Skip to content

feat(lists): load closed items, activity, and runs as you scroll - #2044

Open
rodboev wants to merge 28 commits into
mainfrom
pr/load-on-scroll-8-9
Open

rodboev wants to merge 28 commits into
mainfrom
pr/load-on-scroll-8-9

Conversation

@rodboev

@rodboev rodboev commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Readers of the Closed and All tabs in pull request and issue sidebars can now reach the last row by scrolling.

With 400 closed pull requests, scrolling past row 200 loads the next 200. Open a pull request and go Back to return to its row.

  • A refresh keeps the same number of sidebar rows loaded.
  • Threaded Activity loads older items past its first 5,000 on desktop and phone.
  • Flat Activity loads 200 more events as you scroll, up to 5,000, and keeps its narrow-the-range notice.
  • Actions loads older workflow runs down to the oldest run. Refresh reads at most six pages of 50; scroll for older runs.
  • A failed load keeps the rows, shows the error, and offers Retry.
  • Changing a filter, repository, workflow, or time range starts the list from the top.
  • The shared kit-ui pin supplies the end-of-list loader. Escape clears the user picker filter before it closes.
forge-scroll-recapture.mp4
View, theme, and viewport Before After
Pull request sidebar, Closed tab, dark, desktop Before: Closed pull requests stop at Load more After: Closed pull requests load as you scroll
Activity, dark, desktop Before: Activity ends at Load more After: Activity loads as you scroll
Actions runs, dark, desktop Before: workflow runs stop at Load more After: workflow runs load as you scroll

Readers can follow long histories without stopping at a paging button. Keep the loaded window during refresh so older runs and activity remain available while their status changes.

Show page failures beside the affected list so readers can retry without losing their place.
Loading more activity should keep readers at their current position, while a changed filter or sort should start at the top. Refreshing workflow runs must retain the full loaded history so older runs remain visible and current.
Long run histories made each refresh wait for every loaded page. Keep the existing history and cursor while updating the first page so refresh cost stays bounded.

Changing repositories also starts the sidebar at the top, matching the new list scope.
A long threaded history mounts in batches before the end row can load older subjects. Wait for the last initial subject so the scroll check measures paging after that work finishes.
A refresh can shift page boundaries while older runs remain visible. Continuing from the earlier cursor skips runs that moved into the gap. Use the refreshed cursor and preserve provider order when the next page overlaps the retained history.
A refresh can reveal several full pages before reaching retained runs. Remember the end of the last fetched page so each subsequent page appears ahead of older retained rows, even when the pages have no shared run.
A provider can delete runs after readers load them. An exhausted page confirms that no older runs remain, so keeping the cached tail would leave deleted runs visible indefinitely.
A refreshed page can cover a deleted run while older loaded runs remain beyond it. Keep the history after the last returned run so displaced rows remain visible without restoring rows the refresh has already passed.
Readers need current results throughout the history they have loaded. Reusing a cached tail leaves older statuses stale and can retain deleted runs, so refresh follows the current cursor chain before replacing the visible window.

A failed phone batch also needs to remain reachable for retry. Server errors can be temporary across all list reads, including HTTP 500 responses.
The activity scope already includes pull request attributes from the request parameters. Keeping a second copy in the scroll key makes the same filter govern the key twice.
Paging tests need to exercise the idle state as well as failures. Reset shared state between cases and keep each check at the layer that owns its behavior, so unrelated calls and refresh timing cannot satisfy it.
Changing the loaded depth must keep the current scroll position. The optional limit accepts an omitted value, while an explicit undefined value fails the project's exact optional property checks.
Long run histories made refresh work grow with every page loaded. Limit refresh work while keeping the reader's older rows and their continuation, and let throttled reads be retried.

