Skip to content

armgit / bento

Public Imported from GitHub
0

Bento, the office suite that fits in a file

32.4 KB
Raw Blame

The spreadsheet kind, reviewed as a spreadsheet

What someone who has used Excel, Numbers or Sheets reaches for on a kind: 'canvas' sheet, and what happens when they do.

What this was measured against. A npm run build:single of commit 19a82a4 ("dash: the spreadsheet kind opens, and =SUM below the data lands") with a clean working tree, served from a private copy so nothing could overwrite it mid-measurement. Every "today" below is an observation from that build, not a reading of the source, unless it says code-verified — which means I read the branch and could not exercise it through the UI (a download I could not open, a print I could not trigger).

Work in flight — read this before acting on §2.1. While this review was being written, dash/src/cellprops.ts, dash/src/cellprops.css and scripts/test-dash-cellprops.ts appeared untracked in the tree, and panels.ts gained a buildCellProps call in the spreadsheet arm. That is the per-cell formatting panel — format kind (general/number/currency/percent/date/ text), decimals, thousands separator, currency symbol, custom pattern, live preview, bold, align, text colour, background, and a typed-input reader that keeps leading zeros in a text-formatted cell. Items 2.1 and 1.2 below are therefore probably already answered by somebody else's uncommitted work. They are still described, because the shipped build does not have them and because the parts that pass carry over: that panel does not add italic, underline, borders, wrap, merged cells, or a way to read a cell note.


The one-paragraph answer

The engine is better than the surface. Formulas work, including cross-sheet references, ~90 functions with XLOOKUP/SUMIFS/VLOOKUP, correct cut-versus-copy reference translation on paste, a real unbounded frontier, sparse storage, a complete keyboard set and clean per-edit undo. What is missing is almost entirely reach: nothing on a spreadsheet can be formatted, no row can be inserted, ⌘F finds nothing, six of the eight toolbar buttons throw into the console and show the user nothing, and every export path either drops the sheet or exports a different one. Two of those are silent data loss and should be fixed before anything else on this list.


1 · Loses or misstates data

1.1 Fill down destroys the formula and writes alternating blanks

Today. E4 holds =B4*C4. Select E4:E7, press ⌘D (or drag the fill handle — same code path). Measured result:

before   E4 {"f":"=B4*C4"}   E5 —   E6 —   E7 —
after    E4 {"v":37.5}       E5 —   E6 {"v":37.5}   E7 —

The source formula is replaced by its last computed value, and the fill lands on every other row. With a plain constant it is the same: G4 = 1, select G4:G8, ⌘D → G4 1, G5 blank, G6 1, G7 blank, G8 1.

The cause is in Grid.fillDownSelection (grid.ts:1349): it seeds fillSeries with the top two rows of the selection unconditionally, so a lone value at the top pairs with the blank beneath it and repeats the pair; and the seeds come from valueAt, which returns the computed number for a formula cell, so the formula never travels.

Cost. The highest on this page. This is the second-most-used gesture in a spreadsheet, it is silent, it hardcodes numbers over live formulas, and the alternating blanks are easy to miss in a long column. It is destructive in a way undo will not save you from a week later.

Format? Nothing needed. writeCanvasClip already translates references correctly — a copy/paste of E4 down does the right thing today. Fill should route through the same translation.

Files. dash/src/grid.ts (fillDownSelection, and the canvas fill-handle wiring in wireCanvas); possibly dash/src/select.ts (fillSeries seed contract). Size: half a day. Note this bug is in code shared with the dataset kind and is very likely present there too — I did not measure it there.

1.2 A leading zero is destroyed on entry, with no escape hatch

Today. Typed or pasted, measured:

typed stored shown
007 {v: 7} 7
00123 {v: 123} 123
0044 {v: 44} 44
'007 {v: "'007"} '007

The apostrophe prefix — the universal "keep this as text" convention — is not recognised, so it is stored and displayed as a literal character. There is no per-cell text type to set instead (see 2.1), so there is no way at all to enter a zero-padded code: part numbers, UK phone prefixes, US zips, GL account codes.

Cost. Data loss on typing, immediately visible but unrecoverable without retyping the whole column somewhere else. Ranked below 1.1 only because it is visible.

Format? Supported: CanvasCell.format exists, and '@' is the Excel text mask. This is a canvasValue rule plus a format to consult.

Files. dash/src/grid.ts (canvasValue, canvasCellEdit), plus the formatting UI. Size: hours, once 2.1 exists. Probably already done in the untracked cellprops.ts (readTypedNumber, isTextFormat, TEXT_PATTERN = '@').

1.3 Export Excel silently drops every spreadsheet sheet

Today. Code-verified. exportXlsx (xlsx.ts:1314) begins doc.sheets.filter((s): s is TableSheet => s.kind === 'table'). A spreadsheet sheet is not in that list and produces no finding — the only sheet-skipped finding fires when there are no table sheets at all. I ran Data ▸ Export Excel from a workbook containing my invoice and got no banner, no warning, no findings panel.

So: workbook with one dataset and three spreadsheets → a .xlsx with one worksheet, and nothing said. This is the reverse of the app's stated principle that a lossy export must name what it flattened.

Cost. Data loss on the path people use to hand work to a colleague, with a plausible-looking file as the evidence. Silent.

Format? Not a format question — a spreadsheet sheet maps to an ordinary xlsx worksheet almost exactly (this is what docs/dash-sheet-kinds.md §Export already argues). CanvasCell already carries value, formula, format, bold, colour, bg, align; StyleBook already writes number formats and fills.

Files. dash/src/xlsx.ts (a writeCanvasSheet beside writeSheet). Size: two to three days for a good first cut. A one-line finding ("Invoice was not exported — spreadsheet sheets are not written yet") is one hour and should land immediately regardless.

1.4 Export CSV exports a different sheet than the one on screen

Today. Code-verified (main.ts:1026): exportCsv takes store.doc.sheets.find((s) => s.kind === 'table') — the first table sheet in the workbook, never the sheet you are looking at. On a spreadsheet you get a CSV of some other sheet's data under your workbook's filename; on a workbook with no table sheet at all you get a silent no-op.

Already logged in docs/dash-release.md §3, but its cost rises on this kind: there it exports the wrong tab, here it exports a tab you are not on and the one you are on cannot be exported at all.

Format? No. CSV of a spreadsheet needs a stated convention — Excel's is "the used range, values only", and canvasUsed already computes the used range.

Files. dash/src/main.ts (exportCsv). Size: an hour for the wrong-sheet bug; half a day to add the canvas case with a dialog line saying which range is going.

1.5 An imported .xlsx worksheet always becomes a dataset

Today. Code-verified (xlsx.ts:1080): every imported worksheet is built kind: 'table'. An ordinary Excel worksheet — cell-typed, per-cell formats, formulas anywhere — is forced through a column model that must guess a type per column and cannot hold a format per cell. CellOverride.xlsxF exists precisely because import meets formulas the column model cannot place.

Cost. Every Excel file anyone brings arrives in the wrong kind, and the kind that would have held it faithfully is the one this review is about. Ranked here rather than higher because the loss is visible on arrival.

Format? No — both kinds exist. This is a routing decision at import, plus the honest converse of §1.3.

Files. dash/src/xlsx.ts. Size: two days, and it should wait until export (1.3) exists so the round trip can be tested as a pair.


2 · Cannot be done at all

2.1 There is no way to format a cell — but the format and the renderer already do it

Today. The properties panel on a spreadsheet is one paragraph: "A spreadsheet types each cell on its own, so there are no column properties to set here. Formatting follows the cell." ⌘B does nothing (measured: A3 stayed {v:"Item"}). There is no context menu, no format menu, no toolbar affordance. An invoice reads 12.50 / 4.25 / 99 / 278 / 333.60 — no currency, no aligned decimals, no bold header row.

But the format and the paint are both finished. I committed styled cells straight into the document and the grid drew every one of them:

  • bold: true, bg: '#FFE9B0', color, align: 'center' — all rendered (grid.ts:1818–1823, which even carries a comment saying "setting them is a later pass").
  • format masks resolved through formatValue: '£#,##0.00' → £12.50, £99.00, £333.60; '0.0%' on 0.5 → 50.0%; '0%' → 50%.
  • The document validated clean afterwards.

So this is UI-only work. Nothing about the file format has to change.

Cost. "Looks unfinished" for a scratch pad; "cannot be sent to anyone" for an invoice, a budget or a board pack. Everything a spreadsheet is for ends in somebody reading it.

Files. dash/src/panels.ts (the spreadsheet arm of renderProps), a new cellprops.ts, dash/src/main.ts (⌘B and friends), dash/src/store.ts (a setCanvasCells style patch that touches style without touching value). Size: two to three days — and this appears to be in progress right now (see the note at the top).

Not covered by that in-flight panel, and still absent afterwards: italic, underline, font size/family, per-cell borders, and text wrap. CanvasCell is open ([extra: string]: unknown) so these are additive; borders and wrap need render work, the rest are one CSS declaration each. Rough: a day for italic/underline/size, two days for borders done properly (adjacent-cell resolution is the hard part).

2.2 Six of eight toolbar buttons are enabled, do nothing, and throw

Today. Every one of these is fully enabled on a spreadsheet, gives the normal hover treatment, and on click throws Uncaught Error: grid needs a table sheet out of Grid.sheet (grid.ts:555) with nothing shown to the user. Measured, one console throw per click, at four distinct call sites:

button what happens today
Formula Uncaught (in promise) grid needs a table sheet; no dialog
Chart throws; nothing appears
3D throws; nothing appears
Pivot throws; nothing appears
Story ▸ Capture current view throws; the panel says "Filter, sort and chart the sheet" — a spreadsheet can do none of the three
Dashboard works, and shows a different sheet's data — the canvas is replaced by tiles about Pipeline while the tab strip, the cell reference box and the properties panel all still say you are on the spreadsheet
Import CSV / Excel fine as-is — always adds a new sheet and switches to it; nothing about the current sheet is implied
Export CSV / Excel broken, see §1.3 and §1.4

Grid.sheet's own comment says every caller guards the throw in a try. Six of them do not.

Cost. Not data loss, but it is the thing the owner noticed first and it is what makes the kind read as unfinished. A button that is bright, enabled and inert is worse than one that is greyed with a reason — the app already does the greyed-with-a-reason thing correctly one control away, in the tab menu: "Rename — Only a table sheet can be renamed in this build."

What each should do. The honest answer for most of them is that a spreadsheet has no typed columns, and chart/pivot/3D/story all need typed columns. docs/dash-sheet-kinds.md already names the bridge that fixes this and it does not exist — I searched: there is no "Make this a dataset", no range promotion, and no "open a dataset as a spreadsheet". Without it a spreadsheet is a dead end: no chart, no pivot, no filter, no sort, no conditional format, no SQL, ever.

  • Chart — work differently, and this is the real product question. "Chart a spreadsheet range" should be: select A3:D6, click Chart, and dash reads the first row as labels and each numeric column as a series — the same rule tableToChart already uses in slides. Do not silently promote to a dataset behind their back and do not require it first; a person charting a block of numbers has not agreed to a schema. The honest middle is to chart the range live and offer "Make this a dataset" in the chart's own panel for anyone who wants filters and pivots on it. Until that exists, Chart should be disabled with the reason "Select a range and make it a dataset to chart it" — not silently thrown.
  • Pivot — disabled with a reason. A pivot needs typed columns and a grain; promoting first is the correct extra step and the one that makes the pivot trustworthy. Say so on the button.
  • 3D — disabled with a reason, same argument, plus it needs three numeric columns it cannot identify without types.
  • Formula — hidden. It means "add a formula column", which does not exist on this kind. The cell formula is already the better answer here and it is reachable by typing =. Leaving a button labelled Formula on the one sheet where formulas are per-cell teaches the wrong model.
  • Dashboard — work differently. A dashboard is workbook-scoped, so showing it from a spreadsheet is defensible — but it must say whose numbers these are at the top, not only in a tile caption, and it should not silently replace the sheet you were editing.
  • Story — disabled with a reason. A story step captures a view (filters + sorts + chart); a spreadsheet has no view to capture.
  • Import — works as-is. No change.
  • Export CSV / Excel — work differently, per §1.3/§1.4. Both should export the sheet you are on, and say what they left behind.

Files. dash/src/main.ts (button enablement + reasons, driven off grid.isCanvas), dash/src/grid.ts (a range-to-series reader for the chart case), dash/src/tabs.ts (the promote command). Size: disabling with reasons and hiding Formula, one day and it removes the whole complaint. Chart-a-range, three days. The promotion bridge, a week.

2.3 ⌘F finds nothing on a spreadsheet, and says nothing

Today. ⌘F opens the find bar normally — scope pickers, Aa, Whole cell, Replace and all. Typing Widget on a sheet whose A4 is Widget highlights nothing, reports no count, and throws grid needs a table sheet into the console. FindUI.targets (find.ts:751) calls this.grid.sheet, which throws for this kind; find.ts:761 and :884 also filter to kind === 'table'.

Cost. High, and specifically ironic: find.ts's own header explains that a windowed grid which does not claim ⌘F is a grid that lies. The spreadsheet is windowed in both axes and is unbounded, so it is the kind where the browser's own find is most useless — and it is the kind where dash's find is broken. Replace is gone with it, so there is no bulk edit of any sort.

Format? No.

Files. dash/src/find.ts (a SearchTarget built from the sparse cell map — it is a walk over Object.keys(cells), cheaper than the dataset case). Size: one to two days including Replace.

2.4 No insert or delete of rows and columns, and no context menu at all

Today. Right-clicking a cell on a spreadsheet does nothing — the browser's own menu appears, the selection does not even move. This is deliberate (grid.ts:2008 explains that openCellMenu is a menu about a dataset and would crash), but the consequence is that a row cannot be inserted or deleted at all, by any gesture: no menu, no keyboard binding (⌘⇧+ and ⌘− are unbound — confirmed against the shortcut card), no panel control. Same for columns.

Half the machinery is already written and unused: Grid.shiftFormulas (grid.ts:470) exists precisely to renumber canvas cell formulas across an insert or delete, with a comment about committing it in the same step as the structural patch. Nothing calls it for this kind, because rowcol.ts is TableSheet throughout.

Cost. Very high for real use. "I need another line on this invoice" is the first thing anyone does, and today the only answer is to select the block below and move it by cut-and-paste — which does not update formulas elsewhere that pointed into it (writeClip's own comment admits this).

Format? No. A canvas insert is a re-key of the sparse map plus shiftFormulas, plus the same shift over cols / rows size maps.

Files. dash/src/rowcol.ts (canvas variants), dash/src/store.ts (the patch op), dash/src/main.ts (openCellMenu for this kind), dash/src/select.ts (the keys). Size: three days, and it is the single largest usability win after §2.1.

While that menu is being built, it is also the natural home for: clear contents / clear formats, paste special (values only, formats only — no form of paste-special exists today), hide row/column, and insert note.

2.5 Dates are inert text — no arithmetic, no formats, no sorting

Today. canvasValue deliberately stores 2026-01-14 as a string (the comment says so, and says storing text is the answer that cannot be wrong). Measured consequences on a spreadsheet:

  • =A18-A19 between two date cells → #VALUE!
  • =A18+1 → #VALUE!
  • format: 'yyyy-mm-dd' on a serial 45678 → renders 45678; formatValue has no date branch at all (if (t === 'date') return String(v))
  • TEXT() is not implemented → #NAME?, so there is no workaround

TODAY() works and returns an ISO string, which cannot then be added to.

Cost. "Due date = invoice date + 30" is not expressible. Neither is "days outstanding", ageing buckets, or a month rollup. For anyone doing invoices, budgets or plans this is a hard wall, and the wall is invisible until you hit it.

Format? Partly. The storage decision (ISO string vs Excel serial) is a format decision and worth making explicitly — ISO strings are more honest and diff better; Excel serials are what round-trips. My reading is that ISO strings plus date-aware arithmetic in formula.ts is the right answer and does not change the format at all; EDATE, EOMONTH, DAYS, WEEKDAY, YEAR/MONTH/ DAY and DATE are already implemented and already speak this shape.

Files. dash/src/formula.ts (date coercion in the arithmetic operators), dash/src/format.ts (a date branch). Size: two days. Uncertain: I did not test what the existing date functions accept, so the coercion may already be half-present.

2.6 Freeze panes: absent, and the model has no place for it

Today. A dataset sheet's panel offers Frozen rows and Frozen columns (I can see both on the Pipeline tab). The spreadsheet panel offers nothing, and CanvasSheet has no field for it — freezeAt / readFrozen in rowcol.ts are TableSheet only. A spreadsheet, being the unbounded kind, is exactly where a header row scrolls away first.

Format? Yes — a small additive one: CanvasSheet.frozen?: {rows, cols}.

Files. dash/src/model.ts, dash/src/grid.ts (the canvas paint already computes colLefts and rowSizes, so a frozen band is layout not architecture), dash/src/panels.ts. Size: two days.

2.7 Merged cells: absent, and the model cannot express one

Today. Nothing in the model mentions merging; docs/dash-sheet-kinds.md explicitly puts it out of scope for the first cut, which was the right call.

Cost. Low for correctness, real for presentation — a title across A1:D1 is the first thing on every invoice ever made, and today it simply spills.

Format? Yes, a real format decision — a merge is a range that behaves as one cell, and it touches selection, navigation, paint, formulas, copy/paste and xlsx. CanvasSheet.merges?: string[] (A1 ranges) is the cheap shape.

Size: a week done properly. Worth deferring; worth deciding the field name now so it does not arrive as an escape hatch later (PLATFORM §3).

2.8 Conditional formatting cannot reach a spreadsheet

Today. Rules live at sheet.condfmt[colId] (grid.ts:869) — keyed by column id. A spreadsheet has no columns, so the whole feature is inapplicable, and there is no UI entry point.

Format? Yes. A spreadsheet needs range-keyed rules, e.g. CanvasSheet.condfmt?: Array<{ range: string; rule: Rule }>. condfmt.ts's evaluation (evaluateRules, colorScale, data bars, top-N, duplicates) is already pure over a value vector and would be reused unchanged.

Files. dash/src/model.ts, dash/src/grid.ts, a UI. Size: three days once the cell panel (2.1) exists to hang it off.

2.9 Data validation and named ranges do not exist anywhere in dash

Today. I searched the whole of dash/src: no data-validation concept (the validate.ts in the tree is document validation — findings about a corrupt file, not a rule on a cell), and no named ranges. a1.ts lexes qualified names like Sheet1!A1 and Table1[Col] but there is no name table.

Cost. Moderate. Validation is what stops a shared budget filling with "n/a"; named ranges are what makes a model readable (=Revenue-Costs). Neither is a first-week item, but both are format additions and so worth naming early.

Format? Yes, both. CanvasSheet.validations? (range → rule) and DashDoc.names?: Record<string, string> are the obvious shapes; the second is workbook-scoped and would serve the dataset kind too.

Size: three days each.

2.10 A cell note is a stripe you cannot read or write

Today. CanvasCell.note renders as dg-noted → a 2px accent stripe down the cell's left edge (styles.css:830). There is no title attribute, no hover, no popover, and no way to enter one. I set a note through the document API and got the stripe with no way to read the text back.

Separately, the comments feature refuses this kind honestly: the 💬 panel says "No table sheet is open." — which is at least a refusal, though it names a wire word and denies that the sheet you are looking at is open.

Cost. Low today (nobody can create one), but it will become misinformation the moment an imported .xlsx brings comments in.

Files. dash/src/grid.ts (a title, two lines, ship it now), dash/src/comments.ts (canvas anchors keyed by A1 rather than rid). Size: ten minutes for the tooltip; two days for real threads.

2.11 No point-and-click reference building

Today. Measured: press = in F13, then click D4 to pick up the reference. The edit commits, F13 is stored as {v: "="} — a text value, not even a formula — and the selection moves to D4.

There is also no function autocomplete, and no highlight of the cells a formula being edited refers to.

Cost. High, and easy to underrate. Most people do not type B4*C4; they type =, click, type *, click. Without it, the 90-function library is reachable only by someone who already knows A1 notation, and dash's own shortcut card has to teach = as a keystroke because there is no other door.

Format? No.

Files. dash/src/grid.ts (editCanvas — while an edit is open, a click on a cell inserts its reference instead of moving the cursor; a drag inserts a range). Size: two days, three with the referenced-range highlight.

2.12 No sort, no filter, no hide, no grouping

Today. canvasHeader (grid.ts:1854) emits a letter and a resize grip and nothing else — measured DOM confirms: no caret, no menu. Its comment says so plainly ("a spreadsheet's columns are not typed and cannot be"). There is no row/column hide and no outline grouping on either kind.

Cost. Moderate. Sorting a block of rows is a normal spreadsheet act and does not need column types — Excel sorts a selected range by a chosen column without knowing what anything is. The typed-column argument justifies leaving filter out; it does not justify leaving sort out.

Format? No — a range sort rewrites cells in place.

Files. dash/src/grid.ts, plus the context menu of §2.4. Size: two days for sort-a-range, and it should be a menu item, not a header caret.

2.13 No print, no PDF, no page setup

Today. Code-verified: grep '@media print\|@page' across dash/src/*.css returns nothing. Already in docs/dash-release.md §2 for the app as a whole. For a spreadsheet the specific need is repeated header rows, page breaks, and a print area — and the print area matters more here than on a dataset, because the sheet is unbounded and "print everything" has no meaning.

Format? A print area would be a small addition (CanvasSheet.printArea?: string), the rest is CSS and a page builder.

Size: a week, shared with the dataset kind.

2.14 A spreadsheet sheet cannot be renamed

Today. Tab menu ▸ Rename is disabled, with the reason shown: "Only a table sheet can be renamed in this build." applySheetProps narrows through table(doc, id) and throws otherwise. Duplicate, Move left/right and Delete all work.

Cost. Low in isolation, high in aggregate — every spreadsheet in a workbook is called "Spreadsheet". Cheap to fix and it is the last of the kind-discrimination bugs.

Files. dash/src/store.ts (applySheetProps), dash/src/tabs.ts. Size: an hour or two. Already logged in docs/dash-release.md §3.


3 · Takes more steps than it should, or reads as unfinished

  • The help card tells two lies on this kind. The "Good to know" panel says "Right-click a cell for rows, columns, fill and conditional formatting" and "a column header's caret to sort or filter". Neither exists on a spreadsheet. dash/src/help.ts, one hour.
  • The kind chooser truncates its own explanation. Measured: + ▸ menu is 220px wide, .dx-tab-menu-title and .dx-tab-menu-why are both inline on one line, and the "why" needs 322px — so the one place the two kinds are explained renders as SpreadsheetTyped cells — for a scr. The Save menu one control away gets this right with <span> + <small> on two lines. dash/src/tabs.css, ten minutes.
  • The fill handle's tooltip promises something it does not do. It reads "Drag to fill the selection down or across"; the canvas handler (wireCanvas) only ever extends rows, and there is no fill-right (⌘R is unbound). Either implement across or change the string.
  • #,##0 shows two decimals. readPattern (format.ts:49) only reads a decimal count from an explicit .0 group, so a mask asking for whole numbers falls through to the type default: measured 1234567.891 with #,##0 → 1,234,567.89. Half a day.
  • Negative currency renders £-1,234.50, where every spreadsheet writes -£1,234.50. Ten minutes.
  • A cross-sheet reference loses the source's format. =Pipeline!D2 shows 8200 where the source column shows £8,200.00. Arguably correct (a reference carries a value, not a format) — Excel agrees. Noting it because it will be reported as a bug.
  • TRUE displays as ✓. Consistent with the dataset kind's boolean column, surprising on a spreadsheet where the user typed the word. Worth a decision, not a fix.
  • = alone is stored as the text "=". Harmless, and a symptom of 2.11.

4 · Genuinely fine, and one thing dash does better

These do not need work and should not be padded onto a backlog.

  • The keyboard set is complete and good. Arrows, ⌘arrows to the edge of the block, PgUp/PgDn, Home/End, ⌘Home/⌘End, every shift-extension, ⌘Space / ⇧Space for column and row, ⌘A, F2, Enter/⇧Enter, Tab/⇧Tab with block-aware wrap, ⌘D and ⌘Enter, Del, Esc, ⌘C/X/V, ⌘Z/⌘Y, ⌘S, ⌘PgUp/PgDn between sheets. This is better than most web spreadsheets ship with.
  • Undo granularity is right. Measured: a cell edit is one entry (keystrokes within one cell coalesce via runEdit/endRun), a paste is one entry, a resize is one entry, and redo restores exactly. No complaints.
  • The status bar aggregate works on this kind. Selecting D4:D6 gives Sum 278 · Avg 92.67 · Count 3 · Cells 3, and it is capped so ⌘A on an unbounded sheet cannot freeze the tab.
  • Copy/paste of formulas is correct, including the hard part. A copied formula's references translate; a cut formula's do not. Excel's rule, easy to get backwards, got right. (The documented gap — formulas elsewhere that pointed at cut cells do not follow them — is real but rarer.)
  • fillSeries itself is good where it is reached: ISO dates step by day, months clamp correctly (31 Jan + 1 month = 28 Feb), Q1→Q2, trailing numbers increment, and a lone number sensibly repeats rather than counting. It is the caller that is broken (§1.1), not this.
  • Row height and column width drag live and commit once, keyed by letter and 1-based number so the JSON reads like an address. Stored sparsely.
  • The unbounded frontier is exactly right, and it is the thing that fixes the complaint in docs/dash-release.md §2: rows past the data are real, selectable and typeable, =SUM(D4:D6) below a block works, and the file stays sparse (measured: my whole invoice was a 2.7 KB document, and clearing never-written cells writes nothing).
  • The document validator passes on a spreadsheet with formulas, formats and styles.
  • Cross-sheet references work (=Pipeline!D2 → 8200), which docs/dash-sheet-kinds.md correctly called a requirement of the kind rather than a later nicety.
  • The function library is strong: ~90 functions including XLOOKUP, VLOOKUP, INDEX, MATCH, SUMIFS, COUNTIFS, MINIFS, IRR, NPV, PMT, PERCENTILE, CORREL. Verified live: VLOOKUP → 42.50, SUMIF → 80, COUNTIF → 1, TODAY() → 2026-08-14, =D8/0 → #DIV/0!.

Better than Excel: the tab menu's refusal pattern — a disabled Rename that says why it is disabled — is better than Excel's greyed-out silence, and it is the model the toolbar in §2.2 should copy verbatim. Also fillSeries's deliberate asymmetry (a lone number repeats, a lone date counts) is documented reasoning that Excel has but never explains.


Suggested order

  1. §1.1 fill down — silent, destructive, half a day.
  2. §1.3 / §1.4 export findings — an hour each to stop the silence, even before the real export exists.
  3. §2.2 toolbar — one day to disable-with-reasons and hide Formula. This is the complaint that prompted the review and it is the cheapest thing here.
  4. §2.1 cell formatting — appears to be in flight; check before starting.
  5. §2.4 insert/delete rows and columns + a context menu — three days, and the largest remaining usability win.
  6. §2.3 find and replace — one to two days.
  7. §2.11 point-and-click references — two days; it is what makes the formula engine reachable.
  8. §1.3 real xlsx export of a spreadsheet sheet — the first thing that makes the kind shareable.

Everything below that (dates, freeze, sort-a-range, conditional formats, print, merges, validation, named ranges) is a real backlog rather than a defect list, and the format decisions among them — merges, range-keyed conditional formats, validations, named ranges, a print area — are worth naming in docs/DECISIONS.md before any of them is built, so the field names are not chosen twice.


What I could not verify, stated as uncertain

  • §1.3 export is code-verified only: I clicked Data ▸ Export Excel and saw no warning, but I could not open the resulting file to confirm the sheet is absent. The filter at xlsx.ts:1314 is unambiguous; the absence of a finding is what I am inferring.
  • §2.13 print is code-verified only (no @media print rules exist); I did not trigger a print preview.
  • §1.1 almost certainly affects the dataset kind too, since fillDownSelection is shared. I did not measure it there.
  • §2.5 dates: I tested arithmetic and formatting, not the existing date functions' input tolerance, so the coercion work may be smaller than stated.
  • Everything in this document describes commit 19a82a4. Three other agents were editing this tree while it was written, and at least one of them is building the cell-properties panel. Re-check §2.1 and §1.2 against the tree before acting on them.