fix(web): preserve PDF scroll and zoom across autoreloads#25555

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — This focused PDF viewer fix preserves the current page across workspace preview reloads and adds browser coverage for revision and clamping behavior. Human review is required because it introduces a new oxlint suppression directive for the page-restoration effect.

^{You can add or adjust custom eligibility rules. Learn more.}

Review in Change Stack →

No actionable comments were generated in the recent review.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

Regarding the draft workspace identity finding:

absolutePath is already resolved against the workspace root. Changing roots changes the identity for relative file paths. If the absolute path remains the same, it still refers to the same PDF, so preserving its view is intentional.

thanks @edemaine. the page-relative offset and measuring preceding pages preserve the reader's place even when page sizes change. i kept your implementation and commits, merged current main, reused its extracted layout helpers, and removed the obsolete broad navigation test that main replaced with focused coverage. the reviewers found fit-width drift with normal scrollbars; a stable scrollbar gutter fixes it, and the existing regression now checks a later page's offset at fit width. no settings or workflow changes.

verification:

• the existing regression fails against main: page 3 at custom zoom reloads as page 1 at fit width. it passes with the fix, including mixed page sizes, shrinking pages/documents, horizontal scroll, fit width, and changing documents.
• blacksmith vp run check:changed passed: format, lint, both knip checks, affected typechecks, and 165 tests in four files.
• full app, system chrome, built-in PDF.js path: an actual codex provider command replaces the PDF. page 3, 107% zoom, scrollTop 1750 and scrollLeft 20 survive; the new budget text appears. fit-width also preserves page 2 and its within-page offset (83.65625 px) in a full-app provider reload with normal scrollbars. matching main interaction resets to page 1 and fit width. uploaded images load; recording playback and seeking checked.
• native browser PDF plugins, the electron shell, and native mobile were not exercised. native PDF plugins and native mobile are outside this change's scope. web and desktop share the changed built-in viewer.

all seven reported checks pass and the branch is conflict-free. github's CI and preview workflows are action_required, with no jobs run, pending human approval: https://github.com/pingdotgg/t3code/actions/runs/40440119261. full CI remains unverified. both independent reviewers cleared final head 57d3363583.

review response: CodeRabbit asks for progressive metadata loading on deep-page reloads. this work is already bounded by the 2,000-page and 50 MiB limits, and reads page sizes, not page canvases. awaiting preceding sizes is intentional: the mixed-height regression changes earlier pages and requires their new dimensions to restore the exact saved offset. displaying estimated offsets first would need additional anchor-correction state. i kept the smaller implementation; the suggested 30-second timeout has not been reproduced. largest-document reload performance remains unverified and is a known limit for human review.

before: main resets after the provider replaces the document.

main resets the updated PDF to page 1 at fit width

after: the same updated document stays on page 3 at 107%.

updated PDF preserves page 3, zoom, scroll and shows the new budget text

full-app recording of the provider replacing the PDF while page 3 and zoom stay in place

github does not render this external recording inline. playback and seeking were verified at the uploaded video; inline github playback is unverified.

Thanks @maria-rcks for the edits. LGTM!

Verified in a real dev environment

verified unchanged head 57d3363583a6c0b933f95579d9eef4ff69efa469 on this linux boat vm. reproduced first on github/main (07f7f97609) in a separate worktree. ran vp i and a background vp run dev on each checkout, with only one dev server running at a time.

used repository playwright and /usr/bin/google-chrome, with chrome's "download pdfs" preference selecting the real built-in pdf.js fallback (navigator.pdfViewerEnabled === false). created the project and threads through the app. real codex gpt-6.1-sol turns authored and regenerated the pdfs in a real git repository. no mocked provider, fixtures, or replay data.

• main: page 3 at 162%, scrollTop 2665 and scrollLeft 40 reset to page 1, fit width, and zero scroll after the provider update. navigated back to page 3 to confirm the new v2 title had loaded.
• pr: the same v1-to-v2 provider update preserved page 3, 161%, scrollTop 2649, scrollLeft 40, and the exact 43.828125 px within-page offset. the v2 title appeared. the 1% difference between main and pr starting zoom comes from their fit-width sizing before the same two zoom-in clicks.
• nearby behavior: search found the updated title (1/1). fit width preserved page 2 and its offset within 1 px when the provider enlarged page 1 from 792 to 1000 points. reducing five pages to two clamped page 5 to page 2 and retained 136% zoom. opening another pdf reset to page 1, fit width, and zero scroll.
• exercised the discussion's deep-page concern: a real provider update of a 2000-page report restored page 2000 and its within-page offset within 1 px, showing the new title without a timeout. this is one document, not a general performance claim.

runtime measurements from the actual client:

blacksmith vp run check:changed FETCH_HEAD passed against the pr's merge base daa720a986ca: formatting, lint, both knip checks, affected typechecks, and 165 tests in four files. installed the worker's missing chromium dependency before rerunning. no source changes or pushes were needed.

before: main resets after the update.

main resets the updated planning pdf to page 1 at fit width

after: the pr retains page 3, zoom, and scroll while showing the new title.

pr retains page 3 and 161 percent zoom after the real codex update

real codex update changes the pdf title while preserving page 3 and zoom

inspected the captures and verified uploaded video playback and seeking in t3's browser.

Not verified

electron's native shell, macos/windows, native ios/android, browser-native pdf plugins, and remote/relay/tunnel connections were not exercised on this linux vm. native plugins and native mobile are outside this fix's scope. other providers were not exercised; the changed viewer is provider-independent.

full github ci remains unverified: the head's CI workflow is action_required, pending approval. the scoped blacksmith checks above passed. reopened the synced github comment and confirmed both screenshots render. github does not render the external video inline; playback and seeking work at the uploaded video in t3's browser.

merged main into the existing contributor head and pushed b116aa35b8 without force. preserved both sides of the conflicts, including main's word-wrap regression. updated the lint suppression name and used utility classes for the PDF test wrapper, addressing the review nit. no new files or features.

blacksmith vp run check:changed passes, including 167 tests and 3 compiler cases. both independent source reviewers cleared this head. full-app Chrome verification retained page 3 at 182%, scrollTop 2452 and scrollLeft 20 through a server revision; updated PDF text appeared. current evidence is in the PR body.

the preceding-page metadata finding remains declined for the existing concrete reason: exact restoration after earlier mixed-height pages change requires their new sizes. the work is bounded by 2,000 pages and 50 MiB; progressive correction would expand this focused fix. maximum-size latency remains unverified.

all seven reported checks pass. github reports no conflicts. CI and preview workflow runs are action_required, with no job logs, pending human approval: https://github.com/pingdotgg/t3code/actions/runs/40983714051. macroscope also requests human approval. full CI is unverified, so this is waiting rather than ready.