazily imported chunks. A page that installs a newer build while open no longer reloads under your hands: if nothing has been touched it reloads as it is put away, otherwise the new build waits for the next launch. tools/sw-reload-test.js covers launch (old build painted, page not fetched beyond an empty preload, new build installs, no reload, next launch opens it), refresh (new build at once, one load), an untouched open page (reloads only when hidden) and a used one (never reloads, older chunks still cached)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
A suite run let ~90 throwaway test wallets subscribe and publish on public
relays from one home address, and nos.lol and relay.damus.io banned it.
tools/puppeteer.js wraps puppeteer-core: every socket to a non-local host
gets an in-page stand-in relay (EVENT -> OK, REQ -> EOSE). Tests that read
real content launch with { relays: 'read' } (real reads, publishes answered
in place); feed-live, chat-layout and cordn-ui launch with { relays: 'live' }.
LIVE_RELAYS=read|1|host,... overrides per run.
profile-key-qr: measure the QR again if a repaint swapped it out.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
- unread-dot: the front door's way to import is Create a new wallet → Have an
existing seed; the dot is the Messages bottom-nav button's
- sw-reload: the split build loads app-<token>.js, so a restamped build asks
for app-deadbeef….js; serve the real entry for it
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
|
|
The app is inlined in the page, so when a new service worker took
control the page reloaded — otherwise it would keep running the old
build. But navigations are network-first: the refresh that triggered
the update had already fetched the new page, so the reload landed a
second or two later as an unasked-for second refresh.
The page and the worker now carry the same build token and the page
asks which build took over, reloading only on a real mismatch. The
token is hashed before substitution, since a page cannot contain a
hash of itself.
tools/sw-reload-test.js pins all three cases against a swappable
server: a first visit doesn't reload itself, a refresh after a deploy
is the only refresh, and a page left sitting open across a deploy is
still reloaded — exactly once. The previous build takes two loads for
one manual refresh; this one takes one.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Wvj4aoeMZwA7VUGdJP9Swi
|