Skip to content

armgit / macro

Public Imported from GitHub
0

Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.

209 KB
Raw Blame

Documents

For DSS sync rollout checks, use PostHog flag sync-via-gateway (dev and production) and recreate document sessions after the client receives the updated flag. A reload can initially read PostHog's cached value; open sessions do not change transport when the remote value changes. Off or unknown flag values use a direct sync-service WebSocket per document; on multiplexes document frames over a dedicated DSS /sync/ws socket. Confirm all DSS replicas have the sync endpoint before enabling; notification traffic still uses connection gateway. Verify two clients converge after concurrent edits, a large update, and a forced sync disconnect with pending edits. The historical flag name is unchanged. Local stacks can override with VITE_SYNC_VIA_GATEWAY=true or false.

The document header has separate Share, Copy Share Link, and Side Panel buttons. They are borderless with a soft rounded background on hover. Share opens the sharing dialog; copying a link is a separate action.

In the Share dialog's Optional message, paste an image with Cmd/Ctrl+V or drop a file into the composer. Share composers have no paperclip button on desktop or mobile. Images show a removable preview and are sent inline beside the shared item to each selected person or channel. Share (including Cmd/Ctrl+Enter and the phone drawer's Share button) waits for uploads; Enter still inserts a newline. Removing an uploading image keeps it removed, and closing the dialog discards its attachment draft. Verify paste, drop, removal during upload, upload failure, and sending to both a person and a channel. On touch, attachment previews scroll horizontally and removal remains available without hover. Composer fixture screenshots: desktop and phone.

Verify Public → View links in a fresh signed-out browser context. The /app/drive/md/<id> route (including folder/tab document links and legacy /app/md/<id> redirects) must show the document without private Drive lists or folder navigation, on desktop and touch devices. The login banner may remain; it must not replace the public document. Signed-out readers see Sign in to view comments, not a comment-fetch error or composer. While identity is unresolved, the discussion shows neither comments nor a sign-in prompt; throttle the preview auth response to check this on desktop and touch devices. Account-only lists, contacts, notifications, activity, and gateway/storage WebSockets must stay idle until sign-in; public document metadata, properties, and sync still load normally. Check that signing in restores account services and signing out stops them. Private files still prompt for sign-in. Signed-out Drive list URLs (My Files, Recent, Shared with me, or folders without a document target) show Sign in to view your files with Login and Sign Up actions, never an indefinite loader.

Opening a private file link while signed in shows You don't have access to this file, the signed-in email, and Request access. View is selected by default; choose View, Comment, or Edit where supported, submit, then verify Request sent and Cancel request. On touch devices, document links open the full-page document route; the same request controls appear there, including at phone widths. The document stays unavailable until its owner approves. The open denied view reloads when it receives the grant update. Signed-out visitors are prompted to sign in first; ordinary load failures retain Try again. An already-open document reloads after a permission upgrade on desktop and touch devices. Removing its last grant closes the editor and returns to Request access; a failed background refresh must not restore the revoked document. The denied view shows Loading access requests… until its own request-state read lands, never a request row or Could not load access requests left over from before the change, and a document deleted while it is open shows the not-found view.

In People with access, a non-owner can use Request access beneath the owner to request a supported upgrade (Comment or Edit). PDFs offer Comment; view-only email sharing has no upgrade. Requests can be sent or changed once per minute; cancellation stays available immediately. Approval allows a further upgrade immediately. After a denial, the requester must wait 24 hours to request again. This workflow also applies to the existing chat, folder, project, database, call, and agent-session Share menus, using the access levels that each supports.

Reviewers see a pending count on Share. Entering an item with pending requests opens Share once for that visit; requests arriving during the visit update the badge. On mobile, review requests in the People tab. Permission requests lets a reviewer choose an access level and Grant access, or Deny. Team owners/admins and the human responsible for a bot-owned item review on its behalf. Only the first review of a request succeeds. Grants notify the requester; denials do not. Approved users appear in People with access, with controls to change or remove their direct grant. A pending request also puts the item, projects included, on the reviewer's Home feed until it is resolved. A project row has no Home preview: clicking it opens the project in place, Shift-click opens it in a new split, and when the project is already open in another split that split is focused. Opening a permission-request notification navigates to its item and opens Share.

Both Edit with AI and selected-text AI edits open the shared usage-limit dialog when the backend refuses a request with a usage-limit code. Free users can choose View plans; paid users can choose Open usage settings to add credits. The failed edit leaves the document unchanged and clears its running state. Other edit failures still show AI edit failed. The dialog also opens for usage-limit failures from live document-editing tools in chat.

Snippet owners manage team sharing under Share → Team access. The details panel has no separate Sharing section. Choose Edit to grant the access the old snippet toggle provided, or None to remove team access.

Hover or focus a document title in the header to see its owner, created time, and last-updated time. This details card is shared by documents, tasks, snippets, canvas, and file blocks; unavailable metadata is labeled rather than inferred. Canvas and file breadcrumb names follow live renames, including the tab title and title menu. Rename an open canvas and confirm its header updates immediately without reloading or remounting the editor.

Editable documents show matching Add tags and Add property pills below the title. Once tags are applied, Add tags becomes the existing tag name or count pill; clicking it reopens the picker. Add property opens the shared property selector for that document. Adding or reselecting a property from the title row pins it there. Adding from the side panel leaves it unpinned. The side panel shows all assigned properties, including unpinned ones. In document, task, snippet, and skill side panels, hover or focus a property row to find Pin below title beside Remove from entity. These controls stay visible on touch devices. A pinned row shows Unpin from title; unpinning keeps both the saved value and the side-panel row. Pinning is unavailable for read-only documents, metadata, and properties still being assigned, except the task's default Project placeholder. Long values truncate or wrap within the value column without covering the pin or remove controls. Escape cancels an inline text or number draft without saving it. Hover or focus an inline property pill to find a filled, icon-only Unpin button separated by a gap (keeps its saved value). In a single-row hover card, it sits to the right of the value. In a column hover card, it sits to the right of the key, above a full-width divider; values keep the full width below. Long keys wrap without displacing the Unpin button. Delete from item stays below, where removal is allowed (removes the assignment without deleting the shared property definition). Task builtins cannot be deleted from these pills. Pins are saved with the document, including an explicitly empty pin list. Tasks initially show Status, Priority, Assignees, Due Date, and Project when enabled. Defaults apply only without a saved pin list; customized pins stay unchanged. Changing pins keeps the other defaults unless explicitly unpinned.

Markdown code blocks have a Copy Code button in both editable and read-only views. Successful copies briefly animate the icon to a solid green check-circle; they do not show a success toast.

Embedded code files, canvases, and PDFs in Markdown documents show their content below the card header, including when the document opens in a drive detail view. Project cards remain header-only. To verify, reopen a document with existing embeds and check that their preview bodies load; also check that a code or canvas mention offers Convert to Embed in its preview menu. Drag an embed's bottom-right resize grip, then release and select the card. Its height should stay at the dragged size and persist after reopening the document. Zoom to fit and menu/keyboard zoom use the full canvas viewport center; floating inspectors and details overlays do not reserve centering space. Verify with content on both sides of the canvas and the inspector visible. Press . to fit the current selection, or all content when nothing is selected. The zoom menu keeps Zoom to fit for all content and adds Zoom to fit selection, disabled with no selection. Verify one item and multiple selected items, then deselect and check that . fits everything again. Canvas Ask Macro, including the Drive detail header, attaches the open canvas to the existing agent composer on desktop and mobile. Ask it to create or edit a diagram: ReadCanvas reads the current live scene and EditCanvas applies bounded commands through the same collaborative room. Expand the tool result to inspect its outcome; edits appear in the canvas while the agent works. A viewer can read but cannot edit. New media/document references and removal of embedded assets remain UI actions. If an edit is cancelled or loses its acknowledgement, inspect the canvas before repeating the request because earlier snippets may already have applied.

To verify AI editing, open the same canvas as an owner, a second editor and a viewer. Use Ask Macro to create two labeled boxes and a bound arrow. Confirm all three see the changes, the viewer's edit request is refused, and a second editor's unrelated shape/style edits survive. Move an endpoint shape and check the connector follows it. Reload both editing sessions and inspect the tool's expandable result. At phone width, tap the document title to open its file menu for Ask Macro; the floating zoom/inspector sits below that header. Use the responsive action and verify the composer attachment and result disclosure remain usable.

All canvases open in Canvas Next, including drive details and Markdown embeds; there is no rollout flag or legacy editor. Unversioned/version 1 files migrate in memory. Opening a file alone does not save the migration; the first edit saves version 2 while retaining the original JSON for recovery. Unsupported legacy records and unknown versions show an error without changing the saved file. To verify migration, open a supported legacy canvas, edit it, then reopen it and check that its shapes, text, and connectors remain intact.

Canvas spatial paste keeps the layout of the copied selection. Copying from another canvas, or pasting Excalidraw data, centers the selection in the current viewport at the current pan/zoom. Same-canvas copies retain the cumulative 24-unit paste offset, including copies from another tab of that document. Shift+Cmd/Ctrl+V (or Paste in place in the context menu) keeps original world coordinates and bypasses grid snapping. Verify cross-canvas, same-canvas repeated paste, and Shift-paste with a grouped selection; ordinary text/editor paste is unchanged.

The editable drawing toolbar has a paperclip Attach button with an I shortcut hint. Clicking it opens a standard dropdown for Media, Upload media, and Reference. Its pickers open above the toolbar; attachments land at the canvas center. Verify clicking Attach, choosing each type, and Escape. The Attach tooltip stays hidden while the insertion menu or picker is open. Radial menu buttons retain an opaque surface when their slice is highlighted.

With an editable canvas focused, press I to open a radial menu of separate floating buttons around the mouse position: Media, Upload media, and Reference. Equal directional slices highlight the corresponding button and ring arc on pointer movement; clicking anywhere in the viewport selects the active slice, including well beyond the buttons; outside clicks do not dismiss the radial menu. Escape closes it. The center is a neutral dead zone. Media and Reference replace the radial menu with searchable mentions-style pickers. Upload media opens the OS picker directly for local images/videos. References insert document preview cards; creating new embeds is unavailable. Arrow keys or Ctrl+J/K navigate; Enter selects and Escape closes. Selected rows scroll with the same inset as the menu padding. Picker height stays fixed through loading, search results, and no results, capped by the available viewport space. Verify typing an unmatched query and clearing it does not move the search field or resize the menu. Media retains its dashed drop area and Add files button, with no folder option. Insertion uses the position captured when I is pressed. Verify keyboard selection, closing/reopening, upload cancellation, and placement after pan/zoom. The shortcut is inactive during text/embed editing and in read-only canvases. Reference cards register with the shared reference system when the canvas saves. In both editable and read-only canvases, click a reference card's title to open its document; selecting the rest of the card leaves the canvas in place. Verify the backlink on the target, duplicate cards retaining one reference, and removing the final card followed by undo/redo. Reload must preserve tracking. Existing untracked cards register on their next content save. Pasting a same-host Macro file URL uses the same link resolver as Markdown and inserts a reference card at the viewport center, including nested Drive routes and copied split layouts (the rightmost pane). Legacy /app/md/<id> links still work. Verify a /app/drive/md/<id> paste becomes a card and registers a backlink; foreign-host links and unrecognized routes remain text.

Pasting or dropping image files shows local previews while the static-file upload finishes. Previews become selectable image nodes after upload and remote loading succeed; a failed upload removes its preview and reports the failure. Verify with a slow upload, then undo/redo the inserted batch and reopen the canvas. Blob URLs must never enter the saved scene. When opening a canvas, image nodes show a skeleton-shimmer bg-skeleton placeholder at their saved size until the image loads; this also applies to read-only embeds. Failed image requests display “Media unavailable” instead of an indefinite skeleton.

Canvas object snapping is enabled by default. Drag a shape or selection near another object's edge or center to align it; dashed guides show the matched axes. Axis-aligned resize handles also snap. The tolerance stays six screen pixels at any zoom, and objects more than 160 screen pixels away on the perpendicular axis are ignored. In Zoom options, toggle Snap to objects independently of No grid snapping, Snap to px, or Auto snapping. Object alignment takes priority over grid snapping on the matched axis. Verify dragging with snapping on/off, resizing, Shift-constrained movement, zoom changes, and undo. Rotated resize frames retain their existing local-axis resizing behavior.

Canvas embeds are read-only: Select selects shapes without moving them, and Hand pans the canvas. The toolbar contains only those two tools; property inspector details and drawing/editing controls are absent. Check toolbar clicks, drag-to-pan, and the V/H shortcuts with focus inside the canvas. These interactions must keep focus in the embed and must not type into or modify the outer document. Embeds hide the dot grid. The bottom-left zoom toolbar has minus and plus buttons around the current zoom percentage. Click the percentage for the shared canvas zoom menu: Zoom In/Out, Zoom to fit, and 25%, 50%, 100%, or 200% presets.

Create a document from an editor

Type /document in a Markdown slash menu and choose Document. The popover has a New document title, a Markdown body, and a Tags pill. It starts without task property pills. Tags can also be applied with # in the title or body. Tags sit on the left of the bottom row beside Create Document. Enter in the title moves into the body. Opening mentions keeps the composer at the same height.

Choose Create Document or press Cmd/Ctrl+Enter. The composer closes and the source editor shows a creation placeholder at the insertion point, then replaces it with the new document mention. If creation fails, the placeholder is removed and the composer reopens with the title, body, and tags for retry. Continue editing in split creates the document with the current draft and tags, opens it in an adjacent split, and inserts its mention in the source editor. An empty title uses New Note when expanding. Create → Document continues to open a document directly.

Create a task from a Markdown selection

Select text in a Markdown document and choose Task in the selection toolbar (desktop) or touch toolbar. The task composer opens with the selected text as the title (truncated to 70 characters) and a Snippet mention of the selection as the body; opening that mention scrolls to and highlights the range. Editors also see a Link in document toggle beside Create More; it is remembered and starts off. While on, creating the task inserts its mention right after the selected text, inside the same checklist item when the selection is in one; a selection ending between list items links after the preceding item's text and never adds list items. Verify both toggle states and that other text edited while the composer was open is preserved.

When an editor's selection spans two or more checklist items, the button reads N tasks and creates one task per item without opening the composer; each item is replaced by its task mention and a Created N tasks toast appears.

Images in lists

An image inserted or dropped while the caret is in a list item goes under that item's text, left-aligned with it, and stays resizable; the list is not split and numbered items keep counting. On the empty item Enter leaves after an item, the image goes under the item above and the caret stays on the empty item; pressing Enter again leaves the list, and an image inserted then sits below it. Enter at the end of an item's text starts the next item below its images. Dropped files target the item under the pointer's lower half, or the item above. Markdown saves the image as an indented line under its item. Verify with list-images.browser.e2e.ts in apps/web/src/lib/core/component/LexicalMarkdown/component/core/browser-test.

Markdown outline

On desktop, Markdown documents with at least three headings show a tick rail in the left margin. The rail appears only where the text column's left margin is wide enough to hold it clear of the text, so it never overlaps the checkbox markers of a checklist; a document whose margin comments have pushed the text column against the container padding shows no rail at any window width. Every section whose content overlaps the editor viewport is highlighted, including a section whose heading has already scrolled above it. Hover a tick (or Tab to its button) to expand it and nearby ticks and show a rounded preview with the section heading and up to three lines of body text. While a preview is open, only its tick is emphasized; visible-section highlights return when the preview closes. The preview overlays the text column, but it does not take the pointer: a click that lands on it reaches the document underneath, so checkboxes it covers still toggle, and moving off the rail closes it even if the pointer crosses the card. Click a tick or press Enter to jump immediately to its heading without closing its preview or collapsing the expanded ticks. Scrolling, resizing, and editor updates refresh the visible-section highlights.

On a local HTTPS stack, document and image downloads use /local-storage/ on the app's HTTPS origin. A request to HTTP localhost indicates a stale storage URL or stack configuration; hard-refresh after updating the stack. Markdown also needs a successful /sync/document/.../connect WebSocket upgrade. A 403 there indicates the sync origin check, which the local proxy handles for HTTPS machine hostnames; verify the proxy configuration before retrying.

Databases in documents

Opening the slash menu with / lists Normal Text, headings, and the other Markdown options before Query and Database. Typing /database selects Database: it creates an Untitled database with its first table and Name column, then inserts its live editor at the command's position. Rename the database in the embedded header. A failed create clears its pending placeholder; removing the placeholder while creation runs does not insert elsewhere.

Selecting a database from @ inserts a compact mention. Use its Expand database icon to open the inline editor. Databases created with /database start expanded. Edit cells, add rows or columns, select a table with the Database table picker, choose a saved view with Database view when the table has views, filter, and sort in the document using the same toolbar controls as the full database app. The Filter and Sort icons open their panels. The pickers support keyboard navigation and mark the selected table. Rows added or edited elsewhere, including by AI, appear through the database's gateway subscription. The editor catches up after reconnecting or returning to the tab after a long absence. Edits write to the original database. Open opens the full database; opening a row (its row-number button, a board card or a timeline bar) opens the full database at that row. Collapse converts the editor to a small mention; Expand brings it back. Removing either reference leaves the database and its rows intact.

Document permission controls changing the reference. Database permission controls editing its rows, columns and name, independently of document permission. An expanded editor in a read-only document can still edit a writable database; its collapse/expand is local to that reader. A database viewer sees read-only records. Sharing the document does not itself grant database access. Filters, sort order and widths are local to this visit; the selected table and view are saved in the Lexical node when the document is editable and sync to other readers.

Live database answers

With Databases on, type /query and choose Query to insert a live answer to a question about a database. Answers run with each reader's database access and refresh when referenced tables change. See Databases for the question box, source picker, displays, and editing.

Spreadsheets

Spreadsheets are an internal pilot controlled by the enable-spreadsheets PostHog flag in every environment. Team targeting is configured in PostHog; ordinary document permissions continue to control access to each workbook.

Choose Create → Spreadsheet, or New → Spreadsheet in Files or a folder. From Files the new workbook opens inline in the open folder or tab (/app/drive/.../spreadsheet/<uuid>) without reloading Drive. /app/spreadsheet/<uuid> links redirect into Drive on desktop. Native workbooks use the normal document title bar with Ask Macro and Share at the top right. The File actions ellipsis beside the title uses the same menu as documents, including rename, favorite, move, copy, and permission-appropriate file actions. Native spreadsheets use a green grid icon in file lists and search. The grid fills the panel beneath the formatting and formula bars. While a workbook opens, a shimmering placeholder grid (status "Opening spreadsheet…") stands in for it, and formula cells show a short shimmer bar until their first results arrive; wait for real cell text before reading values. Once it loads, the grid takes keyboard focus with A1 selected, so typing immediately edits A1 and arrow keys move the selection; it does not take focus from a dialog or control reached while loading, nor in an inline preview or when the sheet was opened by J/K list navigation. They have the .spreadsheet file type; uploading an Excel (.xlsx, .xlsm) or CSV file in Files or a channel opens a read-only spreadsheet preview, including existing /app/unknown/<uuid> links and CSV code routes. Review Import notes, then choose Edit in Macro to create a collaborative native copy. The original file and its link remain intact. Conversion waits for a durable save before opening the copy; a failed save can be retried without creating another copy. The normal document download action retrieves the original; the spreadsheet footer exports the imported representation.

You can also open a native spreadsheet and use the bottom-right Import and export → Import… menu to import its sheets. A large workbook shows a progress bar on Import workbook while it is written; the page stays responsive, and one undo removes the whole import.

Imported Excel workbooks keep conditional formatting, data validation and notes. Conditional formats recolor cells and draw data bars and icons, and they update as values change.

To add an in-cell dropdown, select the cells and choose Format and data → Dropdown… in the toolbar (sliders icon), or Dropdown… in the cell context menu. The Dropdown dialog takes Type the choices (one per row, Choice 1, Choice 2, …: Enter on a filled choice adds the next one and Enter on an empty one saves, Add choice appends a row, and pasting several lines adds a row for each) or Use the values of a range (Range of choices, such as A2:A10 or 'Sheet 2'!A2:A10, whose current values are listed below it), and Reject values that are not a choice (on by default). Each choice has a color button (Color of Low) offering No color and nine light colors. New choices take the next unused color; a dropdown saved without colors stays plain until one of its choices is colored. Cells showing a colored choice draw it as a chip, and the open list shows the same chips. Save replaces any validation on the selected cells; Remove clears it from them. Reopening the dialog on a dropdown cell shows its current choices and colors. Agents add the same dropdowns with EditSpreadsheet's set_dropdown operation (choice colors in colors: [{ item, color }]), see them in ReadSpreadsheet, and remove them with clear_validation. Excel downloads write choice colors as conditional formatting fills; importing such a file restores them as choice colors.

Every cell with a dropdown (from the dialog, an agent, or an Excel list rule) shows a small arrow at its right edge; clicking it selects the cell and opens its choices. The selected cell's arrow sits just outside its right edge (Choose a value for B2); click it or press Alt+Down. The list opens below the cell, or above or beside it when there is no room, without covering the cell or dimming the sheet. Up/Down and Enter pick a choice and Escape closes it. While a list is open, one click or tap on another dropdown cell opens that cell's list; any other cell closes it. Typing a value the rule does not allow shows the rule's message in the footer, and a "stop" rule keeps the previous value. A red corner marks a cell with a note; selecting the cell shows the note, and any input message, beside it.

To make a table, select a cell in the records (or the records with their header row) and choose Format and data → Convert to table (or Convert to table in the cell menu, or ⌘/Ctrl+Alt+T). It converts at once, with no dialog: one cell converts the block of filled cells around it, an empty cell becomes a blank table headed Column1–Column4, and the table is named Table1, Table2…. Blank or repeated headers are renamed (Column3, Amount2) and columns take the type their values suggest. While the active cell or the pointer is in a table, its name shows as a tab above the header row (as an icon in the column header bar for a table in row 1), a button named Table Table1 menu whose menu has Rename table and Adjust table range… (dialogs with Table name or Range and Save), a Table color radio group of ten swatches, Formatting checkboxes (Alternating colors, Banded columns, Bold first column, Bold last column, Footer row with totals), Group by view…, Clear all filters when filtered, Convert to range (cells keep the table's look) and Delete table (removes its cells too). In a table cell, Table options in the cell menu or Format and data opens the same menu.

Each header cell has a button (Region column menu; a filtered column's ends in , filtered and shows a funnel) whose menu shows the column's type, a Column type submenu (None, Text, Number, Percent, Currency, Date, Date time, Time, Checkbox, Dropdown, People), Sort A → Z, Sort Z → A, Filter…, Clear filter when filtered, Group by this column, and Insert column left, Insert column right and Delete column. A type formats the column's data cells; checkbox makes blanks unchecked boxes and dropdown offers the column's values as colored chips. A cell whose value does not fit shows an orange corner and a tooltip such as "Expected a number". Filter… opens Filter "Region": a Condition select (Text contains, Greater than, Is between…) with Value (and To) fields, and By values with Search values, Select all, Clear and a checkbox per value; OK applies it and hides whole sheet rows. Group by view… opens Group by view · Table1 with Group by, Summarize and As selects and a table of groups with their row counts and summary; choosing a group filters the table to it and Show all rows clears that filter.

With Footer row with totals on, selecting a footer cell shows its arrow (Choose a value for B5) listing None, Sum, Average, Min, Max, Count, Count numbers, Empty, Unique, Percent filled, Percent empty, Percent unique, Standard deviation and Variance; the cell shows the aggregate's name beside its value. Typing a value directly below the table (without a footer) or directly to its right grows the table; a new row takes the formatting, dropdowns and calculated formulas of the row above. Formulas use structured references such as =SUM(Table1[Amount]), =[@Qty]*[@Price] and =ROWS(Table1); typing Table1[ suggests its columns. Renaming a header rewrites formulas that use the old name, and references to a missing table or column show #REF!. Scrolling past a table's header row shows its column names in the column headers. Agents create tables, with optional columnTypes, through EditSpreadsheet's create_table operation, and ReadSpreadsheet lists each sheet's tables with their columns and types. Excel tables with a header row import with their names, styles, totals and value or condition filters and download as Excel tables; a table without a header row imports as ordinary cells, with its references converted to cell ranges.

To add checkboxes, select the cells and choose Checkbox in the toolbar's Insert chart, image or checkbox menu (chart icon), or Insert checkbox in the cell context menu. Blank cells become unchecked boxes backed by FALSE; cells holding TRUE or FALSE keep their value, and other values stay visible as text. Each box is a checkbox named after its cell (B2) and sits centered in the cell unless the cell is aligned left or right; it takes the cell's text color and font size. Click a box, or press Space, to toggle it; Space over a range checks every box, or unchecks them all once all are checked. Shift+click on a box extends the selection without toggling it. Delete unchecks any checked boxes and, pressed again, removes them; Remove checkbox in the cell menu (or Clear formatting) removes them and keeps their values. Formulas read the values, so =COUNTIF(C:C,TRUE) counts checked boxes. Agents add them with EditSpreadsheet's insert_checkboxes operation and toggle them by setting TRUE or FALSE. Downloads use Excel's checkbox format, and imported Excel checkboxes stay checkboxes.

Double-clicking a cell that holds a date (typed, imported or a date pill, but not a formula), or pressing Enter or F2 on it, opens a calendar below its editor; an Excel date validation rule opens it on its cells even when they are empty, starting at today. The editor keeps focus: arrows move the outlined day by a day or a week, Page Up and Page Down by a month (with Shift, a year), Enter writes the day and moves down, and Escape closes the calendar (a second Escape cancels the edit). Clicking a day writes it and keeps the cell selected. Typing closes the calendar and edits the text as usual. The day is written the way the cell stores dates, so its format stays: a serial number keeps its time of day, a date pill stays a pill and ISO text stays ISO. The calendar is [data-cell-date-picker]; its days are gridcell buttons marked data-active (the keyboard's day), data-selected (the cell's date) and data-today.

Images and charts (column, bar, line and area, each clustered, stacked or 100% stacked, plus pie, pie of pie, bar of pie, doughnut, scatter, combo, radar, bubble, stock, surface drawn as a contour, and Excel 2016's histogram, pareto, waterfall, funnel, box and whisker, treemap and sunburst) are drawn over their cells and move with them; charts redraw as the cells they read change. A 100% stacked chart fills each category to 100% of its positive values and to 100% of its negative values. Scatter charts plot markers without connecting lines. A combo chart draws every series but the last as columns and the last series as a line on a second axis. Each is a figure named after the chart title ("Chart: Revenue") or the image description. The toolbar's Insert chart, image or checkbox menu (chart icon) adds a chart of the selected cells — or, from one cell, of the table around it, placed beside it — or an image file (PNG, JPEG, GIF, WebP or BMP up to 2 MB) at the active cell. Its More charts submenu holds Radar, Filled radar, Bubble, Stock, Contour, Histogram, Waterfall, Funnel, Box and whisker, Treemap and Sunburst. Edit chart's type list also offers 100% stacked bar, stacked line, 100% stacked line, 100% stacked area, Combo, Pareto, Pie of pie and Bar of pie. A histogram bins one column of numbers (Scott's rule, as Excel's automatic bins) or, with labels, sums values by label; a treemap or sunburst reads every column before its values as a level, a blank label continuing the one above. Excel 2016 charts download as chartEx parts with Excel's fallback note for older versions; Excel 2016 maps import as that fallback. A stock chart needs three or four series (high, low, close, with open first for up-down bars), a bubble chart reads x values from the first column, then values and sizes in pairs, and a combo chart needs at least two series. Imported shapes, text boxes, lines, groups and SmartArt are drawn too, as figures named after their text (or their name, such as "Straight Connector 2"); a shape linked to a cell shows the cell's value. They move, size and delete like images, and download as Excel wrote them (SmartArt as a group of its shapes). EMF and WMF pictures show a picture drawn of them and download as the original metafile. Click a drawing to select it: drag it to move it, drag a handle to size it, press Delete to remove it (undo restores it) or Escape to return to the cells. Double-click a chart, press Enter, or use its pencil button to open Edit chart (type, title, legend, the cells it charts, series in rows or columns). From the keyboard, Ctrl+Alt+5 (also in the Insert menu) selects the first drawing, Tab and Shift+Tab move between drawings, arrow keys move the selected one (Shift sizes it, Alt by one pixel). Pivot tables show their last values as ordinary cells; the Excel download keeps them, and Excel rebuilds them from their data when the file opens; pivot tables over other workbooks or data connections download with the data Excel saved with them (connections without saved passwords), for Excel to refresh from their source. GETPIVOTDATA formulas that read a kept pivot table calculate from its cells, so editing a value in the table updates them. Deleting a sheet whose data a chart on another sheet reads keeps the chart: it shows the values it had, which no longer change.

Drag across cells to select a range. Holding the pointer near or past an edge of the grid (including over the row numbers or column letters) scrolls it, faster the farther out it is, so one drag reaches off-screen cells; the fill handle, header drags, formula reference picking, and phone selection handles scroll the same way.

Select a cell to inspect its address and input in the formula bar. Double-click a cell, start typing, or use the formula bar to edit its value. Formulas begin with = and may refer to cells or ranges, for example =SUM(B2:B5). Check both the rendered result and formula bar: a formula's displayed result differs from its stored input. Paste a rectangular selection copied from another spreadsheet to populate multiple cells. The toolbar offers number formatting, bold, undo, and redo; sharing uses the same document permissions as other Macro files.

On a phone, tap once to select and tap the same cell again promptly to edit. Swiping the grid scrolls without extending selection. Drag Move selection start or Move selection end to select a range. The formula bar exposes Apply edit and Cancel edit while editing, so a software keyboard is sufficient. Swipe the formatting ribbon horizontally to reach more controls. On narrow screens, Add rows is in the active sheet's actions menu; Import and export stays at the bottom right.

Both editors offer formula autocomplete. Type a function prefix such as =SU (a bare = shows nothing), use Up/Down to choose a suggestion, and Tab or Enter to insert it. Clicking a suggestion also keeps focus in the editor. The compact popup lists each function with a one-line description and the highlighted function's signature; after ( it shows the signature with the current argument highlighted, including inside nested formulas. Help opens only after typing: focusing or clicking into an existing formula, switching sheet tabs, or picking a reference leaves it closed. Pressing outside the editor closes it, and Escape dismisses help first, then cancels editing on a second press; either keeps help closed until the next keystroke. Suggestions do not appear inside quoted text or in view-only mode.

While editing a formula, click a cell or drag across cells to insert a reference at the caret (for example, type =SUM(, then drag B4 through B7). The draft updates to =SUM(B4:B7 without committing or moving the active cell. A dashed outline shows the referenced range. Release, type ), and press Enter to calculate. This works in both the cell editor and formula bar, including reverse drags, replacement of an existing reference, and subsequent arguments after a comma or operator. To reference another sheet, click its tab while the formula is awaiting a reference (for example after =IF(D2 <), then click or drag the source cells. The other sheet opens at its own selection, even when the formula's row is past that sheet's last row. The draft stays in the formula bar; Enter commits it to the original sheet and cell. Names with spaces are quoted automatically. Escape cancels and returns to the original sheet. While a formula is being edited, each cell or range it refers to, including whole columns or rows such as D:D, is outlined in its own color on the grid, and the reference text in both editors uses the same color. A repeated reference keeps its color. Only references to the sheet being shown are outlined; the colors disappear when the edit is committed or cancelled. On touch screens, tap a cell while editing a formula, then drag Move reference start or Move reference end to extend its reference. Tapping a suggestion or adjusting a reference should keep the input focused and the software keyboard open.

Drag across cells, Shift-click, or use Shift + arrow keys to select a range. Drag across row/column headers or Shift-click a second header to select multiple whole rows/columns. Arrow keys then move from the selection's active cell. The focused grid owns typing and navigation; app navigation shortcuts do not run while it has focus. The active cell keeps a complete border while editing; a range has a translucent tint over its cells' own fills and an outer border. The active cell is never tinted: selected or being edited, it keeps its fill, font, text color, and explicit alignment (automatic alignment edits from the left, as in Excel). Verify selection in both drag directions and after scrolling, including near the last row and column. Only nearby rows are mounted. Use Go to cell or keyboard navigation to reach off-screen cells; verify the target is visible below the sticky column header.

Copy within Macro and paste elsewhere to translate relative references: copying =B4-C4 down becomes =B5-C5, while $B$4 stays fixed. Drag the small handle at the selection's bottom-right corner to fill down/up or right/left. Select a range and use Format and data → Fill down / Fill right or Cmd/Ctrl+D / Cmd/Ctrl+R. Drag fill continues arithmetic number sequences and daily, weekly, monthly, or quarterly date sequences (including month ends). Text and irregular patterns repeat; relative formula references translate. Keyboard/menu Fill down/right explicitly copies the starting row/column. Plain-text paste from other apps keeps formulas as supplied.

With a mouse, hover the selection's border until the cursor becomes a move cursor, then drag it to move the cells; a dashed outline and its address (for example D6:E9) show where they will land, and the grid scrolls when the pointer nears or passes its edge. Escape cancels the drag. Values, formatting, notes and validation go with the cells, the moved cells stay selected, and formulas that refer to them follow: =SUM(B2:B3) becomes =SUM(D6:D7) after moving B2:B3 to D6. Moving onto cells that hold values asks Replace cells? first (Cancel is focused); formulas that referred to the replaced cells show #REF!. One undo reverses the move. Like insert/delete, moving is unavailable while a saved workbook is offline, and a collaborator's edit made meanwhile cancels it with a footer notice. Moving part of a merged cell is refused. Touch screens keep the selection handles instead.

Drag a column header's right boundary or row header's bottom boundary to resize; double-click the boundary to auto-fit. Row separators support Up/Down arrows and Enter to restore automatic height. Explicit row heights take precedence over wrap. At 100% zoom, default columns are 100 pixels wide and rows are 21 pixels high; larger text and wrapping expand the row. Saved custom column widths take precedence. The resize separator also supports Left/Right arrows and Enter for auto-fit. Use Add rows in the footer to append 100 rows (up to 1,000 total). Resizing and row additions save collaboratively and can be undone.

The document title and Share button sit above a compact formatting ribbon: undo/redo, paste, zoom and view options, currency/percent/decimals/number format, font and size, text styles, text/fill color, borders, alignment, Decrease indent/Increase indent, wrapping, functions, format/data actions, and find. Icon controls expose accessible button names and tooltips. The indent buttons move each selected cell one level (three spaces) from its own level, from the cell's right edge when it is right-aligned and its left edge otherwise; indenting a centered cell aligns it left. Number format → Accounting shows Excel's accounting layout: $ at the cell's left edge, the amount at the right, negatives in parentheses and zero as a dash; the cell's accessible name reads like B2: $(1,234.50). Paste special offers Paste and Paste values only. Select a range first; formatting applies to all selected cells. Toggling bold, italic, underline, or strikethrough on a mixed selection first enables it for the entire selection. Whole-column formatting preserves the viewport. In a cell that is already percentage-formatted, typing 5 means 5%; formulas and AI/API numeric values still use fractional values (0.05 for 5%). Font sizes are points. Wrapped rows grow automatically up to 160 pixels at 100% zoom. Check selection and formula-reference outlines after changing wrapping, font size, or zoom. Excel black text and borders on unfilled cells follow the app's foreground color so imported sheets remain readable in dark mode. Explicit text/fill color pairs remain unchanged; theme changes never alter saved or exported workbook colors.

View options directly toggles gridlines, the formula bar, and formula display. Gridlines are saved with the active sheet, undo like an edit, export as Excel's showGridLines, and are disabled for viewers; the formula bar, formula display, and zoom are local to the editor. Go to cell accepts ranges such as A3:E7. Escape from the address or font-size field returns keyboard navigation to the grid. Functions starts an editable formula: for a selected range it proposes the result in the empty cell below; for one cell it opens =FUNCTION( for reference picking. An occupied result cell is never overwritten.

The Find and replace ribbon button (Cmd/Ctrl+F in the grid) opens a compact find bar in the top-right of the sheet (same pattern as channel find): query, match count, previous/next, and close. The sheet stays visible while searching. Cmd/Ctrl+F is captured on the editor (including the formula bar and find bar) so the browser find dialog never opens; a second press re-selects the query. Press the replace toggle to expand replace-with, Replace / All, and match options (case, entire cell, formulas). Next/previous select each result. Formula results can be searched, but replacements preserve the formulas unless Formulas is checked. Format and data sorts the selected rectangle by its active column, keeping each row's values, styles, and relative formulas together, or trims whitespace in text cells. Goal Seek… in that menu finds the number to put in one cell so a formula cell reaches a value you type: set the formula cell, the goal, and the number cell to change, then confirm to write that number. Cancel leaves the sheet unchanged. The same menu offers fill down/right, clear formatting, and clear values. Select data without its header when sorting. A concurrent edit cancels a pending sort; references elsewhere in the sheet are not rewritten to follow sorted rows.

The footer's bottom-right Import and export → Import… accepts .csv and .xlsx. Right-click a row number or column letter for Macro's contextual menu: clipboard actions, clear, hide/unhide, resize and fit-to-data; columns also offer whole-sheet sorting. The menu keeps an existing whole-row/column selection when opened within it. Insert/delete shifts cells, references and named ranges in local workbooks and in saved workbooks while connected; the commands are disabled while a saved workbook is offline or reconnecting. A collaborator's change that arrives while the cells are being moved cancels the insert/delete. Adding blank rows at the bottom remains available. Hidden cells are skipped by keyboard navigation.

Cells support Macro mentions without Markdown formatting. Type @ in a cell or the formula bar to search people, documents, channels, email and dates, then choose an item with the pointer or keyboard. Pasting a Macro app link renders a document pill. Legacy links retain navigation parameters. Routed links retain compatible block targets, but not workspace paths or pane-local search. Formulas still use @ inside a formula or email address does not start mention search. Other Markdown is literal text. Mentions remain attached through copy/fill, undo and collaboration; Excel/CSV export uses their display text. Plain URLs and email addresses are clickable; web links use the same hover preview as channel messages. Click the surrounding cell or use the formula bar to edit link text. AI set_cells accepts Macro URLs or the same <m-user-mention> / <m-document-mention> / <m-date-mention> encoding as docs. Formula references to mention cells use literal labels, never execute a label as a formula.

Dates: @today, @tomorrow, @next friday or @sep 28 offer a Dates bucket. Choosing one enters the date as text such as 10/16/2026 (in the browser locale's day/month order), not a chip, so a cell holding only that date becomes a date value like any typed date; around other text it stays text. Typing @date shows a calendar ([data-cell-date-picker]) in place of the menu, starting at today: arrow keys and Page Up/Down move the highlighted day, Enter or a click enters it, Escape closes it, and the cell editor keeps focus throughout. Date chips written earlier, or by AI set_cells, still show as clock chips; hover one for the full date, and edit the cell to change it. A cell holding only a date chip is a date value: =A1+7 produces a date, and Excel export writes a dated number rather than the label. Typing a date stores it as a date in a matching format: 01/2026 or Jan 2026 shows Jan 2026, Jan 01 2026 shows Jan 01, 2026, 2026-01-01 stays ISO and 1/1/26 shows 1/1/2026 (day first in day-first browser locales). The snapshot value is the serial (46023) with format: 'date' and a numberFormat such as mmm yyyy; the formula bar and cell editor show 1/1/2026. A cell already formatted as a date keeps its format, and text without a year (Jan 26, 1/2) stays text. =DATE(...) and arithmetic on date cells display as dates without choosing the Date number format. A difference of two dates stays a plain day count, and an explicit number format from the toolbar always wins. Number format → Date formats lists presets (Sep 2026, Sep 17, 2026, 09/17/2026, 17/09/2026, 2026-09-17, 17 Sep 2026, Thursday, September 17, 2026) and Custom date and time…, a dialog with a Format code field, a live Preview: status of the active cell, and Apply (disabled for codes that are not dates or times).

CSV imports a file up to 20 MB into the selection, adding rows if needed within the 100,000 × 16,384 limit. Existing cells in that rectangle are replaced, with undo available. Excel imports accept up to 50 MB, 300 sheets, 2,000,000 filled cells, and 100,000 rows × 16,384 columns (A–XFD) per sheet. The footer shows the sheet's size, and + Add columns appends 26 more. An import preview lists each sheet and warns about unsupported content (for example rich text, ink and sheet protection). Choose Insert new sheets to keep existing work, or Replace workbook to replace it in one undoable operation. Names must be unique when inserting sheets. Canceling leaves the workbook untouched; a replacement is blocked if the workbook changed while the preview was open. Encrypted files are rejected. Macro-enabled .xlsm workbooks import like .xlsx (here and as uploads), without their macros: VBA projects, Excel 4.0 macro sheets and form/ActiveX controls are dropped with an import note, and Download as Excel writes a macro-free .xlsx; an uploaded .xlsm keeps its macros in the original download. A legacy .xls is rejected here with a hint to upload it instead; uploads convert it (see Legacy Office uploads).

Import and export → Download as Excel (.xlsx) (also the header's "…" file menu Download submenu, beside Download as CSV) exports every sheet with formulas, current formula result caches, precise numeric values, custom Excel number formats, fonts, borders, and column widths. Named ranges, named constants, and names defined by formulas are retained and calculate. Imported legacy formulas keep Excel's implicit intersection, shown with @ as current Excel shows it; 3-D references such as SUM('Jan:Dec'!B2) are listed sheet by sheet. Imported merged ranges, hidden sheets/rows/columns, row heights, filters, and frozen panes are retained for export. Macro hides imported rows and columns, shows hidden sheets and individual cells of merged ranges; editing a covered merged cell omits that merge during export with a warning so the edit is preserved. Imported charts, images, shapes, text boxes, pivot tables, conditional formatting, validation and notes are written back. Structured table formulas and rich text are not fully supported; review import notes before conversion. CSV imports preserve long identifiers and leading zeros as text and never execute formula-like strings. Download as CSV in the same menu exports only the active sheet's current calculated values. Clipboard menu actions use the browser clipboard; if access is unavailable, use Cmd/Ctrl+V or Cmd/Ctrl+Shift+V in the grid.

Use + in the footer to add a sheet, select its tab to switch, and open the adjacent sheet menu to rename, duplicate, move, or delete. Double-click a tab to rename it, or right-click any tab for its Rename, Duplicate, Move left, Move right, and Delete actions. Drag a tab with the mouse to reorder sheets; a vertical accent bar marks where it will land, and the dragged sheet becomes active. Touch drags scroll the tab strip instead, so use Move left/right there. Sheet operations can be undone; the last sheet cannot be deleted. Each sheet remembers its selection. Tab navigation supports Left/Right and Home/End; a view-only user can switch tabs and copy cells. Adding or duplicating a sheet focuses its grid so typing immediately edits the new sheet. Formulas can refer across sheets, such as =Sheet1!B9 or ='Launch budget'!B9. Rename and delete are currently blocked when live formulas directly reference that sheet, to avoid breaking those references; INDIRECT text cannot be checked this way. The 10-sheet limit applies to local additions/imports; concurrent offline additions can merge above it without hiding another user's work.

Ribbon dropdowns, the import/export menu, and sheet actions use the shared Macro menu styling. Verify keyboard navigation, checkbox toggles, Escape, and restoring editor focus after menu actions or Escape and after Find/rename dialogs close. Clicking outside a ribbon menu onto the address or formula input should keep focus in that input. In narrow windows, sheet tabs should scroll while Add rows and the compact Import and export button remain visible. Imported-file warnings and dialogs should remain accessible in short or narrow windows.

Calculation runs in a worker. If it times out, source editing and undo remain available; simplify the formula or undo and use Retry. Check that a pending calculation does not freeze selection, and that export waits for current results. Appearance-only changes such as strikethrough do not recalculate. During value or formula changes, the last calculated results stay visible until the next results arrive. The footer shows Calculating… only for work taking longer than 250 ms; quick edits should neither flash raw formulas nor shift the footer.

Edits save through the collaborative document connection. Verify collaboration with the same document open for two users: edit different cells, then the same cell, and confirm both views converge. Also edit A1 on different sheets and confirm they remain independent. Remote selections have a tinted range outline, an active-cell border, and a name label in the same collaborator color. The footer repeats their names/colors. Idle connected selections stay visible; disconnected peers expire. Cursors should only appear for peers on the active sheet; switching a local tab must not move another user's tab. Switching back should immediately restore the remembered cursor for collaborators, without another cell click. A remote sheet deletion must cancel any draft on that sheet instead of committing it into the fallback sheet. Undo should reverse only the local user's edit. Close and immediately reopen after editing (including while offline) to check local recovery; reload to check server persistence. Viewers must be able to select and copy cells without editing them. If another user subsequently changes the same cell property or layout value, undo/redo keeps that newer work and reports a conflict without consuming the history step. Structural history also refuses to remove a sheet name still used by a surviving direct formula reference.

For local UI verification without creating hosted documents, development builds provide /app/component/spreadsheet-demo. It runs the real spreadsheet UI with local collaboration state. Clicking Share saves its current cells, formulas, formatting, all sheets, column widths, and added rows as a native document, then opens the normal sharing dialog. It waits for the save acknowledgement before leaving the sample; a failed save keeps the sheet editable and supports retry. Native creation, sharing, and network persistence require the spreadsheet-enabled document and sync services.

Phone verification should cover portrait and landscape, native grid/ribbon swipes, range handles, formula suggestions, sheet rename, view-only controls, and the software keyboard. The isolated browser fixture has separate Android Chrome and iPhone WebKit projects. Chromium uses trusted touch drags; WebKit uses native taps and mouse-pointer handle drags. A reduced test viewport only checks layout; verify actual keyboard resizing and iOS gesture behavior in the simulator or on a device.

Right-click a cell for the Macro cell menu: Cut, Copy, Paste, Paste values only, Clear values, Clear formatting, Fill down/right, and Comment on saved workbooks. Right-click inside a selected range to act on that range; outside it targets the clicked cell. Shift + F10 or the keyboard context-menu key opens the same menu for the selection; Escape returns focus to the grid. Fill requires a multi-cell range along that direction. Commenters can copy and comment without editing cells. Right-click inside the cell text editor retains the native text-editing menu.

Opening a saved spreadsheet from Files/Drive (including a favorite) keeps the Drive navigation sidebar in place. Collapse it with the sidebar control; the spreadsheet header then shows the navigation toggle to reopen it. Opening a spreadsheet does not change the saved sidebar preference. The header keeps the Files location breadcrumbs before the sheet title; click a location breadcrumb to return to that file listing.

Spreadsheet comments

On a saved spreadsheet, select a cell or range and choose Comment in the formatting ribbon (or ⌘/Ctrl + Alt + M) to compose beside the cell. A comment captures the sheet ID and selected range; later selection changes do not move the draft's attachment. The top-right triangle marks the first cell of a commented range. Hover any cell in that range to read its threads; choose Reply in the card to respond without opening the sidebar. Clicking the triangle also opens the card on touch devices. Interacting with a card keeps it open until dismissed so a reply is not lost when moving the pointer.

Comments in the document header opens the workbook commenting sidebar. Range labels navigate to the corresponding sheet and cells; deleted-sheet threads remain readable. Range threads are filtered by Open, Resolved, or All; Open is the default. Discussion contains comments about the whole workbook and is the sidebar's new-comment composer. To start a cell comment, select the range and use the ribbon, keyboard shortcut, or cell context menu. Both surfaces use the shared message controls for replies, edits, deletion, reactions, and attachments. Range threads also offer Resolve and Reopen. Mentions and replies use inbox notifications and message links. Older spreadsheet annotation comments are not displayed. Opening an inbox notification opens the sidebar and targets its comment/range. Range links leave Discussion on its normal timeline; workbook links open that discussion at the linked message and clear the previous range highlight or navigation error. A resolved range link switches the filter to Resolved so the targeted thread remains visible. Comment-only access can post/reply; view-only access can read. Edit/delete applies to the author's own comments, and failures retain the input draft. Draft demos must be saved before persistent comments are available.

Ask Macro about a spreadsheet

Ask Macro sits immediately left of Share. Select the relevant cells, then click it to open the agent composer in a split beside the workbook. The composer starts with the workbook mention followed by one space; nothing sends automatically. The mention captures the active sheet ID/name and normalized selected range at click time. Changing the selection later does not change that draft attachment. A viewer can ask questions; editing still requires edit permission.

In the local spreadsheet demo, Ask Macro first saves the entire workbook and waits for acknowledgement. A failed save leaves the draft editable and supports retry, without opening an empty chat. This path needs the updated native-document backend, just like Share.

The AI can use ReadSpreadsheet to inspect sheet names, used ranges, tables, raw inputs, formulas, typed results, errors, and formatting. CalculateSpreadsheet evaluates scratch formulas and what-if inputs without changing the workbook. EditSpreadsheet applies a validated batch of cell/formula/format edits, fill, row additions, column resizing, dropdowns, checkboxes, table creation, and sheet creation/rename/duplication/deletion. format_cells shows dates in any Excel date code with { "format": "date", "numberFormat": "mmm yyyy" }. Edits require a revision from a fresh read; a concurrent change rejects the entire batch so the AI can reread. Tool rows expand to show the actual results and warnings.

Verify with a saved workbook: select B4:E9, click Ask Macro, check the adjacent chat's mention and trailing space, and ask for a total or a what-if calculation. The workbook should stay unchanged for scratch calculations. Ask for an edit and check both source/formula and displayed result in the sheet and another connected client. A viewer's edit must fail; a concurrent manual edit must force a fresh read. These tool calls require the updated AI backend, AI editing worker, and sync service; the frontend alone cannot test their hosted path.

Desktop file drops

In a rebuilt macOS desktop app, drag a PNG and a PDF from Finder into a disposable Markdown document, then into channel and AI chat attachment inputs. Check that each drop shows the upload feedback and adds each file once, with images rendered inline in Markdown and attachments visible in the composers. Try multiple files and a second drop after the first completes. Discard unsent drafts after checking. Also check that the file picker still attaches files.

Use an actual Finder drag for this check: synthetic browser events bypass Tauri's native interception and cannot verify the desktop window configuration. The setting and native rebuild requirement are documented in the web agent guide.

Folder uploads

Dropping a folder (or choosing one with the folder picker) zips it in the browser and imports it as a folder tree: each subfolder becomes a folder, including empty ones, and each file becomes a document. The folder picker cannot see empty subfolders, so only dropped folders keep them. Office lock files (~$…), Thumbs.db, desktop.ini, __MACOSX and hidden files or folders are skipped. Extension case does not matter (REPORT.DOCX imports as .docx). If some files fail, the rest of the tree still appears and a toast says some files could not be uploaded. The unzip runs in the upload extractor Lambda, which the local stack does not run; verify folder imports against a deployed backend.

Legacy Office uploads

Uploaded legacy binary Office files are converted to their OpenXML equivalents on the server: .doc → .docx, .ppt → .pptx, .xls → .xlsx. This also applies to files inside an uploaded folder. The document keeps its id and name; its file type changes and the converted file becomes its latest version. Converted Word files then go through the regular DOCX pipeline, and converted workbooks open in the uploaded-workbook preview, where Edit in Macro imports them into a native spreadsheet.

While a legacy file uploaded in the last 10 minutes is converting, its view shows Converting … to a PowerPoint presentation… (or Word document / Excel workbook) with Download still available for the original. When the conversion lands, the split reopens in the right block (workbooks switch in place). If conversion fails or takes over 3 minutes, the view falls back to No preview available, and the original stays downloadable. Legacy files uploaded before conversion existed are not converted.

Conversion runs in the convert service through the upload finalizer Lambda, neither of which the local stack runs, so verify it against a deployed backend.

Presentations (PowerPoint)

Uploaded .pptx files open in the pptx block (/app/pptx/<documentId>). With the enable-pptx-editor PostHog flag (on by default in development builds; ENABLE_PPTX_EDITOR overrides it) the block is a full editor; with the flag off it offers the file for download, as before. The deck is parsed, rendered, and edited by the Rust pptx_engine compiled to WebAssembly in a lazily created module worker, so the first open of a session pays a one-time ~6 MB module download.

Layout and test hooks:

  • Ribbon (pptx-toolbar): tabs pptx-tab-<id> for Home, Insert, Design, Transitions, Slide Show, Review, and View, plus contextual Shape Format, Picture Format, Table Design, Layout (tables), Chart Design, and SmartArt Design tabs that appear for the selection (pptx-tab-shape-format, pptx-tab-picture-format, pptx-tab-table-design, pptx-tab-table-layout, pptx-tab-chart-design, pptx-tab-smartart-design). Undo/Redo sit left of the tabs; the save state (pptx-save-state), Download, and Present (pptx-present) sit right. Home: clipboard and Format Painter (pptx-format-painter: click, then click a shape or select text to paint the selection's look and text formatting; double-click keeps it armed until Escape; the stage carries data-format-painter while armed), New slide (pptx-new-slide), layouts, font (pptx-font-family) and size (pptx-font-size) boxes, Bold (pptx-bold) and the other run toggles, text and highlight colors (pptx-text-color), Character Spacing (pptx-char-spacing: Very Tight to Very Loose as pptx-char-spacing-<points>, or More spacing pptx-char-spacing-custom), bullets and numbering, list levels, line spacing, alignment, Text Direction (pptx-text-direction: Horizontal, Rotate all text 90°, Rotate all text 270°, and Stacked as pptx-text-direction-<horz|vert|vert270|wordArtVert>, More Options… pptx-text-direction-more), the shape gallery (pptx-insert-shape, then pptx-shape-<preset>), Arrange (pptx-arrange), Shape fill (pptx-fill) and outline (pptx-outline), Find, and Replace. Insert: table grid (pptx-insert-table, then a cell of pptx-table-grid), Pictures (pptx-image-input), Shapes, Chart (pptx-insert-chart, then pptx-chart-<kind>-<grouping>), SmartArt (pptx-insert-smartart; see SmartArt below), Video and Audio (pptx-insert-video, pptx-insert-audio; file inputs pptx-video-input, pptx-audio-input; clips up to 50 MB, or 2 MB in a shared presentation because the sync service keeps presentations under 4 MB, are embedded with a poster frame or speaker icon), Comment (pptx-insert-comment), Text box (pptx-insert-textbox), and Link (pptx-insert-link). A selected video or audio shape shows Play (pptx-media-play), which plays it over the shape (pptx-media-player, with pptx-media-video or pptx-media-audio) until the slide is clicked; in a slide show a click on the clip plays it in place (pptx-slideshow-media) instead of advancing. Design: slide background, and Variants that restyle every slide through the theme: Colors (pptx-theme-colors, then pptx-theme-colors-<set name>, such as Red Violet) and Fonts (pptx-theme-fonts, then pptx-theme-fonts-<pair name>, such as Georgia). Transitions: pptx-transition-<kind>, Effect options, duration (pptx-transition-duration), automatic advance, and Apply to all (pptx-transition-all). Color menus are PowerPoint's theme grid with tints, standard colors, Recent colors (pptx-recent-colors, custom colors picked lately), More colors… (pptx-more-colors), and Eyedropper (pptx-eyedropper, in browsers with a screen color picker).
  • Header & Footer (Insert ▸ Header & Footer pptx-insert-header-footer, Date & Time pptx-insert-date-time, or Slide Number pptx-insert-slide-number) opens pptx-header-footer, starting from what the current slide shows: Date and time (pptx-hf-date) updating automatically (pptx-hf-date-auto, format select pptx-hf-date-format listing today's date in datetime1–datetime13) or Fixed (pptx-hf-date-fixed, text pptx-hf-date-fixed-text), Slide number (pptx-hf-slide-number), Footer (pptx-hf-footer, text pptx-hf-footer-text), and Don't show on title slide (pptx-hf-not-on-title). Apply (pptx-hf-apply) changes the selected slides; Apply to All (pptx-hf-apply-all) every slide. A layout without the placeholder cannot show the element.
  • Slide Size (Design ▸ Customize ▸ pptx-slide-size): Standard (4:3) pptx-slide-size-standard, Widescreen (16:9) pptx-slide-size-widescreen, or Custom Slide Size… pptx-slide-size-custom, which opens pptx-slide-size-dialog (preset pptx-slide-size-preset, width and height pptx-slide-size-width|height in inches, or cm in metric locales, orientation pptx-slide-size-portrait|landscape, OK pptx-slide-size-ok). When content must change shape, pptx-slide-size-scale asks Maximize (pptx-slide-size-maximize) or Ensure Fit (pptx-slide-size-fit); a proportional change scales without asking. Stage, thumbnails, sorter, show, and print follow the new size.
  • Slide rail (nav "Slides", data-testid="pptx-slide-rail"): one pptx-thumbnail button per slide, labelled Slide N: <title>, with aria-current="true" on the current one. Hovering a thumbnail shows Duplicate slide, Hide slide/Show slide, and Delete slide; thumbnails reorder by dragging; right-click opens cut/copy/paste, new, duplicate, delete, layout, background, and hide. New slide is at the bottom. Shift-click selects a range and Cmd/Ctrl-click adds or removes a slide (aria-selected; the status bar shows pptx-status-selected); duplicate, delete, hide, layout, transitions, background, copy, cut, and dragging then act on every selected slide. Cmd/Ctrl+A selects all slides and Cmd/Ctrl+D duplicates them while the rail has focus.
  • Sections (rail and sorter): a deck with sections shows a pptx-section-header row (data-section-id, aria-expanded) before each section's slides with its name and slide count. Its caret (pptx-section-toggle, or double-click) collapses the section's thumbnails (view only); clicking the header selects its slides; dragging it onto another header moves the section there. Right-click a header for Rename Section (in place: pptx-section-rename, Enter keeps, Escape cancels), Remove Section, Remove Section & Slides, Remove All Sections, Move Section Up/Down, Collapse All, and Expand All. The slide menu's Add Section starts "Untitled Section" at that slide and opens its name for typing. Home ▸ Section (pptx-section-menu): Add (pptx-section-add), Rename (pptx-section-rename-current), Remove (pptx-section-remove), Remove All (pptx-section-remove-all), Collapse All (pptx-section-collapse-all), Expand All (pptx-section-expand-all).
  • Slide Sorter (View ▸ Slide Sorter, or pptx-view-sorter in the status bar; pptx-view-normal returns): every slide in a grid (pptx-slide-sorter, same pptx-thumbnail buttons and menu as the rail), with the same selection and drag-to-reorder; ★ marks a transition; double-click or Enter opens a slide in Normal view.
  • Slide Master view (View ▸ Master views ▸ Slide Master, pptx-view-slide-master) edits the slide masters and their layouts with the same stage, text editing, and Home/Insert tools as slides; every slide on a layout (or, for a master, on any of its layouts) follows. It opens on the current slide's layout. The rail becomes pptx-master-rail: one pptx-master-thumbnail per page (data-kind="master|layout", data-page-id the engine id, aria-current on the one edited, and a PowerPoint tooltip such as "Title Only Layout: used by slide(s) 3-8"), masters numbered with their layouts indented beneath. Placeholders show dotted outlines (pptx-placeholder-outlines). The contextual Slide Master tab (pptx-tab-slide-master, first) has Insert Layout (pptx-master-insert-layout: a "Custom Layout" with a title and the master's footers after the selected layout), Delete (pptx-master-delete, disabled with a reason in its tooltip while slides use the layout, or for a master's last layout or the last master), Rename (pptx-master-rename, dialog pptx-rename-layout-dialog with pptx-rename-layout-name and pptx-rename-layout-ok), Insert Placeholder (pptx-master-insert-placeholder, then pptx-master-placeholder-<content|text|picture|chart|table|smartArt|media>, -vertical for the vertical content and text ones; it lands in the middle of the layout, selected), the layout's Title and Footers checkboxes (pptx-master-title, pptx-master-footers), theme Colors and Fonts (pptx-master-theme-colors, pptx-master-theme-fonts), Background Styles (pptx-master-background-styles: the theme's twelve styles, pptx-master-background-style-<1-12>, and Reset Background, pptx-master-background-reset), Format Background (pptx-master-format-background, the background pane for the master or layout), Hide Background Graphics (pptx-master-hide-background), Slide Size, and Close Master View (pptx-master-close; the status bar's Normal button does the same). Right-click a page for Insert Layout, Duplicate Layout, Delete Layout/Master, Rename Layout/Master, and Format background…; Delete in the rail deletes the page, and Cmd/Ctrl+M inserts a layout. Formatting a whole placeholder (or whole paragraphs) of a layout or master also sets the text style its slides inherit, so bolding a layout's title bolds the titles of its slides. The status bar reads "Slide Master" (pptx-status-master); Design, Transitions, Animations, Slide Show, the Slides groups, notes, and find are not shown there. Engine and AI spelling: the outline's masters (ids, names, layouts with slideIds and placeholder types), any slide-id operation on shapes, text, tables, pictures, and backgrounds with a master's or layout's id, and addLayout, renameLayout, deleteLayout, insertPlaceholder, setLayoutOptions, and setBackgroundStyle.
  • Stage (pptx-stage, focusable): click selects a shape (pptx-selection, handles pptx-handle-<nw|n|ne|e|se|s|sw|w> and pptx-rotate-handle); Shift/Cmd/Ctrl-click adds to the selection, dragging on empty slide draws a marquee (pptx-marquee), and several selected shapes show pptx-selection-outline boxes inside one handle box; drag moves, handles resize (several shapes scale together) and rotate. Right-click opens a menu for what is under the pointer (shapes, text being edited, a table, a chart, or the empty slide). The stage covers exactly the slide, so slide point (x, y) is at stage.left + x × stage.width / slideWidth.
  • Text: double-click (or Enter/F2 on a selected text shape, or just start typing) starts editing; keystrokes go to a hidden textarea (pptx-text-input, "Slide text"). The caret is pptx-caret, an SVG line of zero width, so assert it with toBeAttached(), not toBeVisible(). Escape stops editing. Vertical text (Text Direction) is edited in place too: the caret lies across the rotated line, and arrow keys follow the screen (in text rotated 90°, Down is the next character and Left the next line). Turning a text box that resizes to fit its text swaps its width and height.
  • Tables: press a table that isn't selected and drag to move it, or release without dragging to type in the clicked cell in place (same caret); Tab and Shift+Tab move between cells (Tab in the last cell adds a row); drag across cells to select a range (pptx-cell-range), which the Table Design/Layout tabs and the table menu act on (merge/split, shading pptx-cell-shading, borders pptx-cell-borders, styles pptx-table-styles, insert/delete rows and columns, distribute, alignment, and Text Direction pptx-cell-text-direction, items pptx-cell-text-direction-<value>; a vertical cell's row grows to its text). Drag a column or row border of a selected table to resize it. To move a table that is already selected, drag near the frame's edge, or click outside it first.
  • Charts: double-click a chart, or Edit data (pptx-chart-edit-data) on Chart Design, opens the data grid (pptx-chart-data, cells pptx-chart-cell-<row>-<col>, Apply pptx-chart-apply); Chart Design also changes the type (pptx-chart-type), title, legend, data labels, and colors.
  • Equations (Office Math, typeset by the engine): Insert ▸ Equation (pptx-insert-equation, or Alt+=) starts a new equation at the caret of the text being edited (selected text on one line becomes its text), or in a new text box centered on the slide when no text is edited. Its arrow (pptx-insert-equation-menu) inserts a built-in equation at once (pptx-equation-prebuilt-<id>: area-of-circle, binomial-theorem, expansion-of-a-sum, fourier-series, pythagorean-theorem, quadratic-formula, taylor-expansion, trig-identity-1, trig-identity-2). The equation is written in a floating editor (pptx-equation-editor): LaTeX-style linear text (pptx-equation-input; \frac{a}{b}, x^2, \sqrt[n]{x}, \sum_{i=1}^n, \int_a^b, \left( \right), \begin{pmatrix}…\end{pmatrix}, \begin{cases}, Greek, \mathbb{R}, \text{…}), a live preview (pptx-equation-preview), errors (pptx-equation-error), a Display toggle (pptx-equation-display: own line, centered, versus inline), Insert/Done (pptx-equation-insert, or Enter; Shift+Enter is a new line of text), and close (pptx-equation-close, or Escape). An equation counts as one character (U+FFFC) of its paragraph: clicking it while editing the text selects it whole and shows its text in the editor, where every valid change lands on the slide as it is typed (one undo step per equation); double-click focuses the editor; Delete or Backspace removes a selected equation, and typing replaces it. The contextual Equation tab (pptx-tab-equation, shown while an equation is written or selected) has the built-ins (pptx-equation-prebuilt, which replace the equation's text), Linear (pptx-equation-linear, reopens the editor), Display (pptx-equation-display-toggle), Symbols (pptx-equation-symbols, then pptx-equation-symbol-<command> such as pm, alpha, infty, or pptx-equation-symbol-u<hex> for a character without a command), and the Structures galleries pptx-equation-structure-<fraction|script|radical|integral|large-operator|bracket|function|accent|limit-and-log|operator|matrix> (item n is pptx-equation-structure-<gallery>-<n>). Symbols and structures go into the linear text at its caret; a structure wraps the selected linear text and puts the caret in its first empty slot.
  • SmartArt: Insert ▸ SmartArt opens Choose a SmartArt Graphic (pptx-smartart-dialog): categories pptx-smartart-category-<all|list|process|cycle|hierarchy|relationship|pyramid>, layouts pptx-smartart-layout-<id> (default Basic Block List, vList2, hList1, process1, chevron1, cycle2, radial1, hierarchy1, orgChart1, venn1, pyramid1), the picked name pptx-smartart-picked, and OK pptx-smartart-ok. The graphic is inserted selected with its Text Pane open (pptx-smartart-pane, one pptx-smartart-pane-line input per node, its li carrying data-level; close pptx-smartart-pane-close; the tab on the graphic's left edge pptx-smartart-pane-toggle shows and hides it). Typing in a bullet edits the node live; Enter adds a node after it (splitting at the caret), Tab/Shift+Tab demote/promote, Backspace on an empty bullet deletes the node, and the arrow keys move between bullets. Empty nodes show "[Text]" (pptx-smartart-prompt, editor only). With the graphic selected, click a node to pick it (pptx-smartart-active-node, data-node), then type (replaces its text), press Enter/F2, or double-click to edit it in place (pptx-smartart-node-input; Escape ends); Delete removes the picked node. SmartArt Design: Add Shape (pptx-smartart-add-shape, then pptx-smartart-add-<after|before|above|below|assistant>), Text Pane (pptx-smartart-text-pane), Promote/Demote (pptx-smartart-promote, pptx-smartart-demote), Move Up/Down (pptx-smartart-move-up, pptx-smartart-move-down), Layouts (pptx-smartart-layouts, then pptx-smartart-layout-option-<id>), Change Colors (pptx-smartart-colors, then pptx-smartart-colors-<id>: accent0_1…, colorful1…colorful5, accentN_1…accentN_5), SmartArt Styles (pptx-smartart-styles, then pptx-smartart-style-simple1…5), Reset Graphic (pptx-smartart-reset), and Convert (pptx-smartart-convert, then pptx-smartart-convert-shapes or -text). The right-click menu offers Add Shape, Change Layout, Change Colors, Reset Graphic, and Convert to Shapes/Text. SmartArt in other layouts (from PowerPoint) keeps its drawing: text edits apply in place and structural edits say "this layout's structure can't be changed here".
  • Effects: Shape Format's Shape effects (pptx-shape-effects) and Picture Format's Picture effects (pptx-picture-effects) open PowerPoint's Shadow, Reflection, Glow, and Soft Edges flyouts (hover pptx-effects-<shadow|reflection|glow|soft-edges>). Tiles preview each effect: pptx-effect-shadow-<preset> (outerBottomRight, innerTop, perspectiveBelow...), pptx-effect-reflection-<preset> (tightTouching ... full8pt), pptx-effect-glow-<accent1-6>-<5|8|11|18> (More Glow Colors is pptx-effect-glow-more), pptx-effect-soft-edge-<1|2.5|5|10|25|50>, and -none for each. A choice applies to every selected shape as one undo step. Text effects (pptx-text-effects, Shape Format's WordArt styles) give the selected text, or whole selected shapes, a shadow or glow (pptx-text-effects-<shadow|glow>, tiles pptx-text-effect-shadow-<preset> and pptx-text-effect-glow-<accentN>-<size>).
  • Pictures (Picture Format): Corrections (pptx-picture-corrections, a 5 × 5 brightness × contrast grid of live previews, tiles pptx-picture-correction-b<±n>_c<±n> such as b+20_c-40), Color (pptx-picture-color, Recolor plus dark and light accent variations, pptx-picture-recolor-<value> with : written -, such as duotone-accent1), Transparency (pptx-picture-transparency, tiles pptx-picture-transparency-<0|15|30|50|65|80|95>), Change picture (pptx-picture-change), Reset (pptx-picture-reset, then pptx-picture-reset-picture or pptx-picture-reset-size), Picture border (pptx-picture-border), Arrange, and Size. Crop (pptx-picture-crop) toggles crop mode; its arrow (pptx-picture-crop-menu) offers Crop to Shape (pptx-crop-to-shape, then pptx-shape-<preset>), Aspect Ratio (pptx-crop-aspect, then pptx-crop-aspect-<w>x<h>, which crops the centered part and enters crop mode), Fill (pptx-crop-fill), and Fit (pptx-crop-fit). The picture's right-click menu has Crop and Format picture….
  • Crop mode (pptx-crop-overlay): the slide without the picture, the whole image ghosted outside the frame, and black crop handles pptx-crop-handle-<nw|n|ne|e|se|s|sw|w> (Shift keeps the aspect ratio, Ctrl/Alt crops both sides; dragging past the image pads it). Dragging the picture (pptx-crop-frame) or arrow keys move the image under the frame, and the round pptx-crop-image-handle-<nw|ne|se|sw> scale it. Enter, Esc, a click outside, or Crop again apply the crop as one undo step; rotated and flipped pictures crop in place too.
  • Edit Shape (Shape Format ▸ pptx-edit-shape): Change Shape (pptx-change-shape, then pptx-shape-<preset>) and Edit Points (pptx-edit-points; also right-click a shape ▸ Edit Points). Edit Points (pptx-edit-points-overlay) draws the outline as a red path (pptx-edit-points-path) with black square vertices (pptx-edit-point, data-selected on the clicked one, whose Bézier handles show as white squares pptx-edit-points-handle-in|out; on a straight side they sit a third of the way along it). Drag a vertex or handle to reshape, drag a segment to bend it; Ctrl/Cmd+click a segment adds a point and Ctrl/Cmd+click a vertex (or Delete) removes it. The path follows the pointer live and the shape updates on release, one undo step per gesture (Cmd/Ctrl+Z works in the mode). Right-click a vertex for Delete Point, Open/Close Path, Smooth/Straight/Corner Point; a segment for Add Point, Delete Segment, Open/Close Path, Straight/Curved Segment; anywhere for Exit Edit Points. Esc or a click away from the outline leaves the mode. Rotated, flipped, and grouped shapes edit in place; a preset becomes custom geometry (setCustomGeometry) keeping its fill, outline, effects, text and text area, and the box follows the outline.
  • Merge Shapes (Shape Format ▸ pptx-merge-shapes, enabled with two or more shapes, text boxes, or pictures selected): pptx-merge-<union|combine|fragment|intersect|subtract>. The result takes the first selected shape's formatting, text, rotation, and id (a picture stays a picture with its image in place), replaces the shapes, and is selected (Fragment selects every piece); curves stay curves. Engine and AI spelling: mergeShapes with shapes in selection order.
  • Format pane (pptx-format-pane): Format shape… in menus opens fill and line, size and position (with alt text pptx-alt-text), and text box and paragraph settings (Text box's Text direction select is pptx-pane-text-direction); Format background… opens the slide background. Its tabs (pptx-pane-tab-<shape|effects|size|picture|text>) include Effects (shadow, reflection, glow, and soft edge presets and values, such as pptx-pane-shadow-blur and pptx-pane-glow-size) and, for pictures, Picture (pptx-pane-brightness, pptx-pane-contrast, pptx-pane-recolor, pptx-pane-transparency, crop edges in percent pptx-pane-crop-<left|top|right|bottom>, and pptx-pane-picture-reset). Sliders apply while dragged, one undo step per drag.
  • Find and replace (pptx-find, Cmd/Ctrl+F and Cmd/Ctrl+H): find input pptx-find-input, pptx-replace-input, count pptx-find-count, and pptx-replace-all; Enter steps through matches, selecting each in its shape or cell.
  • Animations tab (pptx-tab-animations): with shapes selected, the gallery (pptx-animation-gallery, tiles pptx-animation-<entrance|emphasis|exit|path>-<effect>, pptx-animation-none) replaces their effect; Add Animation (pptx-animation-add) adds another; Effect Options (pptx-animation-options) holds directions (pptx-animation-option-<value>) and, for text, Sequence As One Object / By Paragraph (pptx-animation-sequence-object|paragraph). Timing: Start (pptx-animation-start select: onClick, withPrevious, afterPrevious), Duration and Delay in seconds (pptx-animation-duration, pptx-animation-delay), Move Earlier/Later (pptx-animation-earlier|later). Preview (pptx-animation-preview) plays the slide in place. The Animation Pane (pptx-animation-pane-toggle, pptx-animation-pane) lists pptx-animation-rows (click picks one and selects its shape; Delete or pptx-animation-remove removes it), and numbered pptx-animation-tags mark animated shapes while the tab or pane is open.
  • Export as pictures (header pptx-export-open) opens pptx-export: PNG or JPEG (pptx-export-png, pptx-export-jpeg), this slide, the selected slides, or all (pptx-export-current, pptx-export-selected, pptx-export-all), and a width (pptx-export-width); Export (pptx-export-run) downloads one picture (<deck> - SlideN.png) or a zip of SlideN pictures. Right-click a shape ▸ Save as picture… downloads it alone as a PNG cropped to its bounds.
  • View ▸ Show: Ruler (pptx-view-ruler: inch rulers along the slide's top and left edges, pptx-ruler-horizontal/pptx-ruler-vertical, measured from the center, with the selection shaded as pptx-ruler-span), Gridlines (pptx-view-gridlines, drawn as pptx-gridlines), and Grid settings (pptx-view-grid-settings: Snap objects to grid pptx-view-snap-grid, spacing pptx-view-grid-spacing in points, and smart guides pptx-view-smart-guides, and Display drawing guides pptx-view-drawing-guides). Snapping moves the dragged box's top-left corner to the grid where no smart guide is within reach; Alt-drag snaps to nothing. These choices are remembered in the browser.
  • Guides (View ▸ Show ▸ Guides pptx-view-guides, or Alt+F9) draw the deck's drawing guides as dashed lines over every slide (pptx-guides, one pptx-guide line per guide with data-orient and data-position in points; layout and master guides are pptx-layout-guide and do not move). Dragging a guide moves it in 1/24" steps (Alt: freely) with a tooltip of its distance from the slide's center in inches (pptx-guide-tooltip, such as ← 2.50); Ctrl+drag copies it; dragging it off the slide deletes it. A selected shape or the text being edited keeps clicks over a guide. Right-click a guide for Add Vertical Guide, Add Horizontal Guide, Color, and Delete; the empty slide's menu has Grid and Guides ▸ Guides, Gridlines, Smart Guides, Add Vertical Guide, and Add Horizontal Guide (new guides go to the center, or half an inch beside guides already there). While guides show, dragged shapes snap their edges and center to them. Guide changes are setGuides edits: saved in the file (PowerPoint's p15:sldGuideLst) and undoable.
  • Selection Pane (Arrange ▸ Selection Pane… pptx-selection-pane-toggle, or Alt+F10) opens pptx-selection-pane: one pptx-selection-row per object (data-shape-id), topmost first with groups nested. Click selects (Cmd/Ctrl adds), the eye (pptx-selection-eye) hides or shows, Show All / Hide All (pptx-selection-show-all, pptx-selection-hide-all) do every object, double-click or F2 renames (pptx-selection-rename), and dragging a row or Bring Forward / Send Backward (pptx-selection-forward, pptx-selection-backward) reorders it among its siblings.
  • Comments (Review tab pptx-tab-review): New Comment (pptx-review-new-comment, Insert ▸ Comment, right-click ▸ New Comment, or Ctrl+Alt+M) opens the Comments pane (pptx-comments-pane) with a draft attached to the selected shape, else to the slide (pptx-comment-draft, its marker pptx-comment-marker-draft; Post pptx-comment-post or Ctrl+Enter, Cancel pptx-comment-cancel). Each pptx-comment-thread (data-comment-id, data-resolved, aria-current when picked) shows the author (pptx-comment-author), a relative time (pptx-comment-time, "A few seconds ago"), the text (pptx-comment-text), replies (pptx-comment-reply-item), and a reply box (pptx-comment-reply, send pptx-comment-reply-post). Each comment's "…" (pptx-comment-menu) has Edit comment (pptx-comment-edit, then pptx-comment-edit-input and pptx-comment-save), Delete thread or comment (pptx-comment-delete), and Resolve or Reopen thread (pptx-comment-resolve, pptx-comment-reopen); resolved threads collapse and grey out. Speech-bubble markers (pptx-comment-marker, data-comment-id, aria-pressed on the picked one) sit at each thread's anchor: beside its shape's top-right corner, at its position, or at the slide's top-left corner; clicking one opens its thread. Review ▸ Delete (pptx-review-delete) offers this comment, all on the slide, or all in the presentation (pptx-review-delete-comment|slide|all); Previous and Next (pptx-review-previous|next) walk threads across slides; Show Comments (pptx-review-show-comments) toggles the pane, and its menu (pptx-review-show-menu) toggles Show Markup (pptx-review-show-markup, the markers). Thumbnails of slides with comments show pptx-thumbnail-comments. Comments are saved as PowerPoint for Microsoft 365 writes them (threaded comments with ppt/authors.xml); comments in the pre-2021 format show too and can be edited or deleted, but not replied to or resolved. They are signed with the host's user name (the fixture's ?author=, default "Alex Morgan"). Engine and AI: addComment, replyComment, editComment, resolveComment, deleteComment, and deleteAllComments; slide outlines list comments.
  • Spelling: misspelled words on the slide being edited get red wavy underlines (pptx-spell-squiggles, a pptx-squiggle per line piece with data-word), checked against an en-US dictionary that loads a moment after the editor opens. Words with digits, ALL-CAPS words, and web and e-mail addresses are skipped; the word being typed waits until the caret leaves it. Right-clicking an underlined word puts the caret in it, and the menu starts with suggestions (pptx-spelling-menu-suggestion), Ignore All (pptx-spelling-menu-ignore-all), and Add to Dictionary (pptx-spelling-menu-add); added and ignored words are remembered in this browser. Review ▸ Spelling (pptx-review-spelling, or F7) opens pptx-spelling-pane, which walks the deck from the current slide, selecting each word on its slide: the word (pptx-spelling-word), Ignore Once, Ignore All, and Add (pptx-spelling-ignore-once|ignore-all|add), suggestions (pptx-spelling-suggestion, aria-selected), and Change and Change All (pptx-spelling-change|change-all), ending with pptx-spelling-complete ("Spell check complete. You're good to go!", OK pptx-spelling-ok).
  • Links (Insert ▸ Link, Cmd/Ctrl+K, or Link… in the right-click menu) open pptx-link-dialog. While editing text it links the selection, or the whole link around the caret; at a bare caret it inserts the "Text to display" (pptx-link-text) as linked text. With shapes selected (not their text) it links the shapes themselves. Link to: Web Page or File (pptx-link-kind-web, address pptx-link-address; macro.com becomes https://macro.com), Place in This Document (pptx-link-kind-place: pptx-link-place-nextslide and the other jumps, or pptx-link-place-slide-<n>, with a preview pptx-link-preview), or E-mail Address (pptx-link-kind-email, pptx-link-email, pptx-link-subject). ScreenTip is pptx-link-tip; OK is pptx-link-ok; editing an existing link adds Remove Link (pptx-link-remove). Right-clicking linked text or a linked shape offers Edit link…, Open link (a web page in a new tab, or the linked slide), Copy link, and Remove link; Cmd/Ctrl+click follows a link while editing text. Engine and AI spelling: formatText link / linkTip and setShapeLink, with #slide=<id> or #nextslide-style jumps for places in the deck.
  • Slide show (pptx-slideshow, F5 from the start, Shift+F5 from the current slide): full screen with the slides' transitions; →/Space/click play the next animation step, then advance (pptx-slideshow-canvas carries data-slide-index and data-step; animated shapes are [data-piece] layers), ← goes back, a number then Enter jumps, B/W blank the screen, S shows notes, Esc ends. Over a link the pointer becomes a hand and its ScreenTip shows (pptx-slideshow-link-tip); clicking follows the link instead of advancing. Morph (pptx-transition-morph) plays as a scene (pptx-morph, data-pairs = objects matched): objects on both slides glide, resize, and turn to their new place (matched by a !! name, then the same name and kind, as a duplicated slide keeps them, then the same text, then the same placeholder); text-only boxes move without stretching; the rest fade out or in. With Effect options ▸ Words or Characters, text-only boxes morph word by word or letter by letter: each one travels to where the same word sits on the next slide (data-units = words matched; sprites carry data-shape, data-unit, and data-morph = from/to). A click finishes it.
  • Presenter View (Slide Show ▸ Presenter view pptx-present-presenter, Alt+F5): opens the audience show in a pop-up window (pptx-audience-canvas; double-click it for full screen) and turns the tab into the speaker console (pptx-presenter): current slide (pptx-presenter-current), next slide (pptx-presenter-next), notes (pptx-presenter-notes), timer (pptx-presenter-timer), counter (pptx-presenter-counter), All slides (pptx-presenter-grid-toggle, or G), and End slide show (pptx-presenter-end). Show keys work in either window. A blocked pop-up shows Open audience window (pptx-presenter-audience-closed). A video or audio clip on the current slide has a Play button on the console (pptx-presenter-media-play), and clicking the clip in the audience window also plays or pauses it instead of advancing: it plays on the audience screen (pptx-audience-media) while the console mirrors a video silently (pptx-presenter-media-video) with controls (pptx-presenter-media-controls: pptx-presenter-media-toggle, pptx-presenter-media-stop, seek pptx-presenter-media-seek, time pptx-presenter-media-time). With the audience window closed, the console's copy plays the sound.
  • Keyboard on the stage: Cmd/Ctrl+Z, Shift+Cmd/Ctrl+Z (or Ctrl+Y), Cmd/Ctrl+S saves now, Cmd/Ctrl+A selects all, Shift+Cmd/Ctrl+C and V copy and paste formatting, Cmd/Ctrl+C/X/V copy, cut, and paste shapes (and slides from the rail; also across decks), Cmd/Ctrl+D duplicates, Cmd/Ctrl+G and Shift+Cmd/Ctrl+G group and ungroup, Cmd/Ctrl+] and [ (with Shift: to front/back) reorder, Cmd/Ctrl+M adds a slide, arrows nudge (Shift for 10 pt), Delete removes, Tab walks through shapes, PageUp/PageDown change slides, F7 checks spelling, Ctrl+Alt+M adds a comment, Cmd/Ctrl+±/0 and Cmd/Ctrl+wheel zoom. The status bar shows the slide number and a zoom slider (pptx-zoom fits the slide).
  • Print (pptx-print-open beside Download, or Cmd/Ctrl+P) opens pptx-print: layout (pptx-print-layout-slides|notes|handouts3|handouts6), all/current/range slides (pptx-print-range, e.g. 1-3, 5), hidden slides, frames, and paper. Print (pptx-print-go) uses the browser's print dialog; Save as PDF (pptx-print-pdf) downloads a PDF directly. pptx-print-summary shows the page count or progress.
  • Speaker notes (pptx-notes) sit below the slide.

Changes save automatically 1.5 s after the last edit, when the tab is hidden, and when the editor closes; pptx-save-state reads Saved, Unsaved changes, Saving…, or Save failed. Each save stores the whole file as a new document version through PUT /documents/{id}/simple_save (limit 100 MB). Viewers without edit access get the same view with editing disabled.

Everyone with the deck open edits it live. The deck is shared through the sync service as Loro maps (one entry per shape, slide position, relationship, and part; see pptx_engine::collab), seeded from the stored file the first time someone who can edit opens it. Others' edits appear within a second; edits to different shapes merge, and edits to the same shape resolve to the latest. Undo takes back only your own changes. Presence shows in the stage's top-right corner (pptx-collaborators, one pptx-collaborator avatar per person), as outlines with name tags around the shapes others selected (pptx-peer-selection, data-peer="<name>", "… is typing" while they type), and as colored dots on the thumbnails of slides they are on. A viewer who opens a deck nobody has shared yet, or anyone when the sync service is unreachable, gets the stored file read-only.

Ask Macro, in the header beside Share, opens the agent composer in a split with the deck mentioned (nothing sends automatically). Macro AI reads decks with ReadPresentation (slides, layouts, theme colors, sections, transitions, header and footer, and every shape with its id, kind, placeholder role, position in points, text, table cells with merges and style, chart type and data, picture crop and adjustments, and shadow, glow, soft edge, and reflection effects; an equation stands in text as U+FFFC, the engine outline lists each paragraph's equations (index, LaTeX-style text, display flag), and the insertEquation and setEquation operations write them; ReadContent returns the same description) and changes them with EditPresentation, an atomic batch of the editor's own operations saved as a new version (saveAs creates an edited copy instead). When an EditPresentation result arrives in chat, an open editor of that deck reloads in place and says Updated with changes made elsewhere. If it holds unsaved edits it keeps them and says the deck also changed elsewhere; saving those edits replaces the other version.

To exercise the editor without a backend, run the browser fixture from apps/web:

bunx vite --config src/features/block-pptx/browser-test/vite.config.ts
# http://127.0.0.1:3018/?deck=generated/kitchen-sink-financial.pptx

It mounts the real editor and worker over corpus decks (?readonly for a viewer, ?autosave=0 to save only on demand). window.pptxFixture exposes saved(), saves(), engine(), errors(), notices(), and externalEdit(ops), which applies operations to the stored copy with a second engine instance and announces the change the way an AI edit does. With document, user, worker, socket, and token parameters the fixture is one collaborator on the real sync service (browser-test/sync-server.ts boots the compiled sync Worker in Miniflare; build it once with \cd services/sync-service && just worker-build), and window.pptxFixture.collab exposes the connection status, peers, and shared entries. Its Playwright suites run with bunx playwright test --config src/features/block-pptx/browser-test/playwright.config.ts (set PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH when the bundled browser is not installed); collaboration.browser.e2e.ts opens several people on one deck. The editor sections above were verified on this fixture; the /app/pptx route itself needs a backend with an uploaded deck.

Designs (Figma)

Uploaded .fig files open in the fig block (/app/fig/<documentId>), and Design in the create menu (key I, also in project create menus) makes a new, empty one. Both are behind the enable-fig-viewer PostHog flag (on by default in development builds; ENABLE_FIG_VIEWER overrides it); with the flag off the block offers the file for download. The file is decoded, rendered, and edited by the Rust fig_engine compiled to WebAssembly, in a primary worker plus up to three tile-raster helpers. The canvas composites 512 px tiles; while zooming it shows the nearest cached scale, then sharpens. Opening shows the file's embedded preview (fig-opening-preview, when present) while the engine and live session start in parallel. A coarse page preview appears before detailed tiles; the visible area gets priority over refining the whole-page preview. Reopening a file shares the compiled engine code; a new open starts compiling it while the file downloads. Each file keeps its own worker memory. Hover and remote cursors update the overlay without redrawing an unchanged page canvas. The canvas becomes editable only after the shared changes have been applied.

People with edit access get an editor; others get the same view read-only (no shape tools, no editing shortcuts). Edits save automatically 1.5 s after the last change, when the tab is hidden, and on close, as a new document version through PUT /documents/{id}/simple_save. Saving rewrites only the edited parts of the original file, so everything else in it survives.

Everyone with the design open edits it live, as in Figma. Changes go through the sync service as Loro maps holding the state of each layer edited since sharing began (see fig_engine::collab); the stored .fig stays the base everyone opens, and the first person who can edit seeds the shared copy. Others' edits appear within a fraction of a second; edits to different layers merge, and edits to the same layer resolve to the latest. Undo takes back only your own changes. Of the people editing, one (the lowest peer id) stores the merged file; the others' toolbars read saved while the sync service holds their changes. Presence: other people's pointers with name tags (fig-peer-cursor, data-peer="<name>", "(typing)" while they type), outlines of what they selected in their colors, and their avatars at the top right of the canvas (fig-collaborators, one fig-collaborator button per person, data-peer="<name>"). Clicking an avatar follows that person: their page and view, inside a colored frame (fig-following) until you click, scroll, or press Stop. A viewer who opens a design nobody has shared yet, or anyone when the sync service is unreachable, gets the stored file read-only; for editors a bar above the design (fig-session-notice) says why, with Retry (fig-session-action). Editing offline is not offered: without the sync service's changes, storing would overwrite other people's. The shared changes apply only to files the session produced (each stored file is listed in the shared metadata before it is stored, and only once the sync service has that record; offline, storing waits). Whoever opens a file stored outside the session (a new upload, an AI edit, a restored version from before such a replacement) starts the shared design over on it, dropping the changes made on the file it replaced; people still in the design get a Reload bar and their copy turns read-only.

Macro AI reads designs with ReadDesign (ReadContent returns the same description; attaching a design points the agent at it): the pages, numbered from 1, each page's top-level frames and sections with ids, sizes and positions, the text in each frame in reading order (instances' text included), the components the frame's instances use, and the file's components with their properties and variants, styles, and variable collections (fig_engine::describe). Hidden layers are left out; pages reads only some pages. The tool reads the stored file, so edits still held by a live session appear once one of the editors has saved. Tools cannot edit designs. Search indexes each page's name, frame names, and text as one chunk, so content search finds a design by the words on its canvas.

Layout and test hooks:

  • Opening a design: download, collaboration connection, and decoding use the same editor-shaped loading shell (fig-opening), without a floating thumbnail or Macro spinner. Its canvas has the same bounds as the editor. The canvas status (fig-loading-status) clears when content is first drawn or an empty page is ready; the surrounding controls need not wait for tiles.
  • Dragging layers with composited effects: when a layer cannot be moved as a cached sprite, the canvas updates in complete preview frames rather than tiles from different drag positions. Full detail returns together after release. Other collaborators still receive the movement as it happens.
  • Editor chrome: both sidebars use the app's page background color, including the loading shell, so they follow the active theme. Horizontal and vertical sidebar dividers use the same edge-frame color as the outer navigation rail. The same divider spans the editor below the file header, including while the design is loading. Drag their inner edges (fig-resize-left|right) or focus an edge and use Left/Right to resize. fig-toggle-left|right hides or shows a panel while preserving its width. The file header shows save status (fig-save-state, data-state is saved, unsaved, saving, or error).
  • Left panel tabs: File (fig-tab-layers) and Assets (fig-tab-assets); arrow keys switch tabs. Assets (fig-assets) shows component previews (fig-asset, searchable with fig-assets-search). fig-assets-grid|list switches the local components between thumbnail cards and compact rows; previews load near the visible portion of the list. Editors click one to place an instance in the middle of the view, and the arrow beside it (or a click, read-only) goes to the main component. ⌥⌘K makes the selection a component (a frame becomes one; other layers are wrapped), duplicating or pasting a main component places an instance, and ⌥⌘B detaches an instance into ordinary layers (nested instances too). Components of one set are listed under the set's name; right-clicking selected components offers fig-menu-combine-as-variants.
  • Team libraries (Assets tab): the book button (fig-assets-libraries) opens the Libraries dialog (fig-libraries-dialog). "This file" shows the publishing state (fig-library-status) and Publish… (fig-library-publish), which opens fig-publish-dialog: the changes since the last publish (fig-publish-change: New, Changed, or Removed components, component sets, styles, and variables, with descriptions), a note (fig-publish-note), and fig-publish-confirm. Names starting with _ or . stay private. Below, the other .fig designs the person can open (fig-library-row, data-library is the document id, searchable with fig-libraries-search) each have a switch that turns the library on for this file (stored in the file, so collaborators share it); a row notes its published asset count, "Not published", or "Unavailable". Each enabled library gets a section in Assets (fig-library-assets): its components as thumbnails (fig-library-asset, data-key; click to place an instance in the middle of the view, or drag one onto the canvas), grouped by component set or name folder, then its color, text, and effect styles and its variables (click applies to the selection: a style, or a color variable bound to the first fill). Inserting copies the component with what it uses (nested components, styles, variables, images) onto the internal canvas as Figma does, so it is not listed with the file's own components. When a library publishes newer versions of what the file uses, Library updates available (fig-library-updates, bottom left; fig-assets-updates in the Assets header) opens the review (fig-library-review): each changed asset (fig-library-update) with this file's copy beside the published version, the library's note, fig-library-update-one, and fig-library-update-all. Updating keeps instances' overrides; it is one undo step. Libraries are read when the design opens and again when the Libraries dialog opens.
  • Layers panel (fig-layers-panel, toggled with ⌥1): open layer search with the magnifier (fig-search-toggle) or ⌘/Ctrl+F, including from Assets. Search uses fig-layer-search; results are fig-search-hit. Escape closes search and returns focus to its button. Below are the collapsible pages list (fig-page buttons; pages named only with dashes are dividers; editors get fig-page-add, double-click to rename in fig-page-rename, and a hover fig-page-delete), and the layer tree (fig-layer-row, data-layer-id is the Figma node id such as 12:34, or I12:34;56:78 inside instances). Rows list top-most first, as Figma does; instance and component rows are purple. A click selects a row, Shift-click the rows between it and the selection, ⌘/Ctrl-click toggles a row. After a click, editors' arrow keys nudge the selection as in Figma (Tab, Enter, and ⇧Enter move through the layers); read-only, they move through the tree (fig-layer-tree; Left and Right collapse and expand). ⌥L collapses everything, and the list scrolls to the selection. Editors can double-click a row (or ⌘R) to rename it (fig-layer-rename), toggle visibility and lock on hover (fig-layer-visibility), and drag rows to reorder or move them into frames and groups. Right-click a row for the layer menu (below).
  • Canvas (fig-canvas): click selects with Figma's rules (inside a top-level frame the click selects the frame's child; sections are transparent; ⌘/Ctrl-click selects the deepest layer; double-click goes one level deeper; Shift-click adds). Click a frame's canvas name label to select that frame; Shift-click toggles it in the selection. Drag the label to move the frame. In editable files, double-click the label to rename it in place (fig-frame-rename): Enter or blur commits, Escape cancels, and undo restores the previous name. Dragging empty canvas draws a selection marquee. Hover outlines what a click would select; holding ⌥ measures from the selection to the hovered layer. Scroll pans, ⌘/Ctrl+scroll or pinch zooms, Space-drag or middle-drag pans. When editing: drag a layer to move it (⌥ drags a copy, ⇧ constrains, edges and centers snap to siblings and the parent frame with red guides; dropping inside another frame moves the layer into that frame, and dropping outside frames moves it onto the page; in an auto layout frame the dragged layer takes the slot it is dropped over; the canvas draws the moving layers itself while they are dragged, the design panel's X and Y follow them, and the move is one undo step, which the others in a live design see as it goes), drag the selection's corners or edges to resize (⇧ keeps proportions; several selected layers scale together), drag just beyond a corner of one layer to rotate it (⇧ snaps to 15°), draw with the frame, rectangle, ellipse, and text tools (a click places a default size; new layers go into the frame under the pointer), double-click or Enter on a text layer to type into it (see Text editing below), and drop or paste image files to place image-filled layers. The pen (P) draws a path: click places a corner, drag pulls out a curve's handles, clicking the first point closes the path, and Enter or Escape ends an open one; the result is a Vector layer with a 1 px black stroke. The pencil (⇧P) draws freehand: a drag becomes a smoothed open Vector layer, and the pencil stays chosen until Escape or another tool. Double-clicking a selected shape or vector (or Enter) edits its points: drag points and handles (⌥ breaks a handle's mirroring), Delete removes the selected point, and Escape or a click elsewhere ends editing; rectangles, ellipses, and the like become vectors when their points change. ⌘C also puts the selection on the system clipboard as Figma does (HTML whose comments carry (figmeta) and the layers as a (figma) .fig document, plus Macro's (macroimages)), so layers paste into other designs and tabs, and layers copied in Figma paste here (their images arrive as gray placeholders, since Figma's clipboard does not carry them). Pasted instances whose component is not in the file are detached. In a live design, the others receive pasted layers as the properties Macro models, so fields it does not model (prototype links, plugin data, and the like) survive only while the person who pasted is the one storing the file. Paste places layers by Figma's rules: into a selected frame (where they were if that is inside it, else centered), beside a selected layer, or on the page where they were when that is in view and in the middle of the view otherwise. Layers inside instances cannot be moved, resized, or deleted, but their name, fills, strokes, opacity, visibility, and text (typing included) are overridden on the instance, as in Figma; text bound to a component text property sets the property. A layer row's padlock toggles that layer. Lock/unlock selection (⌘/Ctrl⇧L or the menus) unlocks when all selected layers are locked, otherwise locks the selection. It reads current engine state after pending local and remote edits; verify multi-selection, repeated toggles, and undo/redo.
  • Main menu, layout grids, and guides: the menu at the top of the layers panel (fig-main-menu) groups commands into File, Edit, View, Object, and Arrange. Open Edit (fig-main-edit) for Undo/Redo (fig-undo, fig-redo); open File (fig-main-file) for fig-menu-export-frames-pdf ("Export frames to PDF": the page's top-level frames as a multi-page vector PDF), fig-menu-export-selection. View (fig-main-view) contains the Layout grids and Rulers toggles (also in the zoom menu, fig-layout-grids-toggle). Submenus open with hover or Right Arrow; Left Arrow returns to the parent. Keyboard shortcuts remains a top-level menu command. Frames' layout grids draw over the canvas (never in exports) until toggled off. With rulers shown (⇧R), editors drag from the top ruler for a horizontal guide or the left one for a vertical guide; a guide dropped on a top-level frame belongs to it, otherwise to the page. Dragging a guide moves it, dropping it on its ruler deletes it, and guides hide with the rulers. Moving and resizing layers snap to guides and to the edges of visible columns and rows of the frames they are over, as well as to other layers.
  • Context menu (fig-context-menu): right-clicking a layer on the canvas selects it (a selection it belongs to is kept) and lists Figma's layer actions with their shortcuts, as fig-menu-<action> items: copy, paste-here, paste-replace, copy-png, bring-to-front, bring-forward, send-backward, send-to-back, group, ungroup, frame-selection, add-auto-layout / remove-auto-layout, create-component, detach-instance, toggle-visible, toggle-locked, flip-horizontal (⇧H), flip-vertical (⇧V), rename, and delete (read-only viewers get copy, copy as PNG, and zoom to selection). On empty canvas: paste-here (pastes copied layers with their top left at the click, into the frame there), toggle-ui, toggle-rulers, toggle-pixel-grid, toggle-outline, the zoom actions, and select-all. paste-replace puts the pasted layers in place of the selection, centered on it, in one undo step. Both paste what is on the system clipboard (layers from any file or tab, as ⌘V does) when the browser lets the page read it, and otherwise the layers copied in this design.
  • Design panel (fig-design-panel, toggled with ⌥8): the fixed selection header offers Create component (fig-create-component) and Boolean operations for editable selections. Design and Prototype tabs support arrow-key navigation. Zoom and Present are in the top right; the selected layer's heading stays visible while properties scroll. Position groups alignment, X/Y, rotation, 90° rotation (fig-rotate-90), and flip controls (fig-flip-horizontal|vertical). Layout groups dimensions, resizing, auto layout, and clip content. The aspect-ratio lock (fig-lock-aspect-ratio) constrains inspector width/height edits together. Appearance groups opacity, radius, and visibility (fig-appearance-visible). Alignment buttons (fig-align-<left|center|right|top|middle|bottom>), name (fig-name), position, size, rotation, radius, opacity (fig-field-<x|y|w|h|rotation| radius|opacity|font-size|stroke-weight>; type a value or arithmetic, or drag the label to scrub; a flip is kept apart from rotation: after ⇧H the angle and X/Y read as before, a vertical flip reads as 180°, and typing a rotation or position keeps the flip), fills and strokes (fig-fills, fig-strokes, rows fig-fill-<n> / fig-stroke-<n>, top paint first, with a hex input fig-fill-<n>-hex, opacity -opacity, -visibility, and -remove; "+" adds; drag a row's grip to reorder), stroke weight, position, and dashes (fig-field-dash: 4, 2, or None), clip content (fig-clip-content), the Typography section for text layers (fig-type: family fig-font-family (opens the font picker fig-font-picker: search fig-font-search, options fig-font-option with data-family, the file's fonts first, then Google Fonts, each previewed in its face; use Up/Down to browse results and Enter to apply; a missing font shows fig-font-missing), weight fig-font-weight, and size on one row, line height fig-field-line-height as Auto, a percentage, or pixels, letter spacing fig-field-letter-spacing, and horizontal/vertical alignment. Open Type settings (fig-type-settings) for fig-italic, fig-underline, strikethrough, paragraph spacing, and case (fig-text-case). Auto width / auto height / fixed size (fig-text-resize) is in Layout; Text content is a disclosure in Typography), auto layout (fig-auto-layout-section: "+" or ⇧A adds it, wrapping other layers in a new frame; "−" or ⌥⇧A removes it; direction fig-layout-direction, gap fig-field-gap as a number or Auto, padding fig-field-padding-h|v or independent sides via fig-padding-independent (fig-field-padding-left|top|right|bottom), and the alignment grid fig-layout-align), width and height sizing (fig-sizing-w|h: Fixed, Hug, Fill, beside each dimension) and fig-absolute ("Ignore auto layout") for layers in auto layout, constraints (fig-constraint-h|v; children follow them when their frame is resized, and groups scale theirs), effects (fig-effects: "+" adds a drop shadow; rows fig-effect-<n> open fig-effect-<n>-settings; inside that popover, fig-effect-<n>-type picks the kind alongside offset, blur fig-effect-<n>-blur, spread, and color in a popover beside the inspector), boolean operations (the selected-layer header menu fig-boolean-menu; a boolean layer also shows fig-boolean-row: fig-boolean-<union|subtract|intersect|exclude>, pressed for the boolean's current operation, which a click changes, and fig-flatten), a Layout guide section for frames (fig-layout-grids: "+" adds a 10 px grid; rows fig-grid-<n> with color fig-grid-color-<n>, kind fig-grid-type-<n> (GRID, COLUMNS, ROWS), fig-grid-visible-<n>, fig-grid-remove-<n>, and fig-grid-settings-<n> opening count fig-grid-count-<n> (a number or Auto), type fig-grid-align-<n> (Stretch/Left/Center/Right), width or height fig-grid-size-<n>, margin or offset fig-grid-offset-<n>, and fig-grid-gutter-<n>), and the Export section (fig-export): "+" adds a preset (1x PNG, then 2x, 3x…; rows fig-export-row-<n> with size fig-export-size-<n> typed as 2x, 0.5x, 512w, or 300h, format fig-export-format-<n> (PNG, JPEG, SVG, PDF). Open fig-export-options-<n> for fig-export-suffix-<n> and format-specific settings: fig-export-outline-text-<n>, fig-export-include-id-<n>, or fig-export-quality-<n>; "−" is fig-export-remove-<n>). Presets are saved in the file (read-only viewers keep theirs for the session). fig-export-button ("Export ") downloads one file named as Figma names it ([email protected]) or a ZIP of several, fig-export-preview-toggle shows a preview (fig-export-preview), and the Export header’s "Copy as" menu contains Copy as PNG and fig-copy-svg (Copy as SVG); ⌘/Ctrl+⇧E exports the selection with its presets (1x PNG without). Inspector choices use custom listboxes: click the trigger, then its named option; use arrow keys/Enter to choose and Escape to dismiss with focus restored. Read-only viewers see the same values as text. The Code tab is Dev Mode's inspect panel (fig-dev-inspect): size and position (fig-dev-width, fig-dev-height), auto layout padding and gap (fig-dev-padding), typography with its text style (fig-dev-text-style, fig-dev-font), colors with the style or variable they come from (fig-dev-colors, rows fig-dev-color), effects, code with a language switch (fig-code-lang-css|tailwind|swiftui|compose; the code is fig-css or fig-code-<language>, copied with fig-code-copy), and the layer's and its layers' export presets as downloads (fig-dev-asset). While the Code tab shows, hovering a layer measures from the selection without ⌥, and a selected auto layout frame's padding and gaps are shaded with their sizes. With nothing selected it shows the page name and canvas color (fig-page-color opens the picker), then the file's local styles (fig-local-styles: color, text, effect, and grid styles by folder; a fig-local-style row, with data-style-name, opens to rename it in fig-local-style-name, change a color style's color (fig-local-style-color) or a text style's size (fig-local-style-size), which every layer using it follows, or fig-local-style-delete). With several layers selected (fig-mixed) it shows what they share; differing values read "Mixed", typing sets all of them, and "+" replaces mixed fills or strokes with one.
  • Components and styles in the design panel: for an instance, fig-instance-section names its main component (fig-main-component; fig-go-to-main goes to it, fig-swap-instance picks another component in fig-component-picker with fig-component-search and fig-component-choice items, preferred ones first, and fig-reset-instance is "Reset all changes") and shows its variant properties as menus (fig-variant-<Name>; switching keeps the overrides that still apply), boolean properties as checkboxes, text properties as fields, and instance swap properties as component pickers (fig-prop-<Name>, spaces as dashes; a changed one gets fig-prop-reset-<Name>), then the same for nested instances whose properties are exposed. A main component, component set, or variant gets fig-component-section: "+" opens "Create component property" (fig-new-property-kind: Variant, Boolean, Text, or Instance swap; fig-new-property-name; fig-new-property-create; a variant property on a lone component makes it a component set), a set's variant properties are renamed in fig-variant-property-<Name> and removed with fig-variant-property-remove-<Name>, a variant's values are typed in fig-variant-value-<Name>, fig-add-variant adds a variant (a lone component becomes a set of two), and each property (fig-component-property) has its name (fig-property-name-<Name>), default (fig-property-default-<Name>, which the bound layers show), and fig-property-delete-<Name>. A layer inside a main component gets fig-bindings-section: which property drives its visibility, text, or instance (fig-bind-VISIBLE|TEXT|INSTANCE_SWAP; the last item creates a property from the layer), and for nested instances fig-expose-instance. The Fill, Stroke, Text, and Effects sections show the shared style in use (fig-style-FILL|STROKE|TEXT|EFFECT, fig-style-detach-<kind> detaches it, keeping its values) and a style picker (fig-style-picker-<kind>: fig-style-option items by folder, and fig-style-create-name with fig-style-create to make a style from the layer and apply it). Fill and Stroke also show the color variable the first paint uses (fig-variable-FILL|STROKE, fig-variable-detach-<kind>) and a picker of the file's color variables (fig-variable-picker-<kind>, fig-variable-option). Frames get fig-variable-modes: per collection with several modes, fig-variable-mode-<Collection> picks one (or Auto, inherited), and bound colors inside follow. With nothing selected, fig-variables lists the collections and their variables (fig-variable, data-variable-name) with a value per mode.
  • Color and paint pickers: a paint's swatch (fig-fill-<n>-swatch, fig-stroke-<n>-swatch, fig-effect-<n>-swatch) opens the picker beside the panel (fig-paint-popover): the kind (fig-paint-type: Solid, Linear, Radial, Angular, Diamond, or Image, which asks for a file), for gradients the stop bar (fig-gradient-bar; click it to add a stop, drag a fig-gradient-stop, Delete removes the selected one), and the color picker (fig-color-picker): the saturation and brightness square (fig-color-area), hue and opacity sliders (fig-color-hue, fig-color-alpha), the eyedropper where the browser has one (fig-color-eyedropper), the format (fig-color-format: Hex, RGB, HSL, HSB) with its fields fig-color-field-<n> and opacity fig-color-alpha-field, and "On this page" swatches (fig-color-swatch). Dragging previews live and is one undo step; Escape closes the picker.
  • Toolbar (fig-toolbar): one floating row centered across the editor at the bottom. Tool groups retain their last-used tool: Move/Hand (fig-tool-menu-move), shapes (fig-tool-menu-rectangle: Rectangle, Line, Arrow, Ellipse), and Pen/Pencil (fig-tool-menu-pen). Menu choices are fig-choose-<name>; the group's current button is fig-tool-<name>. Frame and Text have their own buttons. Shortcuts remain V/H, F, R/L/⇧L/O, P/⇧P, and T; ⇧ while drawing snaps lines to 45°. Arrow keys, Enter, and Escape operate the menus. Comment (fig-tool-comment, C; fig-comments-unread counts unread threads) is followed by Actions (fig-actions-button) and Dev Mode (fig-panel-tab-code, a toggle). Selecting a drawing tool exits Comment. The boolean menu is in the selected layer's inspector heading (fig-boolean-menu: fig-menu-boolean-<union|subtract|intersect|exclude> and fig-menu-flatten). Zoom (fig-zoom-menu) and Present (fig-present-button, ⌥⌘↵ / Ctrl+Alt+Enter) are in the inspector header, or float at the canvas's top right while that panel is hidden. The shortcuts dialog (fig-shortcuts, Ctrl+⇧+?) is also in the main menu.
  • Comments (the comment tool, C): Macro comments, the same threads as markdown, PDF, and spreadsheet comments. Pins (fig-comment-pin, data-thread, data-unread) sit at a constant size over the canvas; a click places a comment on the top-level frame there (it moves with the frame) or on the canvas, written in fig-comment-draft with the app's message composer (Enter posts; @ mentions people and documents; attachments; Escape cancels). A pin opens its thread beside it (fig-comment-popover, data-thread, data-resolved): the shared message thread (replies, reactions, edit, delete, copy link) under a header with fig-comment-resolve / fig-comment-reopen and fig-comment-close. The right panel becomes the comments list (fig-comments-panel), with Present and Zoom in its header. Search (fig-comments-search) matches the first comment, the latest replies, and author names. Open fig-comments-filter for fig-comments-filter-<open|resolved|all>. Rows fig-comment-row have data-unread; a row opens the thread on its page. Below the rows, fig-comments-discussion holds comments on the whole design. A copied comment link (?comment_id=) or a comment notification opens the tool at its thread (or at the discussion). Escape closes the thread, then the tool. In the app, a pinned thread's root carries a fig thread anchor; the browser fixture keeps comments in memory, with plain stand-ins for the thread (fig-comment-item, fig-comment-reply) and composer (fig-comment-input).
  • Present (fig-present): the selection's top-level frame (or the first flow's start, or the first frame) scaled to fit (fig-present-screen, data-frame), with the prototype playing: clicks on hotspots navigate (with dissolve, slide, push, move in and out; Smart Animate dissolves), open overlays (fig-present-overlay; a click outside closes one that allows it), go back, close overlays, open links, and hover and after-delay interactions run; a click on nothing flashes the hotspots (fig-present-hint). → ↓ Space and ← ↑ ⇧Space step through the flow (its frames as it reaches them, or every frame in page order when the frame is in no flow), R restarts, Escape leaves. The bottom bar shows the flow (fig-present-flow), frame name (fig-present-name), position (fig-present-index, "2 / 3"), fig-present-previous / fig-present-next, Copy link to frame (fig-present-copy-link, a link with ?present=<frame id> that opens presenting it), and fig-present-exit. Variant changes and Scroll to are not played.
  • Prototype tab (fig-panel-tab-prototype, beside Design; fig-prototype-panel): for a top-level frame, its flow starting point (fig-flow-add, name fig-flow-name, fig-flow-remove); for a layer, its interactions (fig-proto-interaction): editors add a click interaction (fig-proto-add). Click its compact row (fig-proto-open) to open settings beside the inspector (fig-proto-settings); adding an interaction opens its settings automatically. Set its action (fig-proto-action: Navigate to, Open overlay, Back), destination frame (fig-proto-destination), animation (fig-proto-transition), and duration in ms (fig-proto-duration), or remove it (fig-proto-remove); other triggers and actions read as text (fig-proto-summary). The page's flows (fig-flow) present from their start. While the tab is open the canvas draws connections (fig-noodle, data-from, data-to; the selection's when it has any) and flow tags (fig-flow-badge). The first connection on a page without flows starts "Flow 1" at its frame, as in Figma. Edits are undoable, shared live, and saved as Figma's prototypeInteractions and prototypeStartingPoint.
  • Text editing (fig-text-editing): the text box fig-text-box, the caret fig-text-caret, and selection rectangles fig-text-selection are drawn from the engine's layout; typing goes to a hidden textarea (fig-text-editor, which holds the selection as selectionStart and selectionEnd). Click places the caret, drag or ⇧-click selects, double-click selects a word and triple-click a paragraph; ↑/↓ move by line, ⌘←/→ (Home/End) to the line's ends, ⌥←/→ by word, ⇧ extends; ⌘B/⌘I/⌘U bold, italicize, or underline the selected characters (at a caret, what is typed next); ⇧Enter is a line break within the paragraph. With characters selected the Type section and fills show and change only theirs, "Mixed" where they differ. ⌘Z undoes typing in bursts (a pause of a second starts a new step) and keeps editing. Escape or a press elsewhere on the canvas ends editing (the panels keep it); an emptied layer is removed.
  • Fonts: text keeps Figma's layout until edited; edited text is laid out in its own fonts, loaded first: fonts on this computer once permitted, else Google Fonts (the files for the text's scripts, cached by the browser), else Inter. A Missing fonts notice (fig-missing-fonts, rows fig-missing-font) lists fonts neither has, with Use fonts on this computer (fig-use-local-fonts, Chromium's Local Font Access permission prompt).
  • Keyboard, as in Figma: ⇧0 100%, ⇧1 fit, ⇧2 selection, ⌘/Ctrl +/−, N and ⇧N next/previous frame, PageDown/PageUp pages, Enter children, ⇧Enter and Esc parent, Tab/⇧Tab siblings, ⌘/Ctrl+A select all, ⇧R rulers, ⇧' pixel grid, ⌃G (Ctrl+⇧4 off macOS) layout grids, ⌘/Ctrl+Y outline view, ⌘/Ctrl+\ or ⌘/Ctrl+. hide UI, ⌥1 / ⌥2 / ⌥8 layers, assets, and design panels, ⌘/Ctrl+⇧C copy as PNG, ⌘/Ctrl+⇧E export. ⌘/Ctrl+P (or ⌘/) opens the design palette (fig-actions, Figma's actions menu, never the browser's print): type in fig-actions-input, arrows and Enter run an action (fig-action-<action>), Escape closes it. ⌘K stays the app's command menu; opened from a design it offers the palette in a hint row (command-menu-hint, "Did you mean to open the design palette?"), which ⌘P or a click opens. Editing: ⌘Z/⇧⌘Z undo and redo (each brings back what was selected around the step), ⌘D duplicate, ⌘C/⌘X/⌘V, Delete, arrows nudge (⇧ by 10), ⌘G group, ⇧⌘G ungroup, ⌥⌘G frame selection (like ⇧A's wrapping frame, it does not clip its layers), ⌘] / ⌘[ forward/backward, ⌥⌘] / ⌥⌘[ front/back, ⇧⌘H hide, ⇧⌘L lock, ⇧H / ⇧V flip, ⌘R rename, ⌥⇧U union, ⌥⇧S subtract (the top layers from the bottom one), ⌥⇧I intersect, ⌥⇧E (or ⌥⇧X) exclude, and ⌘E flatten (booleans and shapes into one vector layer). ⌥A / ⌥H / ⌥D align left, center, right and ⌥W / ⌥V / ⌥S top, middle, bottom; ⌃⌥H / ⌃⌥V (Ctrl+Alt off macOS) distribute three or more layers; digits set opacity (1 is 10%, 0 is 100%, two quick digits such as 4 5 make 45%); ⇧X swaps fill and stroke; ⇧⌘K places images; ⇧⌘R pastes to replace. A boolean keeps its layers as children and recomputes its shape when they change, and takes the bottom layer's fills and strokes.
  • App hotkeys over a design: while the design has focus its own keys come first, so R, T, O, H, C, \, ⌘\, ⇧H/⇧L, ⌘Z and the others act on the design rather than the app's rename, tag, leader keys, or new split. Keys the design does not use stay the app's (⌘K, ⌘S, ⌘;, G, /, E, M), and so do ⇧←/⇧→, ⇧H, ⇧⌘C, and Escape when there is no selection to act on (focus the next split, copy the link, leave a spotlighted split). A read-only design leaves the editing keys (R, T, O…) to the app.

The viewer has a browser fixture that needs no backend. From apps/web (build the engine first with just ensure-fig-engine-wasm):

bunx vite --config src/features/block-fig/browser-test/vite.config.ts
# http://127.0.0.1:3019/?file=showcase.fig

It opens files from crates/fig_engine/tests/fixtures (the synthetic showcase.fig, design-system.fig with components, a component set, properties, and styles, and prototype.fig, a clickable prototype), or from any directory named by FIG_CORPUS_DIR; the header also opens a local .fig. ?edit makes the file editable and ?new opens a blank design (editable); saves stay in memory, and ?reload reopens each one to check it round-trips. window.figFixture exposes engine(), saves(), errors(), notices(), downloads(), fontRequests(), and comments (the in-memory comment store: threads(), arrive() for someone else's comment, notified()), and ?present=<frame id> opens presenting; its font source serves the bundled Inter for every Google family and has "Nowhere Grotesk" installed locally, so font tests need no network. ?collab (with &people=alice,bob, the default) shows several people editing one file side by side (fig-person-<Name> holds each editor), each running the real shared-design session over an in-page sync server; window.figFixture.collab.people() gives each person's engine(), saves(), status(), and peers(). ?libraries keeps two designs in memory, the library "Design system" (design-system.fig, or &library=) and a blank "App", opened one at a time with fig-fixture-open-<id> (design-system, app; &open=app starts there); saves replace the stored design, and window.figFixture.libraries has open(id), current(), and reads(id). The Playwright suite runs with bunx playwright test --config src/features/block-fig/browser-test/playwright.config.ts (set PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH when the bundled browser is not installed). The sections above were verified on this fixture; the /app/fig route itself needs a backend with an uploaded .fig.

For import fidelity checks, save a local copy from Figma and open that same file in the fixture. Check the page names before switching pages, then use Find layers and ⇧2 to compare the same frame at a useful zoom in both apps. Include nested component logos, colors inherited from the page's variable mode, effects supplied by library styles, repeated pattern fills, and text that extends beyond a group's stored bounds. Compare exports as well as the canvas; an overview thumbnail can hide missing letters, wrong colors, or absent glows. Wait for the tiles to sharpen after zooming. The fixture preserves imported text outlines, but its substitute fonts do not verify the appearance of newly edited text.

Photoshop editor

Uploaded .psd and .psb files open in the psd block (/app/psd/<documentId>), and Photoshop file in the create menu (key H, also in project create menus) makes a new 1920 × 1080 document with a white Background. Both are behind the enable-psd-editor PostHog flag (on by default in development builds; ENABLE_PSD_EDITOR overrides it); with the flag off the block offers the file for download. The file is read, composited, and edited by the Rust psd_engine compiled to WebAssembly, in one worker. The canvas draws 512 px tiles from a pyramid of scales (the nearest one shows while zooming, then it sharpens) and redraws only what an edit changed. How Macro AI and search read these files is under Photoshop and Illustrator documents.

People with edit access get the editor; others get the same view read-only (no painting or editing tools, no editing shortcuts). Documents the editor cannot edit as they are (CMYK, Lab, Indexed, Bitmap, Duotone, Multichannel, 32-bit) open read-only with a bar (psd-notice, reason in psd-notice-readonly) whose Convert to RGB (psd-notice-convert, also Image > Mode > RGB Color) converts them to 8-bit RGB; the same bar lists parts of the file that could not be read (psd-notice-warnings), and psd-notice-dismiss hides it. Edits save automatically 1.5 s after the last change and when the tab is hidden, as a new document version through PUT /documents/{id}/simple_save; psd-save-state (data-state is saved, unsaved, saving, or error) shows where that stands.

Everyone with the document open edits it live, as in the Figma editor: changes go through the sync service as Loro maps of each changed layer's state and pixel tiles (see psd_engine::collab); the stored file stays the base everyone opens, and the first editor seeds the shared copy. Undo takes back only your own steps. Of the people editing, one (the lowest peer id) stores the merged file. Presence: other people's pointers with name tags (psd-peer-cursor, data-peer="<name>"), their selections' outlines in their colors, and their avatars at the top right of the canvas (psd-collaborators, one psd-collaborator per person); clicking an avatar follows their view (psd-following) until Stop. Without the sync service the document opens read-only, and for editors a bar (psd-session-notice) explains why, with Retry (psd-session-action); a file stored outside the session (an upload, an AI edit) turns open copies read-only with Reload.

The editor (psd-editor) is laid out as Photoshop is:

  • Menu bar (psd-menu-bar): File, Edit, Image, Layer, Select, Filter, and View (psd-menu-file … psd-menu-view); items carry ids such as psd-menu-new-layer, psd-menu-image-size, psd-menu-filter-gaussianBlur, psd-menu-adjust-levels, psd-menu-flatten, psd-menu-export-png, and psd-menu-download. Filters (Gaussian Blur, Unsharp Mask, Add Noise, Mosaic, Motion Blur) and Image > Adjustments open a dialog (psd-filter-dialog, psd-adjust-dialog) at the top right that previews on the canvas; OK keeps it as one step, Cancel or Escape drops it. Image Size (psd-image-size), Canvas Size (psd-canvas-size), Resolution, Feather (⇧F6), Expand, and Contract open small dialogs (*-ok, *-cancel).
  • Options bar (psd-options-bar, tool name in psd-options-tool): the tool's settings (brush size, hardness, opacity, flow, spacing, pressure; marquee feather; wand and bucket tolerance; gradient style and method; type size), and Fit Screen / 100% (psd-zoom-fit, psd-zoom-100). During Free Transform (⌘/Ctrl+T) and Crop it shows the box's size with commit and cancel (psd-options-commit, psd-options-cancel; Enter and Escape do the same).
  • Toolbar (psd-toolbar, psd-tool-<tool> with aria-pressed; clicking the tool in use again, or right-clicking, lists its group's other tools, psd-tool-option-<tool>): Move V (auto-selects the layer under the pointer; arrows nudge, ⇧ by 10), Marquee M (rectangle, ellipse; ⇧ adds, ⌥ subtracts, ⇧⌥ intersects), Lasso L (freehand, polygonal), Magic Wand W, Crop C, Eyedropper I, Brush and Pencil B ([ and ] size, digits opacity), Eraser E, Gradient and Paint Bucket G, Type T, Rectangle and Ellipse U (shape layers in the foreground color), Hand H (or hold Space), and Zoom Z (⌥ zooms out). ⇧ with a tool's key cycles its group. Below, the foreground and background swatches (psd-foreground-color, psd-background-color; clicking one opens the Color tab), X swaps them (psd-swap-colors) and D resets them (psd-default-colors).
  • Canvas (psd-canvas, with the zoom and size in psd-status, zoom in psd-zoom): wheel scrolls; pinch, ⌥-wheel, or ⌘/Ctrl-wheel zooms; ⌘0 fits, ⌘1 is 100%, ⌘+ and ⌘− step. Selections show marching ants; guides from the file are drawn. Pasting an image (⌘V) or dropping image files places them as new layers; ⌘C copies the selected pixels as PNG, ⇧⌘C copies merged.
  • Panels (psd-panels): tabs Properties and Color (psd-tab-properties, psd-tab-color) over the Layers panel. Properties (psd-properties-panel) shows the active layer's kind and bounds (psd-layer-bounds) and its settings: text (font, style, size, leading, tracking, color, alignment psd-text-align-<left|center|right>, faux styles, and Edit text… psd-text-edit), fill and shape layers (solid color or gradient with its style and method, the shape's stroke, psd-stroke-add), adjustment layers, masks (density, feather, enabled), and the layer style's effects on or off. Color (psd-color-panel) edits the foreground or background (psd-color-target-foreground, psd-color-target-background) with a saturation and brightness area, hue strip, hex, RGB, and HSB (psd-color-hex, psd-color-r, …).
  • Layers panel (psd-layers-panel): the active layer's blend mode (psd-blend-mode), Opacity and Fill (psd-layer-opacity, psd-layer-fill; drag their labels to scrub), and locks (psd-lock-transparency, psd-lock-pixels, psd-lock-position, psd-lock-all); then the layers top first (psd-layer-row, data-layer-id, data-layer-name), each with visibility (psd-layer-visibility), group disclosure (psd-layer-disclosure), the thumbnail (artboards show a frame, psd-layer-artboard, and Properties titles them Artboard), the mask (psd-layer-mask, with its thumbnail psd-layer-mask-thumbnail; clicking it makes painting edit the mask), and the layer style marker (psd-layer-effects, click toggles). Click chooses, ⌘/Ctrl-click adds, ⇧-click picks a range, double-click renames (psd-layer-rename), dragging reorders or moves into groups, and right-click (or psd-layer-menu) opens the layer's menu (duplicate, delete, group, clipping mask, masks, rasterize, merge down, export). The bottom buttons add a fill or adjustment layer (psd-new-adjustment), a mask (psd-add-mask), a group (psd-new-group), and a layer (psd-new-layer, ⇧⌘N), and delete (psd-delete-layer).

Typing: with Type, click the canvas for new text, or a text layer to edit it. A box under the text (psd-text-editing, field psd-text-input) takes the typing; each keystroke lays the layer out again on the canvas. New text becomes a layer at the first character, named after its first line (the name follows the text until someone renames the layer); Escape or ⌘/Ctrl+Enter finishes, and the whole typing is one undo step. Text left empty is removed. Other shortcuts: ⌘Z / ⇧⌘Z (Ctrl+Y), ⌘A, ⌘D, ⇧⌘I, ⌘J, ⇧⌘J, ⌘E, ⌘G, ⇧⌘G, ⌥⌘G (clipping mask), ⌥⌫ and ⌘⌫ fill with the foreground and background, ⌫ clears the selection or deletes the layer, and ⌘/ (or the keyboard button) lists them all (psd-shortcuts).

