This skill builds on core and table-features. Cell spanning is stateless and fully derived: a memoized table-level span index is rebuilt from the rows that are actually rendered, and per-cell reads are O(1) lookups against it. Sorting, filtering, pagination, and row pinning only change adjacency; spans follow automatically and never persist.

Setup

import { cellSpanningFeature, tableFeatures } from '@tanstack/table-core'

export const features = tableFeatures({ cellSpanningFeature })

Core Patterns

Opt columns into value-based row spanning

const columns = [
  { accessorKey: 'region', spanRows: true },
  {
    accessorKey: 'createdAt',
    spanRows: ({
      anchorValue,
      value,
    }: {
      anchorValue: unknown
      value: unknown
    }) => sameMonth(anchorValue as Date, value as Date),
  },
]

spanRows: true merges adjacent rendered rows whose values match under Object.is. Nullish values never merge under the default comparison; a predicate can opt in. Predicates are anchored: every candidate row is tested against the run's first row (anchorRow/anchorValue), which keeps runs transitive.

Declare column spans per row

const columns = [
  {
    accessorKey: 'label',
    spanColumns: ({ row }: { row: { original: { isSummary?: boolean } } }) =>
      row.original.isSummary ? Infinity : 1,
  },
]

spanColumns counts visible columns in render order and clamps to the end of the cell's pinned region, so Infinity means "the rest of my region" and a span never crosses a start/center/end pinning boundary. Cells only join a vertical run when their column spans match, so a full-width summary row never merges into the data run above it.

Render with the skip-covered-cells loop

{
  row.getVisibleCells().map((cell) => {
    const rowSpan = cell.getRowSpan()
    const colSpan = cell.getColSpan()
    if (rowSpan === 0 || colSpan === 0) return null
    return (
      <td key={cell.id} rowSpan={rowSpan} colSpan={colSpan}>
        {flexRender(cell.column.columnDef.cell, cell.getContext())}
      </td>
    )
  })
}

A span of 0 means the cell is covered; cell.getIsCovered() is the same check as one call. This mirrors the header.rowSpan convention for uneven header trees.

Disable

enableCellSpanning: false as a table option disables everything; the same flag on a column def opts a single column out and wins over the table option.

Common Mistakes

[CRITICAL] Rendering rowSpan={0} instead of skipping the cell

Wrong:

<td rowSpan={cell.getRowSpan()} colSpan={cell.getColSpan()}>

Correct:

cell.getIsCovered() ? null : (
  <td rowSpan={cell.getRowSpan()} colSpan={cell.getColSpan()}>
)

In HTML, rowspan="0" is valid and means "span to the end of the row group", so the wrong version does not render an empty cell. It merges the covered cell down the entire tbody and shifts every later cell out of its column.

[HIGH] Expecting spanning to sort or group the rows

Wrong:

const features = tableFeatures({ cellSpanningFeature })
// data arrives unsorted; the table renders zero merged cells

Correct:

const features = tableFeatures({
  cellSpanningFeature,
  rowSortingFeature,
  sortedRowModel: createSortedRowModel(),
  sortFns: { alphanumeric: sortFn_alphanumeric },
})
// sort by the spanned column, or emit data with equal values adjacent

Value-based spanning merges adjacent equal values only. It does not reorder rows, and it is not grouping; register columnGroupingFeature when collapsible group rows are what is wanted.

[HIGH] Precomputing spans from the source data

Wrong:

const spans = useMemo(() => computeSpans(data), [data])

Correct:

const rowSpan = cell.getRowSpan()

Spans derive from the final row model. A span precomputed from data survives sorting, filtering, and page changes, producing ragged rows the moment any of them changes.

[MEDIUM] Assuming a run continues across a page boundary

A run is clipped by the paginated row model. The next page opens a fresh cell for the continuing value; nothing carries over, and nothing needs to.

[MEDIUM] Reimplementing merge-aware selection on top of cellSelectionFeature

When both features are registered, selection is already merge-aware: ranges expand at derivation time to fully enclose any merge they touch (includes and excludes alike), navigation crosses a merge in one step, and getSelectedCellCount() counts a merge once. Do not pre-expand stored selection corners yourself; stored corners stay stable while sorting or paging changes the merges, and pre-expanded corners go stale.

[MEDIUM] Spanning a grouped column

spanRows is ignored while its column is grouped, and grouped rows never join runs in any column. Grouping already collapses repeated values into group rows; spanning them again would merge twice.

API Discovery

Inspect node_modules/@tanstack/table-core/dist/features/cell-spanning/ for the full API: cell.getRowSpan(), cell.getColSpan(), cell.getIsCovered(), table.getCellSpanIndex(), and the spanRows/spanColumns/enableCellSpanning column and table options.