Keep browser checks for returning to Activity at depth so the remaining position regressions are visible.
Returning from another page could leave readers at the top of a long activity feed. Keep its position until the rows have mounted, and retain the phone return marker while restoring the loaded depth.
The activity view now reads its scroll scope when it mounts. Its test fixture must provide that store method so the existing behavior checks can exercise the view.
The updated UI kit checks avatar usage and handles modal Escape events on the document. Use its avatar component and send test keyboard events through the dialog so the checks exercise the current integration.
The shared overlay claims Escape before the app receives it, so the shortcut dispatcher must honor that claim. Activity also needs to calculate its scroll scope without changing pagination state inside a derived value.
The updated segmented control changes the radio's accessible name. Locate the visible Score option within the row order group so the test continues to check the selected order.
Public test fixtures should avoid names that could identify real people or projects. Use consistent example names across each paging scenario.
Paging controls can render while their scroll callbacks are disconnected, and a refresh test can miss whether updated runs reach the selected repository. Exercise those paths through the observable behavior and retain distinct dispatch regressions.
The phone modal handles Escape at the document before the search field can clear its text. Let the field claim that key first so a second press can dismiss the picker.
The updated UI kit describes excluded conditions as left out of results. Keep the filter tests checking the accessible label that users hear alongside the filtered rows.
The frontend formatter rejected these long assertions. Keep them in the required form so the check can validate the refresh regression.
Keep filter reset and retry coverage together while leaving shared component behavior to its own tests. This reduces duplicate cases without losing the paging regression.
Both Activity views start at the top when filters change. Keep that behavior with the viewport owner so the two views cannot drift apart.
@roborev-ci

roborev-ci Bot commented Nov 18, 2026

Copy link
Copy Markdown

roborev: Combined Review (df0afac)

Verdict: Changes require fixes for 3 findings.

Medium

  • frontend/src/lib/stores/workflow-actions.svelte.ts:415-416: After loading 350 or 1,000 GitHub runs with more available, deleting a refreshed run makes scrolling permanently skip an unseen run. The six-page refresh retains an outdated page number; later refreshes cannot recover it.

    Fix: Follow the refreshed page sequence and reconcile overlapping history in provider order so additions and deletions leave no unreachable gaps, preserving loaded rows while fetching missing history and keeping refresh work bounded. Where: readRuns, frontend/src/lib/stores/workflow-actions.svelte.ts

  • frontend/src/lib/stores/workflow-actions.svelte.ts:400-401: After loading all 350 workflow runs, if 300 new runs arrive, refreshing hides 300 existing runs that scrolling cannot recover. With no overlap, the six-page refresh drops cached rows and retains the previous end-of-list state.

    Fix: When refreshed pages contain no previously loaded runs, preserve cached runs outside their coverage and keep intervening history reachable by scrolling. Where: readRuns, frontend/src/lib/stores/workflow-actions.svelte.ts

  • frontend/src/lib/components/ActivityFeed.svelte:79-81: If threaded Activity loses rows after your last scroll, returning can leave you at the top and stop remembering subsequent scrolling. Restoration waits for the saved row count even when fewer rows remain.

    Fix: Finish restoring the scroll position once all remaining rows have rendered, even if fewer rows remain than were saved, and clamp the restored position to the available content. Where: ActivityFeed.svelte, ActivityThreaded.svelte


Reviewers: 3x codex, codex (security), codex (test-value) | Synthesis: codex, 6m19s | Total: 19m58s

Workflow history changes between reads, so retaining an old cursor can skip runs or leave deleted ones visible. Refresh the visible sequence from the provider and continue from its last page.

Activity can shrink while its reader is away. Restore the saved position once the entries that remain have mounted, so later scrolling can save a new position.
Readers need to retry failed pages and keep their place when a refresh fails. The tests need to exercise those outcomes through each list owner and distinguish query changes from refreshes.

Keep the boundary cases while sharing their setup, so a shorter suite still catches lost rows, stale cursors, and failed position resets.
@roborev-ci

roborev-ci Bot commented Nov 18, 2026

Copy link
Copy Markdown

roborev: Combined Review (f5be5cc)

Verdict: No findings at or above medium severity.


Reviewers: 3x codex, codex (security), codex (test-value) | Synthesis: codex, 1m19s | Total: 12m44s

@rodboev
rodboev marked this pull request as ready for review November 19, 2026 01:29

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Development

Successfully merging this pull request may close these issues.

1 participant