To exercise the editor without a backend, run the browser fixture from apps/web (build the engine first with just ensure-psd-engine-wasm):

bunx vite --config src/features/block-psd/browser-test/vite.config.ts
# http://127.0.0.1:3020/?new

?new opens a new document (&size=800x600 for another size), ?file= a file from the directory PSD_CORPUS_DIR names, ?readonly a viewer, and ?reload reopens each save to check it round-trips. ?collab (&people=alice,bob) puts several people side by side on one document through an in-page sync server; window.psdFixture exposes engine(), saves(), downloads(), errors(), notices(), and collab. Its Playwright suites run with bunx playwright test --config src/features/block-psd/browser-test/playwright.config.ts (PSD_BROWSER_PORT picks another port than 3020; set PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH when the bundled browser is not installed).

Illustrator editor

Uploaded .ai files open in the ai block (/app/ai/<documentId>), and Illustrator file in the create menu (key V, also in project create menus) makes a new one with a single 1920 × 1080 artboard. Both are behind the enable-ai-editor PostHog flag (on by default in development builds; ENABLE_AI_EDITOR overrides it); with the flag off the block offers the file for download. The file is read, rendered, and edited by the Rust ai_engine compiled to WebAssembly, in a primary worker plus up to three tile-raster helpers; the canvas composites tiles at the device pixel ratio over a gray pasteboard and redraws only what an edit changed.

People with edit access get the editor; others get the same view read-only (only the selection, direct selection, hand, and zoom tools; no editing keys). Edits save automatically 1.5 s after the last change, when the tab is hidden, and on close, as a new document version through PUT /documents/{id}/simple_save. Saving writes a standard PDF-based .ai that Illustrator and PDF readers open; an unedited file is written back byte for byte. Names given in the Layers and Artboards panels are saved with it (/MacroName marked content and a /MacroArtboard page key, which other readers ignore). For files Illustrator saved, a dismissible note (ai-illustrator-note) says that Illustrator-only editing data (live effects and the like) is not kept, and what the engine could not read faithfully is listed in ai-warnings.

Everyone with the file open edits it live. Changes go through the sync service as the Loro maps aiMeta, aiDoc, aiNodes, and aiImages (ai_engine::collab); the stored .ai stays the base everyone opens, and the first person who can edit seeds the shared copy. Each person creates objects in an id session (1–4095) picked at random among those the people present do not use, published in their presence. Undo takes back only your own changes. Of the people editing, one (the lowest peer id) stores the merged file, and the saved file becomes the new base. Presence: pointers with name tags (ai-peer-cursor, data-peer="<name>"), outlines of what others selected in their colors, and avatars at the top right (ai-collaborators, one ai-collaborator per person; a click shows their view). When the sync service is unreachable the file opens read-only with a bar (ai-session-notice, Retry is ai-session-action); a file stored outside the session (a new upload) starts the shared copy over, and people still in it get a Reload bar.

Layout and test hooks (root ai-editor, canvas ai-canvas):

  • Left panel: the main menu (ai-main-menu: ai-menu-download for the .ai, ai-menu-export-png and ai-menu-export-png-2x for the active artboard, then the view items) and the tabs ai-tab-layers and ai-tab-artboards.
  • Layers (ai-layers-panel, tree ai-layer-tree): rows (ai-layer-row, data-layer-id is the engine's node id, data-kind layer/group/clipGroup/path/text/image/artwork, aria-selected) list top-most first, as Illustrator does. A click selects (⇧ a range, ⌘/Ctrl toggles) and the canvas follows; a row's ai-layer-toggle expands it, layers show their color (ai-layer-color, a click cycles it), and hovering shows ai-layer-lock and ai-layer-visibility. Double-click renames (ai-layer-rename). Rows drag to restack, and into groups and other layers. ai-layer-new adds a layer on top; ai-layer-delete deletes the selection.
  • Artboards (ai-artboards-panel): ai-artboard-row per artboard (click fits it in the window, double-click renames in ai-artboard-rename, hover ai-artboard-delete) and ai-artboard-new.
  • Toolbar (ai-toolbar, floating on the canvas): ai-tool-<tool> for select (V), direct (A), pen (P), type (T), line (\), rectangle (M), ellipse (L), polygon, star, eyedropper (I), artboard (⇧O), hand (H), and zoom (Z); below them the fill and stroke new objects get, or the selection has (ai-swatch-fill, ai-swatch-stroke, each with a color picker and …-none), ai-swap-colors (⇧X), and ai-default-colors (D).
  • Properties (ai-properties, title ai-selection-title): with nothing selected, the active artboard (ai-artboard-section). Otherwise Transform (ai-field-x, -y, -w, -h measure outlines without strokes, as Illustrator does; ai-field-rotation, flips), Appearance (ai-fill-hex, ai-stroke-hex take hex or none, ai-field-stroke-width, ai-stroke-cap, ai-stroke-join, ai-field-dash, ai-field-opacity, ai-blend; gradients show as swatches without a hex field), Character and Paragraph for text (font picker, ai-font-style, ai-field-font-size, ai-field-tracking, ai-field-line-height, ai-text-align), Align (ai-align-left|center|right|top|middle|bottom, ai-distribute-horizontal|vertical), Pathfinder (ai-pathfinder-unite|subtract|intersect|exclude), and Quick actions (ai-arrange-front|forward|backward|back, ai-action-group, -ungroup, -clip, -release, -outline).
  • Status bar (ai-status-bar, bottom center): ai-undo, ai-redo, ai-save-state (data-state is saved, unsaved, saving, or error), and the zoom menu (ai-zoom-menu: ai-zoom-fit, ai-zoom-fit-all, ai-zoom-100, ai-outline-toggle).
  • Canvas gestures: scroll pans, ⌘/Ctrl or ⌥ + scroll and pinches zoom, Space or the middle button drags to pan. The selection tool selects (⇧ adds), draws marquees, moves (⇧ constrains, ⌥ drags a copy), scales by the eight handles (⇧ keeps proportions, ⌥ from the center), and rotates just outside a corner (⇧ by 45°); double-click enters a group or edits text. Direct selection drags anchor points and, once an anchor is chosen, its handles (⌥ breaks a smooth point). Shape tools drag out a box (⇧ square, ⌥ from the center; a click makes a 100 pt one). The pen clicks corners and drags smooth points, closes on its first point, and Enter or Escape ends an open path. The type tool clicks for point text, drags for area text, or clicks into text to edit it; typing happens in an overlay (ai-text-editing, input ai-text-input, ai-text-caret, ai-text-selection), and Escape or a click outside ends it. The eyedropper gives the selection the clicked object's fill and stroke. The artboard tool chooses, moves (with the artwork on it), resizes, and draws artboards; Delete deletes the chosen one. Image files dropped on the canvas or pasted are placed (at most 1000 pt long).
  • Shortcuts (⌘ is Ctrl off macOS): ⌘Z/⇧⌘Z, ⌘C/⌘X/⌘V (pastes 10 pt further each time), ⌘D duplicate, Delete, arrows nudge 1 pt (⇧ 10 pt), ⌘A / ⇧⌘A select all / none, ⌘G / ⇧⌘G group / ungroup, ⌘7 / ⌥⌘7 make / release clipping mask, ⌘] / ⌘[ forward / backward and ⇧⌘] / ⇧⌘[ front / back, ⇧⌘O create outlines, ⌘Y outline view, ⌘0 fit artboard, ⌥⌘0 fit all, ⌘1 actual size, ⌘+ / ⌘- zoom. While the editor has focus its keys come before the app's; keys it does not use, and Escape with nothing to deselect, stay the app's.

The editor has a browser fixture that needs no backend. From apps/web (build the engine first with just ensure-ai-engine-wasm):

bunx vite --config src/features/block-ai/browser-test/vite.config.ts
# http://127.0.0.1:3021/?sample

?new opens a blank document as the create menu makes it, ?sample a sample built with the engine (an 800 × 600 artboard with a red rectangle, a blue circle, and the text "Hello" on "Layer 1", and an empty layer "Notes"), and ?file=<name> a file from the directory named by AI_CORPUS_DIR (with a trailing slash); the header also opens a local .ai. Saves stay in memory (?reload reopens each one); ?readonly opens it as a viewer would. window.aiFixture exposes engine(), saves(), errors(), notices(), downloads(), fontRequests(), and rowsOf(bytes). ?collab (with &people=alice,bob, the default, and &sample or &file=) shows several people side by side (ai-person-<Name> holds each editor), each with the real shared-document session over an in-page sync server; window.aiFixture.collab has people() (each engine(), session(), saves(), status(), peers()), storeOutside(), reopen(name), setReachable(bool), and stored(). The Playwright suite runs with bunx playwright test --config src/features/block-ai/browser-test/playwright.config.ts (AI_BROWSER_PORT picks another port than 3021; set PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH when the bundled browser is not installed). The sections above were verified on this fixture; the /app/ai route itself needs a backend with an uploaded .ai.

Photoshop and Illustrator documents

Uploaded .psd and .psb files are Photoshop documents (file types psd and psb, the second Photoshop's large document format) and .ai files are Illustrator documents (ai). All three are design files in the vector association, beside Figma designs, so a Photoshop document is never handled as a picture, despite its image/vnd.adobe.photoshop media type: the image block and image attachments for the agent skip it. The backend reads these documents with the Rust psd_engine and ai_engine, the same engines the browser editors run as WebAssembly.

Macro AI reads them with ReadPhotoshopDocument and ReadIllustratorDocument (ReadContent returns the same description, and attaching one points the agent at its tool):

  • Photoshop: the canvas size, color mode, bit depth, and resolution, then the layer tree from top to bottom with each layer's kind (pixels, group, text, shape, fill, adjustment, or smart object), name, id, visibility, opacity, blend mode, position, and size, and the text of text layers with their font, size, and color (psd_engine::describe::outline).
  • Illustrator: the artboards with their names, positions, and sizes in points, then the layer tree from top to bottom with each object's kind, name, id, position, size, fill, and stroke, and the text of text objects with their fonts (ai_engine::describe::outline).

Descriptions stop at about 60,000 characters. The tools open files up to 300 MB; a Photoshop document whose layers decode to more than 512 MB of pixels is reported as too large to read, and Illustrator files saved by version 8 or earlier (PostScript rather than PDF) as unreadable. The tools read the stored file and cannot edit these documents.

Search indexes a document's layer and object names and the text of its text layers as one chunk, so content search finds a poster by its headline. Documents the engines cannot read (old or damaged files, or Photoshop documents beyond 256 MB of decoded pixels) are indexed by name only.

Create and type

Pasting a Macro /app/agents/<uuid> session URL into a Markdown editor converts it to an agent mention, just like the legacy /app/agent/<uuid> URL. Link query parameters are retained, except for referral codes.

Pasting a routed entity link such as /app/drive/md/<uuid> in an app editor creates the same document mention as a legacy link. Routed links retain the entity identity and compatible block targets (for example comment_id), but not the workspace path or pane-local search state. A copied multi-pane URL (/app/.../~/...) references its rightmost pane. If that pane has no supported entity, the URL remains an ordinary link. Project task comment targets, Home PRs, and agent chat links also convert to their respective entity mentions.

Pasting a bare GitHub pull request URL (github.com/<owner>/<repo>/pull/<n>, including subpages such as /files) inserts a link that becomes a PR mention once Macro finds the synced PR, usually within a moment. Hovering the mention shows its status, diff size, comments, and checks; clicking opens the PR in a split. A PR mention always stays on one line: a long title truncates with an ellipsis before the #N, so read the full title from the hover card. In a sent message, a mention an agent wrote by URL alone shows owner/repo#N until its title loads and keeps the same width either way. A PR Macro has not synced stays an ordinary link, and pasting over a selection keeps the default behavior.

Pasting inside a code block inserts the clipboard's plain text verbatim: links stay text instead of becoming mentions, long pastes do not collapse into a pasted chip, and Markdown syntax is kept. Pasted files still attach.

  1. Create → Document D. The app navigates to /app/md/<uuid> with the title field focused.
  2. type_text the title, then submitKey: "Enter" to drop into the body. Tab from the title also moves to the start of the body, without inserting a paragraph. While a document is open, Tab and Shift+Tab never switch the hosting view's tabs (e.g. Drive's My Files / Recent): opened from Drive, focus a non-editable control in the document, press Tab, and the document must stay open.
  3. Type paragraphs with plain type_text; use Enter between paragraphs. Do NOT use fill — the editor is contenteditable and fill does not work on it.
  4. The document auto-saves continuously (collaborative CRDT; no save button). The tab title and header update to the typed title.

On desktop, opening a document through Cmd/Ctrl+K focuses its body editor after the content and title have loaded. Creating a document focuses its empty title; if the new document already has a title, it focuses the body. Ordinary clicks and Back/Forward do not move focus into the editor. A control the user focused while the document loaded keeps focus.

The a11y snapshot exposes the entire body as the contenteditable's value and as paragraph nodes — use the snapshot itself to verify content. For formatting checks, run evaluate_script over [contenteditable] strong etc.

Body placeholder advertises: / for block commands, @ to reference files, ; for snippets. Markdown auto-format works while typing (# heading, [] checklist, > quote, 6. or 6. for an ordered list that starts at 6). Pasting multiple paragraphs inside a quote keeps those paragraphs quoted; see block formatting for verification. Imported Markdown recognizes bare URLs with the same shared matcher as messages; code and escaped URLs remain literal. Remove link persists through Markdown and editor-state saves. See message links for verification.

Every top-level checklist shows a completed count and progress meter on a thin line above its items, left-aligned with the checkboxes. Checklists containing an actual task mention additionally get a filter menu (status, priority, assignee, due date) that reveals on hover and stays while filters are active; filters dim unmatched items in place and persist with the document. Plain checkbox lists have no filter, and removing the last task mention removes it. On Android, use the software keyboard to check : emoji, / commands, ; snippets, and # tags where enabled. Each should open once and filter as you type. Tapping an emoji or command applies it; a second # closes the tags menu and leaves literal ## for Markdown headings.

AI text-writing operations require a paragraph/list-item or text-run ID. A table, row, cell, or list-container ID is rejected with guidance to choose a content block or use setCell. Existing stray inline content directly inside table cells is preserved in paragraphs when the editor opens the document. Bullet, numbered, and check lists stay in table cells; converting a cell paragraph to a list keeps the text.

@ opens the mention menu wherever the caret starts a word, including directly in front of existing text — the menu opens empty there instead of searching for the word ahead of the caret. It also opens immediately after a soft line break: type a word, press Shift+Enter, then type @ and a name to filter the menu. Typed inside a word (he@llo) it stays literal text. On Android, verify this with the software keyboard: tap @, type a name to filter, and tap a result to insert a single mention. The menu should remain visible above the keyboard while typing.

Ctrl+F / Cmd+F opens the in-document find bar. Matches include paragraph text and every inline mention chip (people, groups, tasks, docs, channels, skills, dates, tags, …) by the title shown on the chip.

On touch devices, the text-selection menu (Copy, Cut, Comment, Share, and other available actions) appears above the floating header, comment input, and bottom dock. It stays anchored to the selection while the document scrolls.

On a touch device, swipe a list item right to indent one level (Apple Notes style) or left to outdent. Nested children move with the parent. The first item can indent too, even in a single-item list. Vertical scrolling and taps are unchanged. Items stay still during the swipe and change indentation only when a successful swipe is released; short or blocked swipes leave them in place. Swiping requires permission to edit the document; comment-only access does not allow indentation changes. Losing edit permission during a swipe cancels it. To verify nesting, give a list item a child and grandchild, then swipe the parent right and left: all three should shift one level together, preserving their relative depths and order.

CRM company and contact mentions

With CRM enabled, type @ followed by a company name or domain in an editor or composer. Companies and contacts appear together under Contacts & Companies; People holds Macro users and groups only. Selecting a CRM contact inserts its record link, including when its email matches a Macro user. User-only pickers exclude CRM records. On mobile, the single list includes both CRM record types, while only Macro people and groups are pinned. With ENABLE_GRAPHQL_SOUP enabled, results include cached companies even if they are absent from the first 500 companies in the REST Quick Access feed; the REST feed remains a fallback. Cache search covers synchronized companies, not the entire CRM.

To verify, search for a cached company absent from that REST page, select it, and check that the inserted company mention points to the correct company. Also check searching by domain and that an open picker updates when companies finish hydrating. Discard unsent test drafts rather than sending them.

Project mentions

With Projects enabled, type @ followed by a project name in an editor, composer or spreadsheet cell. Projects (not folders) come from Quick Access, so they appear alongside documents and tasks in the Documents, Agents, & Tasks section and in entity property pickers that accept projects. The command menu searches the same Quick Access projects. Selecting one inserts a document mention with the project's icon and current name, like a channel mention. Clicking it or pressing Enter on it opens the project the way a task mention opens a task: in Tasks, under Projects › the project (on touch devices, as the project view on its own). A project you cannot read shows No Access. Pasting /app/initiative/<id> or a Tasks project link inserts the same mention.

The mention is stored as <m-document-mention>{"documentId":"<initiative id>","blockName":"initiative",…}</m-document-mention> (project is a folder). In a document it is tracked as a reference like other entity mentions. It is deliberately not a channel-message reference, so mentioning a project in a channel never shares the project with the channel's members.

To verify, mention a project in a document and in a channel draft, check the mention opens the right project, rename the project and reload to see the name update, and delete the mention. Discard unsent test drafts rather than sending them.

CRM pipeline mentions

Pipelines are not in the @ menu or the command menu. Pasting a pipeline link into an editor or composer inserts a document mention with a table icon and the pipeline's current name: the pipeline's copy-link action (/app/companies/pipeline/<id>), an older shared view link (/app/companies?crmView=… whose view is that pipeline) and /app/pipeline/<id> all work. Copied Customers links for other views (Companies, People, lists) stay plain links. A pipeline you cannot read shows No Access; a trashed one shows Deleted.

Clicking the mention, or pressing Enter on it, opens Customers on the pipeline at /app/companies/pipeline/<id> (on touch devices too), in the mention's split like other entity mentions. A split already showing that pipeline is brought forward instead.

The mention is stored as <m-document-mention>{"documentId":"<pipeline id>","blockName":"pipeline",…}</m-document-mention>. Like a project mention, it is never a channel-message reference, so mentioning a pipeline in a channel does not share it with the channel's members.

To verify, copy a pipeline's link in Customers, paste it into a document, check the chip's name, click it with Customers closed and again with Customers open on Companies, and open /app/pipeline/<id> directly.

CRM associations

With CRM enabled, any task, document or call can point at CRM records through the Companies and Contacts system properties: side panel Properties → Add property. Both pickers list Quick Access records: the team's companies and its most recently interacted contacts, filtered by name, domain or email. CRM contacts have their own Quick Access bucket, apart from people. They appear under Contacts & Companies in @ mentions and under People in the command menu. Values show the record's name and open the company or contact. An entity can carry the property without listing it (set at creation or through the API); adding that property pins the existing value rather than clearing it. Calls are linked automatically when they end, from their participants and the invitees of the calendar event carrying the meeting link; verify on a finished call's Properties.

Native offline reopening

On native mobile, previously opened Markdown documents/tasks can reopen after an app restart using their cached body and last-known permissions. Warm the document online first, then restart with API traffic blocked: verify the existing body, make a disposable edit, and restart offline again to check local recovery. Restoring connectivity must reauthorize synchronization before queued edits reach the server; verify the server copy, not just the still-cached editor text. Also reconnect after the initial sync's 10-second timeout: a reconnect snapshot must release queued edits without requiring the document to reopen. A document content-readiness timeout is retryable and must not revoke its cached open context; explicit access denial still does. The body should not wait for unrelated CRM metadata, references, duplicate-task suggestions, closed sharing/tag menus, or disabled mention queries. With those requests pending, the cached editor remains visible; optional information can appear when its own request finishes.

For a cold deep link or restored split, document loading waits for persisted user identity before capturing its offline session; a stalled auth request must not delay an identity already restored from IndexedDB. If no identity is cached, the normal auth query must succeed first. A previous logout marker is not a cached identity: after signing in again, it must trigger fresh authentication, not clear the new login cookie. Verify this restart/deep-link flow with a disposable account. Logout during the identity wait prevents the old load from opening under a subsequent login.

Cached open context is scoped to the signed-in user and invalidated at logout; permission tokens are never persisted. A missing body snapshot still requires an online open—metadata alone must not produce an editable empty document. This path does not imply offline coverage for PDFs, attachments, or other binary files.

Reference hover previews

The @ menu includes Recent agent sessions after Channels and before Companies. Search by session or persona name within the 500 most recently updated accessible sessions. Menu rows show the session title followed by a muted persona name, including @Cursor and @Macro for built-in personas. Names from the session API take precedence; older responses use the shared built-in name resolver or cached custom bots. Selecting one inserts an inline reference showing the shared agent icon and an underlined session title. Chips omit persona avatars and status. Click it (or select the node and press Enter) to open /app/agent/<id>. It references an existing session; it does not invoke the persona, attach its transcript to AI context, or grant access. Private and deleted sessions read Private agent session and Deleted agent session; clicking one (a tap on a touch device) opens a card that says why, as other unavailable references do. Mounted references refresh every 30 seconds while the tab is active to update titles and check access. Expanded session references are block-level Magic Chips with the same filled background as document cards. Click the card body to select its editor node (border and selection ring); use Open session to navigate. The collapse action returns the card to an inline session mention. Text before and after an expanded reference remains in separate paragraphs.

