feat(web): virtualize large file previews#21466

🟠 High files/LargeTextViewer.tsx:56

An empty text attachment is shown as a load error instead of the intended empty-file state: the worker fetches it with Range: bytes=0-..., which the server rejects with 416, and open treats that response as a failure. Have the worker fetch without a range for zero-length files or treat this response as empty.

Fixed in d9ce5f98aa. The worker treats only a 416 response with Content-Range: bytes */0 as an empty file; other range failures still report errors.

🟠 High files/FilePreviewPanel.tsx:1550

After an editor or agent atomically replaces a truncated file, WorkspaceLargeTextPreview keeps using the old signed asset grant, so the preview fails to load until it is reopened or the URL cache refreshes. Changing the revision query parameter does not refresh the grant returned by useAssetUrlState; refresh the asset URL when workspaceMutationId changes, as the other workspace media previews do.

Fixed in d9ce5f98aa. WorkspaceLargeTextPreview now refreshes its signed asset grant on workspace mutations using the same useWorkspaceMutationRefresh path as workspace media.

🟡 Medium files/AttachmentFilePreview.tsx:357

Clicking Copy while LargeTextViewer is still loading silently does nothing because copyContents.current is unset and the optional call short-circuits. Check for the callback before invoking it and report that the viewer is still loading when it is unavailable.

Fixed in d9ce5f98aa. Copy now reports that the viewer is loading when its callback is unavailable, and still reports worker copy failures.

🟡 Medium files/OfficeViewer.tsx:330

Sheets with rowCount === 0 no longer display “This sheet is empty,” leaving an empty preview that is indistinguishable from a blank or unloaded sheet. The previous empty-state fallback was removed, and VirtualTablePreview is now rendered unconditionally; restore the message for zero-row sheets.

Fixed in d9ce5f98aa. Zero-row sheets show "This sheet is empty." again; verified with an empty XLSX in the app.

🟠 High files/VirtualTablePreview.tsx:254

With more than 500,000 rows and wrap enabled, scrolling skips the remainder of wrapped lines, making that content unreadable. first maps scroll position using a fixed 29px per row, so each row transition moves the next row to the viewport top regardless of the actual wrapped row height; account for measured row heights or preserve the intra-row scroll offset.

Already addressed by 907526b33d. ScaledRows preserves the fractional intra-row offset using the measured row height, and wheel/key scrolling scales by that height rather than skipping wrapped content.

🟠 High files/VirtualTablePreview.tsx:183

A partially returned page is treated as complete, so when the worker later adds rows, this page is never requested again and its missing tail stays blank. The effect only reloads when rows is undefined; also reload when rows.length is less than the page's expected row count based on props.count.

Fixed in d9ce5f98aa. RowPage reloads when the cached page has fewer rows than the indexed count requires; the existing browser test now covers a growing partial page.

🟠 High files/LargeTextViewer.tsx:48

After a preview load fails, changing src or text leaves the old error set, so the component keeps rendering FileSurfaceFailure even if the replacement worker loads successfully. Workspace revision changes therefore cannot recover the preview without remounting; clear error when the effect starts a new load.

Fixed in d9ce5f98aa. New sources clear the previous worker error and pages; the browser regression test loads binary text, replaces it with an empty source, and observes successful recovery.

🟠 High files/VirtualTablePreview.tsx:49

With measured, nonuniform column widths, horizontal scrolling can leave visible columns unrendered, creating blank areas and making those columns unreachable. firstColumn and lastColumn use the uniform columnWidth even though the table layout uses sizes; derive the visible range from cumulative sizes widths.

Fixed in b2bfd086568f0551698f612048d204ff1a06193c. Horizontal virtualization selects columns from the measured cumulative offsets. The browser check verifies continuous viewport coverage before and after sideways scrolling with narrow columns.

🟠 High files/largeText.worker.ts:42

read() replaces malformed UTF-8 bytes with �, so preview pages and “Copy contents” silently return altered file contents; for example, a Latin-1 0xE9 is displayed as �. Decode with a fatal UTF-8 decoder so invalid input is rejected instead of corrupted.

Fixed in b2bfd086568f0551698f612048d204ff1a06193c. The worker decodes fetched bytes with fatal UTF-8 validation, matching the prefix parser and rejecting malformed trailing bytes.

🟡 Medium files/FilePreviewPanel.tsx:1420

Opening a file:line link to a truncated source file now shows the file from the beginning without scrolling to or highlighting the requested line. This branch replaces SourceFilePreview, which handles onFilePostRender, with WorkspaceLargeTextPreview, and the virtualized viewer receives no reveal request; pass the reveal line/request through and implement it in LargeTextViewer.

Fixed in b2bfd086568f0551698f612048d204ff1a06193c. Workspace line requests reach the virtual source viewer, wait for complete indexing, and scroll to the highlighted row. Verified through a real Codex response linking line 500 of a 30,000-line log, plus browser checks for repeated wrapped-line requests through line 999.

🟠 High files/LargeTextViewer.tsx:94

Rows in prefetched page 1 remain visible after the worker opens a newer version of the file, so the table can show stale rows alongside current rows and copy returns different contents. The meta handler leaves pages intact; clear the prefetched pages when worker indexing reports metadata so they are reloaded from the worker.

Fixed in b2bfd086568f0551698f612048d204ff1a06193c. The first worker metadata clears every prefetched page and pending request before replacing the count. The browser check serves an old prefix and a newer worker file, then verifies current page-one content.

🟠 High files/VirtualTablePreview.tsx:398

A single wrapped line taller than the viewport cannot be scrolled through: when the row-count track is shorter than the viewport, every positive position.top triggers the bottom-alignment branch and jumps the content to the bottom. Only apply that branch when the synthetic track exceeds the viewport.

Fixed in 8f626041ad82240a00d1f1e79f3a877f69c55acd. Bottom alignment now requires a synthetic track taller than the viewport. The new browser check reproduced the jump before the fix and passes afterward; verified scrolling within a 2,413 px-tall spreadsheet row in the real client.

🟡 Medium files/VirtualTablePreview.tsx:403

A wrapped file with count=1 snaps back to the row's beginning after each 29px of scrolling, so pressing End does not reveal the row's end. Because first stays at 0, the % 1 here discards the completed scroll distance; preserve the full offset when there is no virtual row advance.

Fixed in cb56109c71. When every row fits the row window, the content uses native scrolling and End uses the actual scroll height. The regression check now exercises End and Home, and the real client reaches the bottom of the 2,413 px spreadsheet row and returns to its beginning.

🟠 High files/VirtualTablePreview.tsx:270

When wrapped source rows are taller than 29px, scrolling to the end skips rows near the bottom; with 100 rows, a 400px viewport, and 1,000px-tall rows, rows 88–98 are never visible. maxFirst at line 270 reserves only 29px per row, and the bottom alignment at lines 401–402 jumps directly to the final viewport; let the end mapping account for the actual heights of the remaining rows.

Fixed in 0bc4d7872d. The final mounted rows now scroll natively using their measured height, so the virtual track no longer jumps over them. The new browser check failed before the fix and now visits each final tall row; the actual XLSX view reached rows 88–99 in order and aligned the last row bottom on End.

🟠 High files/AttachmentFilePreview.tsx:160

Workspace .txt and .csv files whose read failed with binary_file are shown as empty, and copying them returns an empty string. Their unknown sizeBytes is 0, which selects LargeTextViewer; that viewer treats zero as a confirmed empty file and skips fetching the source. Exclude workspace resources with unknown sizes from this path while preserving it for confirmed empty files.

No longer relevant as of 91bbedecc40c3bfe72462f388a3139a3bea54070

Fixed in 91bbedecc4: unknown-size workspace resources follow the normal read path, while confirmed zero-byte attachments retain the empty shortcut. Both draft and thread workspace cases fetch the source and report binary content in the existing browser suite.

🟠 High files/FilePreviewPanel.tsx:1417

Manual Refresh leaves the large-file rows and Copy contents stale when an external edit changes a file that remains truncated. FileBrowserPanel calls file.refresh, but that updates neither this branch’s key nor workspaceMutationId, so WorkspaceLargeTextPreview does not reload its asset or restart LargeTextViewer; pass a refresh generation into this preview or trigger its asset refresh from the explorer callback.

No longer relevant as of 91bbedecc40c3bfe72462f388a3139a3bea54070

Fixed in 91bbedecc4: manual Refresh increments the worker preview generation, reloading the asset and its full-file copy buffer. In the real app, an external same-size edit appears after Refresh and Copy returns all 1,800,000 updated characters.

🟠 High files/LargeTextViewer.tsx:95

Two Copy requests issued while indexing can replace the preview with a permanent error, leaving the file unusable until retry. The first worker error clears copyPending, so the second error reaches setError here even though it belongs to a Copy request; correlate worker errors with their requests or suppress overlapping Copy requests.

No longer relevant as of 4e0cf34ab6a63ba7c21b1e3919f9ca35787fa54b

Fixed in 4e0cf34ab6 and covered in 230e4005d0: an overlapping copy is rejected before posting another worker request, leaving the first request intact. The browser case verifies the original and subsequent copies succeed and the preview has no error.

🟡 Medium files/VirtualTablePreview.tsx:339

An uncached reveal can be completed against a numbered placeholder before its page loads: the row header makes row.textContent nonempty even when the data row is missing. When the page arrives, preceding tall rows can push the target out of view, and handledReveal prevents another attempt. Check that the corresponding getRows entry exists before processing the reveal.

No longer relevant as of 4e0cf34ab6a63ba7c21b1e3919f9ca35787fa54b

Fixed in 4e0cf34ab6 and covered in 230e4005d0: reveal waits for the corresponding getRows entry before measuring the target. The browser regression loads the wrapped target after its numbered placeholder and verifies it stays visible across repeated reveals.

🟡 Medium files/VirtualTablePreview.tsx:272

When count grows from the worker’s partial metadata to the full file size, the existing top is interpreted using the new scale, so the preview jumps to an unrelated row—for example, roughly line 1,000 becomes roughly line 3,600 for a million-line file. Tall-row previews also jump when maxFirst changes from zero. Preserve the current logical row and within-row offset, then recompute scrollTop for the new scale.

Fixed in 230e4005d0: scroll scaling retains the logical row and within-row offset when indexed counts grow, including the transition from a native tall-row tail. Browser cases cover growth from 5, 100, and 10,000 lines to one million lines and preserve the visible row.

The scaled-region case is fixed, but tail-to-scaled conversion still uses a single measured row height for all native-tail pixels. With mixed-height wrapped rows after the first tail row, it can still map to the wrong logical row, so the original tall-row case is not fully resolved.

fixed in 790ed91b68. native-tail pixels now cross the measured heights of each mounted row before conversion to the scaled track. the existing growth regression covers mixed heights from 5 and 100 rows as well as the scaled region; all 91 browser checks pass.

🟡 Medium files/VirtualTablePreview.tsx:409

Pressing End can leave the final line offscreen when the last rows wrap: the handler sets scrollTop to the current scrollHeight before the native tail is mounted and measured, and does not scroll again when that content expands. Keep the end-scroll pending until the tail is loaded and measured, then scroll to the updated bottom.

fixed in ea11e4599f. End remains pending until the final page is loaded and the wrapped tail is measured, then scrolls to the updated bottom. the existing tall-row browser regression now presses End before the final page arrives and verifies the last row is visible.

External image Major · Performance & Scalability · Quick win

Keep the large-text viewer loaded when the file is unchanged.

src includes workspaceMutationId. The LargeTextViewer key and its load effect both depend on src. Each workspace mutation therefore remounts the viewer and starts a new worker. The new worker downloads and indexes the whole file again, up to 256 MB, and the user's scroll position resets.

The workspace mutation id changes after every agent command, so this repeats while the agent works on unrelated files.

useWorkspaceMutationRefresh also calls refreshAssetUrl(). That call mints a new signed URL, so assetUrl.url changes a second time. One mutation therefore causes two full reloads.

The media-file claim already returns fileRevision. WorkspaceVideoPreview uses that value to keep an unchanged file loaded, and the comment at lines 293-295 explains why. Use the same rule here:

• Use fileRevision as the revision, and fall back to workspaceMutationId when the server does not report one.
• Change the loaded URL only when the revision changes.

✅ Confirmed as addressed by @maria-rcks

Fixed in d359115c8a. The loaded signed URL now stays pinned to the file identity and server revision, with the workspace mutation as the legacy fallback. Retry and the existing manual explorer refresh still reload; the real-worker browser regression verifies unchanged scroll and worker identity across unrelated mutations and URL renewal, then reloads for file changes and manual refresh.

@maria-rcks The supplied code addresses this finding. loaded.url stays unchanged while the file identity and fileRevision stay unchanged, with workspaceMutationId as the legacy fallback. Retry and manual explorer refresh still trigger a reload.

I did not run the browser regression.

📝 I couldn't check this conversation for new learnings, so no new learning was saved.

If you found this review helpful, would you consider giving us a shout-out on X?

Thank you for using CodeRabbit!

✅ Review thread resolved.

_{You are interacting with an AI system.}

External image Minor · Functional Correctness · Quick win

A failed URL renewal replaces the loaded large-text viewer with a placeholder that has no retry action.

In WorkspaceLargeTextPreview, a Failure asset state takes the early return. The branch renders "Unable to load file preview." and unmounts LargeTextViewer. That removes the loaded rows, scroll position, reveal state, and the copy and save actions. The placeholder has no retry control. The viewer's onRetry is passed only to LargeTextViewer, so it is unavailable in this state. The preview returns only after a later successful renewal. That renewal can come from another workspace mutation or from the 30-minute refresh interval.

The Failure state cannot show whether the file changed. A renewal failure also can mean the file was deleted, so keeping the old content is not clearly correct. For that reason, this comment does not claim that the loaded content is still current or that the pinned URL should stay in use. A retry action restores recovery without making the stale content the default.

Possible fix: add a retry action to the failure placeholder that calls refreshAssetUrl, and use the shared button primitive from components/ui.

✅ Addressed in commit 7ba6c39

fixed in 7ba6c39d07: failed authorization now offers the existing file-surface retry action. retry renews the pinned URL even when the file revision stays unchanged; the real-worker browser regression, web typecheck, lint and build pass.