Skip to content

Offline mode for visited/loaded resources#2582

Open
rosa wants to merge 17 commits intomainfrom
offline-mode
Open

Offline mode for visited/loaded resources#2582
rosa wants to merge 17 commits intomainfrom
offline-mode

Conversation

@rosa
Copy link
Member

@rosa rosa commented Feb 20, 2026

This adds offline mode to Fizzy, for web and the upcoming Hotwire Native apps. This is the simplest version where only stuff you've already seen is available offline. Any write action will fail.

This relies on hotwired/turbo#1427, which has been extended with more functionality after testing this on Fizzy, and extends the service worker that was only handling web push notifications to cache resources for offline access, with different rules depending on the nature of the resources.

@rosa rosa changed the title Offline mode for visited pages Offline mode for visited/loaded resources Feb 20, 2026
@rosa rosa force-pushed the offline-mode branch 3 times, most recently from 491e1ae to f4f32f7 Compare February 20, 2026 20:42
rosa and others added 16 commits February 24, 2026 16:18
This change conditionally renders the TurboOffline caching rules based on
whether the request comes from a Hotwire Native app. Web browsers get a
minimal service worker that only handles push notifications and a simple
document fetch fallback.

Why separate behavior for native vs web?
-----------------------------------------
We want offline caching for Hotwire Native apps (where users expect
app-like offline behavior) but not for regular web browsers.

Why use ERB conditional rendering?
----------------------------------
We explored several approaches to detect Hotwire Native requests in the
service worker:

1. User-Agent detection in fetch handler: Would be ideal, but Android
   WebViews override the User-Agent header with the system default when
   requests are intercepted by service workers, stripping the custom
   "Hotwire Native" identifier.

2. Custom header (X-Hotwire-Native) from native apps: Would require
   intercepting ALL requests at the native level using both
   WebViewClient.shouldInterceptRequest() and ServiceWorkerClient.
   Complex to implement and still incomplete coverage for all request
   types (navigation, resources loaded by HTML).

3. In-memory flag with IndexedDB persistence: Service workers can be
   terminated when idle and restart with fresh state. Reading from
   IndexedDB is async, but the decision to call respondWith() in a
   fetch handler must be synchronous.

4. Separate service worker URLs: Same theoretical churn problem as ERB
   rendering, with more complexity.

Why ERB conditional rendering works in practice
-----------------------------------------------
The main concern with conditional ERB rendering was "churn" — the service
worker constantly updating as different client types fetch it. However,
this only happens when web and native share storage on the same device.

In practice, this is rare because:
- Android native apps use isolated WebView storage
- iOS doesn't support service workers in WebViews (yet)
- Web browsers have completely separate storage

So each context gets its own stable service worker without churn.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adds a logout Stimulus controller that sends a message to the service
worker to clear all cached content when the user logs out. This ensures
that cached data from one user isn't accessible after logout.

The implementation:
- Adds logout_controller.js that posts { action: "clearCache" } to the
  service worker via postMessage
- Updates logout buttons to use the controller on form submission

Also fixes data-turbo placement: moved from button to form element where
it actually takes effect for disabling Turbo Drive form submissions.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously, offline caching was conditionally enabled only for Hotwire
Native apps. This removes that restriction and enables offline support
for all users, including PWAs and regular browsers.

This Uses the new `fetchOptions` support in `TurboOffline` handlers
to pass `cache: "no-cache"` for document fetches, which we need
to work around a quite annoying Safari PWA bug
(see #1014).

Also, simplify a bit the cache names and remove the `misc` one.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Cached video and audio files.
1MB and no limit for number of entries except for attachments, where we
limit individual entries to 2MB and total of entries to 500. The number
is based on the following percentiles for Active Storage blobs:

```
p50: 97.1044921875 KB
p75: 236.9140625 KB
p90: 917.7548828125 KB
```
When the service worker is registered for the first time, resources loaded
before it becomes active won't go through the service worker. These resources
may be served from the browser's HTTP cache on subsequent requests, bypassing
the service worker cache entirely.

The new `preload` option accepts a regex pattern. On first visit (when no
controller exists), it waits for the service worker to take control, then
sends a message with URLs from `performance.getEntriesByType("resource")`
that match the pattern. The service worker fetches and caches these resources.
Rename logout controller to clear-offline-cache and attach it to the
magic link verification form so the service worker cache is cleared
when a different user signs in.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Simplify the clear-offline-cache controller to use the new
Turbo.offline.clearCache() API instead of messaging the service
worker directly.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
We don't need the service worker to cache itself, or pages that won't
work offline anyway.
Gate Turbo.offline.start() behind Current.user so the service worker
is only registered for authenticated users. This avoids errors on
unauthenticated pages where /service-worker.js requires auth.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Otherwise, for resources like images loaded via <img> tags, the browser
sets `mode: "no-cors"` (as these aren't CORS requests), so the service
worker gets an opaque response even though the server sends CORS
headers.

We could upgrade all no-cors requests to cors mode, sending a `mode:
"cors"` request to a server that doesn't send CORS headers will fail
entirely: the `fetch` call throws a `TypeError` (network error), and the
browser blocks the response. So the resource wouldn't load at all, not
even as opaque. We wouldn't be able to cache it at all.

By opting in via `fetchOptions`, we can do it only for rules where we
know the server will send CORS headers.
CORS mode doesn't work with redirect-based Active Storage URLs
when the storage bucket uses wildcard Access-Control-Allow-Origin.
Relying on maxEntries alone until specific origin CORS is available.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant