Show a tabular file as a picture, drawn the way the export writes it

[JackTelford]

Previews a spreadsheet in the answer instead of making the reader download it first. Each worksheet is drawn as a figure — header band, lettered column strip, numbered row gutter, and a caption saying what was left out of the window.

The picture is drawn from the formatting the file is written with

Resolving formatting in two places is how a preview ends up showing a different header colour, or a raw 74612.15999999999 where the file shows $74,612.16. The resolution both tools need now lives in TabularFormattingResolver, and ExportTabularDataTool delegates to it, so the two cannot drift. Column formats inherited from the uploaded workbook are applied to the drawing too, which is what makes a column that was currency in the upload read as currency in the preview.

A workbook stores a number and a format code and leaves the presenting to whatever opens it, so the codes are read in TabularPreviewValueFormat. Currency, accounting, plain number, percent and date are covered; anything outside that vocabulary falls back to the general presentation rather than being guessed at.

The general presentation now rounds to the significant digits a spreadsheet shows, instead of printing a double to full binary precision — 74612.15999999999 was wrong even with no format applied.

Two faults found while building it

Read-your-own-write. Figures were addressed before the rows backing them were committed, so every picture arrived at an address that did not resolve yet. The document is now committed before its address is handed out.

The model would not write the markers that place the pictures, through three rounds of prompt hardening. The host now appends a marker for any servable image the answer did not name — the rule this codebase already applies to generated files, which are surfaced even when the model does not cite them.

Reading the picture

A figure is bounded to a fraction of its drawn size so a stack of sheets does not push the conversation off screen, which leaves a grid at roughly half size. Clicking one enlarges it (~45% to ~72%, undistorted).

The download sits over the picture's corner rather than on a line of its own — that band cost a row of vertical space per sheet. It saves a PNG, converted in the browser from the picture already loaded: no second fetch, and nothing converted for a picture nobody saves. Every failure path falls back to the SVG the server holds.

Verification

• TabularPreviewValueFormat covered by 15 cases against the format codes a real workbook carries: $74,612.16, $0.00, -$81,065.04, (1,234.50), 12.3%, 09-01-26 — all pass.
• Zoom measured across three sheets at 1280x900: undistorted, 71-73%, opens and closes, thumbnail unchanged.
• Hover download: 0px added height, 51 ms conversion, real PNG magic bytes, canvas untainted.
• Generated .xlsx validated with OpenXmlValidator — 0 errors; bytes served match bytes on disk by SHA-256.

The second commit fixes a check that located the figure wrapper by the first mention of its class name in the file. That is the markup only while no rule or selector is written above it, so adding a stylesheet for the figures made it read a CSS selector instead. It now searches for the opening tag and checks every figure a file emits.

Known gaps

• The SQL-query preview path gets the correct header colour and the rounding fix, but not per-column formats — attributing a query to a source table needs the export's sql-matching logic.
• ai-chat.js and chat-interaction.js render the same figure markup for the other chat surfaces and still need the zoom and PNG-download treatment, which requires an asset rebuild.

🤖 Generated with Claude Code