Skip to content

[lexical-rich-text][lexical-table] Bug Fix: caret no longer cycles around a trailing table - #8949

Closed
LeSingh1 wants to merge 1 commit into
facebook:mainfrom
LeSingh1:fix-7999-caret-cycles-around-trailing-table
Closed

[lexical-rich-text][lexical-table] Bug Fix: caret no longer cycles around a trailing table#8949
LeSingh1 wants to merge 1 commit into
facebook:mainfrom
LeSingh1:fix-7999-caret-cycles-around-trailing-table

Conversation

@LeSingh1

@LeSingh1 LeSingh1 commented Aug 7, 2026

Copy link
Copy Markdown
Contributor

Repro from the issue: table as the last node of the document, caret in the last cell, press Right repeatedly. The caret cycles — block cursor beneath the table, then the root offset before the table, then back into the last cell. Enter at that middle state inserts a paragraph above the table, which is the second half of the report.

Only the second key press is wrong. KEY_ARROW_UP/KEY_ARROW_DOWN already consume the key at $isSelectionAtStartOfRoot/$isSelectionAtEndOfRoot; the horizontal handlers had no equivalent, so at a block cursor on the outer edge of the root the key fell through to the native caret, which walks around the block cursor element.

Both horizontal handlers now consume the key when the collapsed element point is at the root edge beside a child that needs a block cursor. Non-collapsed selections, and block cursors that still have a sibling to move into, are untouched.

Nothing moves in the model on the bad key press, so the reproduction is a browser test under packages/lexical-table/src/__tests__/browser, plus 4 unit tests.

Fixes #7999

…ound a trailing table

With a table as the last node of the document, right arrow out of the last
cell reached the block cursor beneath the table, the next right arrow jumped
back to the root offset before the table (where Enter inserts a paragraph
above it), and the one after that dropped back into the last cell.

The horizontal arrow handlers had no equivalent of the $isSelectionAtEndOfRoot
/ $isSelectionAtStartOfRoot guards that KEY_ARROW_UP/DOWN already use, so at a
block cursor on the outer edge of the root the key fell through to the native
caret, which walks around the block cursor element. Both handlers now consume
the key when the collapsed element point sits at the root edge next to a child
that needs a block cursor, leaving the selection alone.

Non-collapsed selections and block cursors that still have a sibling to move
into are untouched, so collapsing a selection and stepping into a decorator
keep working.
@vercel

vercel Bot commented Aug 7, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
lexical Ready Ready Preview Aug 7, 2026 9:42am
lexical-playground Ready Ready Preview Aug 7, 2026 9:42am

Request Review

@LeSingh1

Copy link
Copy Markdown
Contributor Author

Consolidated into #9054 with the other PRs that share this defect, per @etrepum's note on #9027 and @mayrang's on #9035. Same fix and same tests, one review.

@LeSingh1 LeSingh1 closed this Aug 10, 2026
LeSingh1 added a commit to LeSingh1/lexical that referenced this pull request Aug 10, 2026
… grid indices, cell state and boundary selection

## Description

Eleven small table bugs, consolidated into one branch per the review feedback
on facebook#9027 and facebook#9035. They fall into three groups.

### 1. Grid geometry

Code that uses a row's child index where the table *grid* column index is
required, or that ignores a span. These agree only while every cell has
`colSpan === 1`/`rowSpan === 1`, so they are invisible until a table has merged
cells.

- `$insertTableColumnAtNode` (`packages/lexical-table/src/LexicalTableUtils.ts`)
  walked leftwards from the insertion column looking for the cell to insert
  after. When every grid position at or before the insertion column is covered
  by a `rowSpan` from an earlier row, the walk fell off the left edge and
  appended the new cell to the *end* of the row, so the inserted column did not
  line up. Replaced with the same left-to-right map scan `$unmergeCellNode`
  uses: track the last row-map entry with `startRow === i` at or before the
  insertion column, `insertAfter` it, or `$insertFirst` when the row owns
  nothing to the left. This also removes the labeled `continue`. (facebook#8836)
- `toggleTableColumnIsHeader`
  (`packages/lexical-playground/src/plugins/TableActionMenuPlugin/index.tsx`)
  fed `$setTableColumnIsHeader`, which indexes `gridMap[row][columnIndex]`, the
  result of `$getTableColumnIndexFromTableCellNode` — the cell's index among its
  row's *children*. With an earlier spanning cell in the row, "Toggle column
  header" applied to a column left of the one clicked. It now reads
  `cellMap.startColumn` from `$computeTableMap`. (facebook#9006)

### 2. Cell and table state

Transforms that rebuild a cell or its paragraph and drop what it carried, plus
the `dir` attribute being written onto an element that the export throws away.

- `$insertTableIntoGrid` (`LexicalTableUtils.ts`) copies the template cell's
  `backgroundColor` onto the destination cell but not `verticalAlign`, which
  was added to `TableCellNode` after that loop (facebook#7077). Pasting a table with
  vertically aligned cells over an existing table dropped the alignment. (facebook#9015)
- `$tableTransform` (`LexicalTablePluginHelpers.ts`) pads short rows out to the
  table width with `$createTableCellNode()` and no header state, so a short
  header row was repaired into a mix of `<th>` and `<td>`. The padding cells now
  inherit the `TableCellHeaderStates.ROW` bit from the row's last cell — the
  same reference `$insertTableColumnAtNode` uses when it appends a column. (facebook#9023)
- `$unmergeCellNode` (`LexicalTableUtils.ts`) recomputes the header state of the
  cells it splits off, correctly, but gives them nothing else, so a filled
  merged cell unmerged into one filled cell and a run of blank ones. The split
  cells now keep the original's `backgroundColor` and `verticalAlign`, which is
  the cell's own presentation and not a description of its row or column. (facebook#9025)
- `TableNode.exportDOM` (`LexicalTableNode.ts`) narrows the export to
  `element.querySelector('table')`. When scrollable tables are active
  `createDOM` returns the scroll wrapper `<div>`, which is where
  `ElementNode.exportDOM` writes `dir` — so the wrapper, and the direction with
  it, was thrown away. The direction is now re-applied to the `<table>`, which
  is where `$convertTableElement` reads it back. (facebook#9029)
- `TableObserver.$clearText` (`LexicalTableObserver.ts`) emptied each selected
  cell by constructing a fresh `ParagraphNode`, so Backspace over a multi-cell
  selection also reset the format, style, direction and indent of those cells.
  It now `$copyNode`s the cell's existing paragraph, which carries the element
  state and returns a childless copy. (facebook#9041)

### 3. Selection and caret at the table boundary

- The `KEY_ARROW_LEFT_COMMAND`/`KEY_ARROW_RIGHT_COMMAND` handlers in
  `registerRichText` (`packages/lexical-rich-text/src/index.ts`) had no
  equivalent of the `$isSelectionAtEndOfRoot`/`$isSelectionAtStartOfRoot` guards
  the up/down handlers already use. With a table as the last node of the
  document the key fell through to the native caret, which walks around the
  block cursor element, so the caret cycled: last cell → block cursor → root
  offset before the table → back into the last cell. A new
  `$isBlockCursorAtRootEdge` consumes the key when the collapsed element point
  sits at the root edge beside a child that needs a block cursor.
  Non-collapsed selections and block cursors that still have a sibling to move
  into are untouched. (facebook#8949, fixes facebook#7999)
- `$handleArrowKey` (`LexicalTableSelectionHelpers.ts`) armed the Firefox
  scrollable-table workaround (`setShouldCheckSelectionForTable`) on every
  ArrowDown. From the block cursor below a trailing table ArrowDown moves
  nothing, so the flag survived to be consumed by the next selection change —
  an ArrowUp back into the table — which snapped the caret to the first cell.
  A new `$isSelectionBeforeTable` arms it only when the focus is before the
  table in document order, the only side ArrowDown can enter from.
  (facebook#8963, fixes facebook#6822)
- `$deleteTableRowAtSelection` (`LexicalTableUtils.ts`) removes the table
  outright when the selection covers every row, but never moved the selection
  first, leaving a `TableSelection` anchored to detached nodes. Added the
  `grid.selectPrevious()` that `$deleteTableColumnAtSelection` and
  `TableObserver.$clearText` already do in their identical branches. (facebook#9037)
- `$adjustFocusInDirection` (`LexicalTableSelectionHelpers.ts`) called
  `getCornerOrThrow` on the rect from `$computeTableCellRectBoundary`, which
  *grows* the rect until it contains every merged cell straddling an edge — so
  the anchor is not guaranteed to be at a corner, and Shift+Arrow killed the
  editor update with an invariant. `getCornerOrThrow` is replaced by
  `getAnchorCorner`, which falls back the same way `$extractRectCorners` 24
  lines below already does: to the corner opposite the focus, then to the
  top-left. (facebook#9045)

## Test plan

Unit tests cover ten of the eleven fixes, two browser tests cover the
caret/arrow-key behaviour at the table boundary, and facebook#9006 is a Playwright spec
(`packages/lexical-playground/__tests__/regression/7266-column-header-merged-cells.spec.mjs`).

### Before

Source fixes reverted, new tests kept:

```
$ npx vitest run --project unit packages/lexical-table packages/lexical-rich-text
     × pasting a table carries the cell backgroundColor and verticalAlign 6ms
     × a short header row is padded with header cells 5ms
     × extends the selection instead of throwing 73ms
     × keeps the paragraph format, style and direction of each cell 89ms
     × inserts the new cell in the correct column for rows spanned by a rowSpan cell 7ms
     × exports dir with scrollable tables active 31ms
     × leaves the selection outside the removed table 13ms
     × ArrowRight at the block cursor after the last block stays put 5ms
     × ArrowLeft at the block cursor before the first block stays put 1ms
     × the cells a merged cell splits into keep its backgroundColor and verticalAlign 13ms
⎯⎯⎯⎯⎯⎯⎯ Failed Tests 10 ⎯⎯⎯⎯⎯⎯⎯

 Test Files  8 failed | 23 passed (31)
      Tests  10 failed | 291 passed (301)

$ npx vitest run --project browser packages/lexical-table/src/__tests__/browser/TableTrailingCaret.test.ts packages/lexical-table/src/__tests__/browser/Issue6822BlockCursorArrowKeys.test.ts
   × the caret stops beneath a trailing table instead of cycling around it 172ms
     × ArrowUp from the block cursor below a trailing table returns to the last row 380ms
     × ArrowUp from the block cursor below a table that is the only root child returns to the last row 355ms
 Test Files  2 failed (2)
      Tests  3 failed | 2 passed (5)
```

The remaining new cases pass on both sides of the change — the unstyled-unmerge,
short-body-row, non-wrapped-export, delete-every-column and
still-owns-a-cell-to-the-left cases — which is what pins each asymmetry rather
than the fix itself.

### After

```
$ npx vitest run --project unit packages/lexical-table packages/lexical-rich-text
 Test Files  31 passed (31)
      Tests  301 passed (301)

$ npx vitest run --project browser packages/lexical-table/src/__tests__/browser/TableTrailingCaret.test.ts packages/lexical-table/src/__tests__/browser/Issue6822BlockCursorArrowKeys.test.ts
 Test Files  2 passed (2)
      Tests  5 passed (5)

$ npx playwright test --project=chromium Tables.spec.mjs TablesHTMLCopyAndPaste.spec.mjs 7266-column-header-merged-cells.spec.mjs
 99 passed, 1 skipped

$ npx tsc --noEmit -p .
(clean)
```

Supersedes facebook#8836, facebook#8949, facebook#8963, facebook#9006, facebook#9015, facebook#9023, facebook#9025, facebook#9029, facebook#9037,
facebook#9041, facebook#9045, consolidated per the review feedback on facebook#9027 and facebook#9035.

facebook#9000 and facebook#9011 were part of an earlier revision of this PR and have been
dropped, because each contradicted an existing e2e expectation rather than
extending it:

- facebook#9000 imported `dir` in the `DOMImportExtension` table rules. Google Sheets
  emits `<table dir="ltr">` unconditionally, so `Copy + paste (Table - Google
  Sheets)` in `TablesHTMLCopyAndPaste.spec.mjs` went from `dir="auto"` to a
  pinned `dir="ltr"` on the table, and the rows below it stopped emitting
  `dir="auto"` at all. Honouring a pasted `dir` is a deliberate behaviour change
  for third-party HTML and needs its own PR.
- facebook#9011 made the right-edge resizer widen the *last* grid column a merged cell
  spans instead of its first. `Resize merged cells width (2)` in
  `Tables.spec.mjs` pins the existing behaviour, and both variants widen the
  merged cell identically — they differ only in which unmerged column absorbs
  the change, so this is a behaviour preference that needs a maintainer call.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. extended-tests Run extended e2e tests on a PR

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Bug: Odd selection behavior with table as last element in document

2 participants