Browser refresh clears form values; underlying tension between page-instance preservation and singleton data freshness

[tkuhn]

Symptom

On both production (https://nanodash.knowledgepixels.com) and locally, when a user partially fills a form (e.g. on /publish?template=...) and presses browser refresh (F5), the page reloads with all fields blank. The previously entered values are gone.

This used to work: refresh restored the partly-filled form.

History / root cause

There are two relevant commits:

1. e03c61594 (PR #401, 2026-03-16) — fix: use ONE_PASS_RENDER so browser refresh creates fresh pages. This switched Wicket's render strategy from the default REDIRECT_TO_BUFFER to ONE_PASS_RENDER in WicketApplication.java. The commit message:

   With the default REDIRECT_TO_BUFFER strategy, Wicket added page instance info to the URL. Browser refresh then loaded the stored page (with its already-built component tree) instead of creating a new one, so updated data from singletons like Space was not reflected in the HTML. Switching to ONE_PASS_RENDER keeps URLs clean, so F5 always creates a new page instance that reads current data.

2. 063c82c2 (PR #449, 2026-04-22) — refactor: drop formobj, use PageReference for preview Back. This removed the formobj session-stored form-map workaround that had been compensating for the lost state in PublishPage since 8d1f2c5a (2025-06-05).

So the chain is:

• Pre-Instances sometimes get slow and eventually unresponsive after a while #401: Wicket default REDIRECT_TO_BUFFER — URL carries a page-id, F5 restores stored page instance, form state preserved for free.
• After Instances sometimes get slow and eventually unresponsive after a while #401: ONE_PASS_RENDER — F5 always builds a fresh page, so form state is lost. formobj masked this for the publish flow.
• After refactor: drop formobj, use PageReference for preview Back #449: formobj removed. Nothing compensates anymore. The regression is now user-visible.

The underlying tension

ONE_PASS_RENDER was introduced to solve a real problem: pages held direct references to mutable singleton data (e.g. a Space object), so after the page was restored from the page store on F5, the HTML rendered from stale fields rather than reflecting current singleton state.

REDIRECT_TO_BUFFER and ONE_PASS_RENDER are opposite trade-offs of the same setting:

Strategy	F5 behavior	Form state on F5	Singleton data on F5
REDIRECT_TO_BUFFER (default)	restore stored page	preserved	stale if held by reference
ONE_PASS_RENDER (current)	build fresh page	lost	fresh

You cannot have both via the render-strategy knob alone.

Proposed solution

The Wicket-idiomatic fix is to leave the render strategy at the default REDIRECT_TO_BUFFER and stop holding live data references in the component tree. Use LoadableDetachableModel (LDM) for anything fetched from a singleton/registry.

An LDM serializes only the identifier (e.g. a space id) and detaches at the end of every request. On the next render it re-loads via load(). So even when Wicket restores a stored page tree on F5, anything wrapped in an LDM re-fetches from the singleton before rendering.

This gives us both:

• F5 preserves the stored page instance → form state intact
• LDMs re-fetch on each render → singleton data always fresh

Concrete pattern (e.g. for Space)

```java
IModel spaceModel = new LoadableDetachableModel() {
@OverRide protected Space load() {
return SpaceRepository.get().findById(spaceId); // only spaceId is serialized
}
};
```

Then pass spaceModel to components or use PropertyModel.of(spaceModel, \"label\"). Label, ListView, etc. all accept IModel natively.

Steps

1. Fix the pages PR Instances sometimes get slow and eventually unresponsive after a while #401 was about. Change page fields from Space space / MaintainedResource resource to String id + an LDM. Replace direct space.getXxx() calls with PropertyModel.of(spaceModel, \"xxx\") or short helper LDMs.
   • page/SpacePage.java:53 — private final Space space field captures snapshot.
   • page/MaintainedResourcePage.java:41 — same pattern with MaintainedResource resource.
2. Fix ExplorePage.java:129-133. Replace Model.of(\"back to \" + …findById(contextId).getLabel()) with an LDM that re-resolves the label.
3. Fix HomePage.java:73. Replace the findById(...) snapshot passed to ViewList with an LDM if the home resource is editable mid-session.
4. Remove the ONE_PASS_RENDER setting from WicketApplication.java:139.
5. Verify by F5-testing. Other pages that hold singleton references (ResourcePartPage, SearchPage, DownloadRdfPage, UserPage, ChannelPage, ListPage, PreviewPage, ViewPage, SpaceListPage) are medium-risk — most hold data that is effectively immutable within a session. Convert any that show stale data in testing, following the same LDM pattern.

Wins of the revert

These form pages will immediately get F5-preserves-state behavior back:

• PublishPage
• GenPublishPage
• All forms under page/connector/

Notes

• This issue is independent of (but in the same neighborhood as) #271 (fixed in PR #455), which dealt with shared template-component models being overwritten by URL params.
• The audit above lists only file:line references for the high-risk pages. A fuller list of medium-risk pages is available on request.

[github-actions]

🎉 This issue has been resolved in version 4.27.1 🎉

The release is available on:

• nanodash-4.27.1
• GitHub release

Your semantic-release bot 📦🚀