Hover a document reference chip to open its preview without navigating. With the preview open, the compact header shows a tinted icon, title, and author/time byline. Click the title to open the document; the Reference actions ellipsis menu contains copy link, split, embed/collapse, AI, and delete actions when applicable. In an editable editor it also offers Convert to Plain Text, which replaces the chip with the item's link as ordinary text, so sending it shares nothing; the No Access and Deleted cards show the menu there too. The preview stays open while this menu is active, and moving over other reference chips must not open their previews. Click outside or press Escape to dismiss the menu; other references can then be hovered again. Images use an inset frame and task chips appear below the header. Long titles wrap in place without a full-name tooltip.

A reference to something deleted reads Deleted, and one the viewer can't open reads No Access. Clicking either keeps its card open instead of doing nothing, and on a touch device a tap opens the same card as an opaque bottom sheet. The card names the item: This document was deleted or doesn't exist., This channel was deleted or doesn't exist., You don't have access to this event, and so on (a calendar event's reads This event was deleted or cancelled.). Skill, task and snippet references name their kind rather than "document". Database, form and routine references keep their own chip while the item opens, and switch to this Deleted or No Access chip and card once it doesn't; a database or form chip that loaded keeps its chip through a failed refresh unless the item is gone. Their detail reads answer non-owners 401/403 for a missing item too, so a deleted database or form reads No Access to anyone but its owner. A reference made in the last minute whose item hasn't reached the preview yet still opens it.

With ENABLE_GRAPHQL_SOUP enabled, the popup reuses the reference's live ItemPreviews batch, including task properties and viewer permission, without another fetch. Explicit refreshes may revalidate that batch, but requests must settle while the pointer stays over the same reference; cache updates must not cause a continuous fetch cascade. Adding or changing a direct sibling of a document reference must leave the reference mounted without flashing or restarting its preview query.

Embedded document cards

Document cards use a compact icon/title row and an actions menu. Full previews sit inside an inset surface; the author's display name and update time appear under the title as a byline. The plain 1rem icon sits in a column to the left of the title, aligned with its first line. Wrapped title lines, the byline, and task chips share the title's left edge. Full previews use the card's full content width with equal left and right insets. Title and byline share a text stack with a consistent 4px gap and 20px title leading, including when the title wraps. Titles and bylines use text-sm, differentiated by semibold and regular weight; smaller details use text-xs. Item.Icon provides the plain first-line-aligned icon slot. The small ellipsis button sits at the top right. Full embeds have a 320px minimum card height and a smaller rounded inset frame. The document-preview overlay uses the same plain icon, title/byline stack, small actions button, and task status control; image previews keep equal side insets. Metadata-only references omit the preview. Tasks replace the type icon with an icon-only status control; click it to change status when you have edit access. Priority and assignee chips remain below, without a duplicate status chip. Status and detail slots share one TaskPropertiesPreviewProvider per card: GraphQL preview data is reused, and REST fallback property/access queries are owned once, not separately by each slot. Non-task cards do not load task properties. Use the title to open the referenced document and the actions menu for the same applicable actions as a mention: Copy Link, Ask Macro for supported document types outside chat, Copy Branch Name for tasks, and Open in New Split when the reference is not already open. Editable cards also offer Convert to Inline Mention and Delete; these actions update when the editor becomes editable after loading. Converting back preserves the referenced document and its location parameters. Collapse/Expand Reference is only available on mentions; cards (including sent-message cards) never offer collapse. Title navigation preserves the reference's block parameters, including message, thread, annotation, and document locations. Click the card frame to select its editor node; controls and embedded content handle their own clicks. Full embeds remain vertically resizable and scroll inside the inset preview. When verifying, check a canvas embed, a metadata-only reference, and an editable task, including resize, menu actions, and keyboard access to the title and property controls.

AI edit

  1. Select text in the document, then click AI edit in the selection toolbar (on touch, Edit with AI in the selection toolbar opens the same prompt in a drawer).
  2. A focused prompt box appears (placeholder Describe changes). Type the instruction, press Enter (or click Send). The edit runs in fast mode and its changes apply client-side, so they land in the user's undo stack. Selected-text AI edit uses the shared Send button: muted and disabled for an empty or whitespace-only prompt, black with a white arrow in light mode once an instruction is entered (inverted for dark mode), on desktop and mobile.
  3. While running, the button row shows a Stop button (a11y text Stop AI edit), and the live cursor walking the text is labelled with the editor's name: Macro (AI) for an inline edit, or the name of the agent or persona whose session asked for the edit (e.g. Grunk (AI)). Edits stream directly into the document — there is no accept/reject step. The editor can insert the same @ mention chips a person can: dates/times, people, documents, channels, agent sessions (including the expanded Magic Chip card), and the other chip types.
  4. Completion signal: the Stop button disappears. Poll for that with evaluate_script; do not rely on wait_for text.

Comments (Discussion)

Below the editor: Discussion section with a Leave a comment... contenteditable. Desktop uses the same compact 15px composer as channels and AI chat, with an Attach images paperclip that opens the image picker directly. Wrapping text or Shift+Enter expands the editor above the controls. Lists, blockquotes, and other non-paragraph blocks always expand while editing. A single paragraph returns to the compact layout once it fits on one line. Touch keeps separate Attach images and Format buttons. Send comment is disabled until text exists. Click the composer, type_text, then click Send comment (Enter also submits). The comment renders above the composer with author + timestamp. @-mentions in comments notify the mentioned user. Editing a discussion comment keeps the attachment and send controls, with no trash button. Deleting a comment's first message deletes the whole discussion — the confirmation reads Delete comment, the replies under it go too, and an anchored comment's highlight clears from the document. Deleting a reply removes only that reply. On mobile, the new-comment composer is docked above the navigation bar, replacing New when commenting is available in documents and tasks. When the comment composer is unavailable, New remains available; these views have no floating Ask AI input. Tap Leave a comment... to expand the channel-style input; use Send comment to submit (Enter inserts a newline on mobile). Submitting clears and unfocuses the mobile input, returning it to its compact state and dismissing the keyboard. The compact input's paperclip opens the native photo library in the iOS app, with a file-picker fallback when unavailable; browsers use the file picker. Cancelling adds no images. While the main document editor is focused with the virtual keyboard open, the floating comment input is hidden; dismissing the keyboard or leaving the document editor restores it with any unsent draft intact. Comments remain in the Discussion section, and collapsing that section does not hide the docked composer. On touch devices, the Discussion section is hidden until it contains a comment; the floating Leave a comment... input remains available. If the discussion becomes empty again, the section disappears. Desktop keeps the empty section and inline input.

Select plain text or text already covered by a comment, then press Comment. The press keeps the editor selection until the draft opens. Cancel with Escape or the composer's close control, then select the same text and try again: the new draft highlight should clear on cancellation and the existing comment should keep its full range. Repeat with a selection spanning paragraphs and with a collaborator editing another part of the document. Draft marks stay local until the comment is posted. Posting first saves the highlight, then the message. Refuse an anchor update and verify that no message is posted, the input survives, and cancelling lets another comment open without changing document text. Document typing briefly pauses during anchor publication so pending text is flushed before the highlight is saved. Repeat posting and starting another comment with both edit and comment-only access; creating a document tests the owner path. For comment-only access, use a document no editor has opened since it was created: a commenter's session writes nothing to the document, so its comments still post. Also withhold a collaborator's update before posting: catch up its missing history before publishing the highlight, preserving both the remote text and the comment. Confirm the posted comment appears in a separate browser, since reopening the same client can restore an unsynced mark from its local snapshot. When messages and notifications exist but another browser has no inline highlights, check WAL acknowledgements and sync-service's comment-validation errors. Cancellation retries unconfirmed highlight cleanup after reconnect or reopening. An uncertain message response preserves its highlight rather than risking a saved comment becoming invisible.

Discussion and thread loading errors use compact muted text and a separate Retry button. Retry is disabled while fetching. A missing parent or thread shows an unavailable message without Retry; a deleted task cannot load comments until it is restored. Verify that an unavailable discussion hides its composer.

Comments anchored to selected text open in a floating margin card on desktop and a Comments drawer on touch devices. Hovering a desktop card reveals its actions without changing the card size or header text wrapping. New comments, replies, and edits use plain inputs on the card or drawer's background. On desktop, with the margin card focused: Enter opens a reply (same as the Reply action; Enter in the composer then sends), and Escape closes the reply composer, then the card. A press anywhere outside the comments and the document text (another split, the sidebar, the document header or empty page) returns a selected card and its highlighted text to their resting state; menus and dialogs opened from the card do not, and an unsent new comment keeps its composer. Clicking the commented text again selects the card again. The pinned reply keeps at least 16px of bottom clearance above the drawer's curve, including while the keyboard is open, and accounts for the home-indicator safe area when the keyboard is closed.

An inline-comment notification or /app/md/<id>?comment_id=<comment-id> link should scroll to the anchor and open its thread on the first click, including when the document has not loaded yet. Verify both comments arriving before the editor and comments arriving after it. Once loaded, click elsewhere in the document, then click the same notification again: it should revisit the comment, while background comment refreshes should leave the user's position alone.

Verify bottom Discussion links from Home too: /app/home/md/<id>?s0.drive.commentId=<comment-id>. Open a long document from another Home item with its caches cold. The first click should reach the comment even when its messages arrive before the editor or loading content interrupts the initial scroll. Repeat the click after scrolling away. Wheel, touch, keyboard input, or dragging the scrollbar during navigation should take control immediately; later background refreshes must not pull the reader back. If the editor never becomes ready, its pending comment jump expires after ten seconds; subsequent recovery must not revive it without a new navigation.

A document row keeps pointing at its newest comment notification until that notification is done, so on desktop the row opens the document at that comment. On touch the thread is a drawer over the whole page rather than a margin card, so a row tap opens the document itself: no drawer, with the comment's highlight still in the text. Verify this from Home, Inbox, and Files with a comment notification that has already been read — the drawer must not reappear on every open. Tapping the highlighted text still offers Show comment, and opening the notification itself still opens the thread.

When checking desktop margin placement, scroll a long document while an embed or image above the highlighted text changes height. Scroll anchoring may keep the text at the same screen position; its comment card should stay aligned through the resize, without briefly jumping upward or downward. Repeat with several rapid height changes, including while scrolling has paused.

Also verify anchored comments in Drive's detail pane: open a document with existing text anchors, then click a numbered comment badge to expand it. The document should stay visible and the thread should open; loading the document with its badges still collapsed does not exercise thread rendering. Comment copy links should retain the document/task route and the selected comment.

Document discussions

Document comments are messages read and written through /dss/messages/document/<id>, and both comment surfaces reuse the channel message components, as channels do.

Below the editor, expand Discussion to see comments without a text anchor. Its Leave a comment... composer is the channel composer: the + (Add) menu, formatting, mentions, and Send message (Enter also submits). Confirm completion by the new message appearing above the composer; a failed send retains the draft. The timeline initially loads a bounded page with up to three preview replies per thread. Expand a thread to load its replies; Load earlier comments pages backward. Live updates preserve unsent replies and edits while updating the surrounding thread.

Discussion and comment headers include the date for older messages (for example, Yesterday at 4:37 PM or 09/24/26 at 4:37 PM). Regular channel timelines retain their date dividers and time-only message headers.

When verifying @ mentions, compare the same person query in the document body and the Discussion composer: shared contacts use the same recent-interaction ranking. The desktop menu keeps up to three People results visible while other result categories load; use View all for the remaining matches. Check that a person stays clickable after document and email results arrive. Type and backspace through a query that keeps the same matches: existing rows should stay mounted and the menu should not collapse while cached results refresh. Check document titles and their order as well as People: type and backspace between a name's prefixes and verify the top document does not disappear and return while the result count briefly drops. Changing the total number of matches should not change a category's preview slots when it still has enough rows to fill them; use View all for the full list. Clear the unsent draft after testing.

Select text and choose the comment action to create an anchored comment. These threads appear beside their text in the margin (or in the active thread drawer on phones) and never in the bottom Discussion, including after live updates or reloads. Links to these threads open the margin without changing the bottom Discussion's timeline or expanding it. Existing highlights locate threads by their stable mark IDs. Replies, attachments, reactions, and editing use the same message controls as channels. In every floating comment card (the Markdown, PDF, and Word margins and expanded minimized threads), a long comment folds to a faded preview with Show more; Show less folds it again. A press outside the card does the same, including the press that closes an expanded minimized card. Presses on the card, its margin badge, and menus or dialogs the card opened leave the body expanded. In that card, Open session is the icon-only button from a narrow agent card, so the label does not take a line of its own. Channels keep the labeled chip. Short comments, the phone drawer, the bottom Discussion, and channels always show the full text. Markdown and Word margin cards stack without overlapping: a tall or expanded card pushes the cards below it down (past the end of a short document if needed), and no card starts above the top of the document. Check this with a long thread active near the bottom of a short document that has another thread above and below it. Removing the last marked text moves its retained conversation to Discussion, where it remains after reload. Removing only part of a marked range keeps the conversation anchored to the remaining text. On phones, the active Markdown thread opens in a drawer with a pinned reply composer; long-press any message for edit, delete, copy-link, and reaction actions.

A document thread carries no thread-level controls above it. Deleting the root message deletes the whole discussion, replies included, and answers with the root's tombstone. Copy link targets the specific comment with comment_id=<message id>. Previously copied numeric links still resolve under current document permissions. Deleting an anchored Markdown discussion removes its mark while preserving the document text and any overlapping comments. If deletion happens while the document is closed, its next editable view removes the retained mark when the document loads. Read-only viewers see plain text without a dead comment highlight; the stored document and overlapping live comments stay intact.

PDF comment threads in the right margin use the channel composer (Leave a comment..., Enter sends) and the message thread controls. Highlight comments come from selecting text and choosing the comment button in the selection menu; placeable comments come from the toolbar Comment tool and a click on the page. Discussions read and post through /dss/messages/document/<id>; anchor geometry still loads from /dss/annotations/anchors/document/<id>. Deleting a highlight's discussion keeps the highlight as a plain highlight; deleting a placeable's discussion removes the placeable. An anchor bound only to a legacy annotation thread that was never imported stays hidden rather than shown as a bare highlight.

PDF signatures

Choose Signature in the PDF toolbar, then click the page where it should go. The Draw signature dialog opens over the active PDF. Drag across the canvas with a mouse, pen, or finger; Use signature becomes available after drawing. Confirm to place the signature, then drag it to adjust its position. Cancel signature, Escape, or clicking outside discards the unfinished placement. Signatures are saved into that PDF through its normal save flow; reusable saved signatures are not available yet.

When verifying this flow, draw near the right and bottom edges as well as the center. Check at desktop and phone widths and at 2x pixel density; strokes should stay under the pointer, survive resizing, and appear on the PDF after confirming. Reopen the dialog and verify it stays centered during its opening animation and starts with an empty canvas. Cancelling must leave no empty signature box.

Word (DOCX) editor

When enable-docx-editor is on, uploaded .docx files open at /app/write/<id> in an editor instead of the PDF preview. The flag is a PostHog flag that is on by default in dev mode; set VITE_ENABLE_DOCX_EDITOR to override it locally. The header label has a Beta badge. Pages are laid out and drawn by Macro's own DOCX engine (Rust compiled to wasm, in a worker) as <canvas> sheets with Word's pagination, so the document's text is not in the DOM. Edits sync through the sync service as you type: every open copy updates live and shows each collaborator's caret with their name ([data-docx-peer]).

  • Open a newly uploaded DOCX while it is still processing. The view stays loading and retries the original export once per second for up to 30 seconds, then renders as soon as extraction finishes without needing a reopen. If processing takes longer, it asks you to try opening it again shortly. A first-open sync /exists 404 is expected: the editor seeds collaboration from the stored upload. The export endpoint returns 409 while parts are missing and never publishes an empty archive.
  • Pages are [data-docx-page="<index>"] elements. Click a page to place the caret, drag to select, double-click for a word and triple-click for a paragraph. On a touch screen a swipe scrolls, a tap places the caret, and a double tap or a held press selects a word. Keystrokes go to a hidden textarea, [data-docx-input] (labelled Document text); it must have focus, which a click on a page gives it. Read text back from another tab or after a download, not from the page.
  • Editors get a toolbar labelled Document formatting, docked full width under the header (like the spreadsheet toolbar): Undo, Redo, the Paragraph style, Font and Font size selects, Bold, Italic, Underline, Strikethrough, Superscript, Subscript, the Text color and Highlight menus ([data-docx-menu="color"], [data-docx-menu="highlight"]), Clear formatting, Bulleted list, Numbered list, the alignment buttons, Decrease indent, Increase indent, the Line spacing menu, Insert table (inside a table also the Table rows and columns menu, [data-docx-menu="table"], to insert or delete rows and columns or the table), the Insert footnote or endnote menu ([data-docx-menu="notes"], in the body), Track changes, Hide tracked changes / Show tracked changes, Comment on selection, Find and replace (viewers get Find and replace). The header's three-dot file menu has Download on desktop and mobile, which saves the live .docx, or the server's export before the editor has loaded. While tracking is on (or the caret is on a tracked change) it also shows Accept change, Reject change, Accept all changes and Reject all changes.
  • The browser's own find cannot see canvas text, so Mod+F in the document (or Find and replace) opens the editor's find bar ([data-docx-find]) in the top right; Ctrl+H (Cmd+Shift+H on a Mac) opens it with the replace field. The Find in document field ([data-docx-find-query]) searches as you type and shows <n> of <total> ([data-docx-find-status]); matches are highlighted on the pages ([data-docx-find-match]). Enter and Shift+Enter (or the arrow buttons, or Mod+G) move between matches and select them; Match case and Whole words only narrow the search. Straight and curly quotes match each other. The Replace toggle shows Replace with ([data-docx-find-replacement]) with Replace ([data-docx-replace]) and Replace all ([data-docx-replace-all]); replacements follow tracked changes and one undo takes back a replace all. Escape closes the bar with the current match selected.
  • Arabic and Hebrew paragraphs lay out right to left as in Word (joined Arabic letters, mixed-direction lines in visual order); the left and right arrow keys move left and right on the page.
  • Mod+Z, Mod+Shift+Z and Ctrl+Y (or the toolbar buttons) undo and redo your own edits only, never a collaborator's. Mod+B/I/U format, Tab and Shift+Tab indent list items, Enter splits paragraphs and Shift+Enter inserts a line break.
  • Track changes turns tracking on for the whole document (it is saved in the file, as in Word): every editor's typing then shows as an underlined insertion and deletions stay visible struck through, each under its author's name. Formatting changes (bold, alignment, lists, indents) are recorded too: the text looks formatted, and Accept and Reject appear when the caret is in it. A thin bar in the left margin marks every line that holds a change. Accept and reject act on the selection, the change at the caret, or every change.
  • Double-click a page's header or footer area to edit it. The body dims, the area gets a dashed edge and a Header (or Footer) label with a Close button; Escape or a click on the body returns to the body. Header and footer edits reach collaborators and the download like body edits.
  • Click a footnote or endnote at the bottom of the page (or after the body) to type in it, as in Word; nothing dims. Escape or a click on the body returns to the body. Comments stay with the body text.
  • To add a footnote or endnote at the caret, use the toolbar's asterisk menu (Insert footnote or endnote → Footnote / Endnote) or Word's shortcuts (Ctrl+Alt+F / Ctrl+Alt+D; Cmd+Option+F / Cmd+Option+E on a Mac). The number appears in the text and the caret moves into the new note at the foot of the page (endnotes go after the body).
  • Clicking a DOCX in the Home list opens the editor in the Home preview pane. Viewers and commenters see the same paginated pages, read-only.
  • To comment on any text, including table cells: select it, then click the floating Comment button beside the selection ([data-docx-comment-button]). You can also use the toolbar Comment on selection button or Mod+Alt+M. The draft opens a thread card in the right margin. Posting creates a normal document discussion (markdown anchor with mark_id), so it also appears in channels and notifications. Commented text is highlighted by overlay elements ([data-docx-comment-highlight]) above the canvas. Comments anchor in the body only, not in headers or footers.
  • Comments written in Word (stored in the file) show in the same margin as read-only cards ([data-docx-word-comment="<id>"]): author, date, text and replies, marked In the document (and Resolved when done). Their text is highlighted too ([data-docx-comment-highlight="word:<id>"]). They stay in the file on download; reply with a Macro comment.
  • Threads whose text was deleted are listed under Comments on text that has changed, above the Discussion composer.
  • AI CommentOnText works on DOCX by quote. The editor pins each quote to the first matching text the next time someone opens the file.
  • AI EditWordDocument can record its edits as tracked changes (trackChanges, or by default when the document tracks changes) and add Word comments (addComment), both attributed to the requesting user by name rather than to the agent. Its comments are written into the file, so they appear as In the document cards and travel with the download.
  • Download exports the current collaborative state with every edit, including headers, footers and tracked changes. Macro comment threads are not written into the file; Word comments (including the agent's addComment ones) are.
  • The stored upload is not rewritten yet. Search, the PDF export and AI ReadContent still see the original file.
  • Ask Macro, in the header beside Share, opens the agent composer in a split with the document mentioned (nothing sends automatically).
  • AI ReadWordDocument and EditWordDocument read and edit the live copy, so open editors patch agent edits in as they land. ReadWordDocument lists every paragraph, table cell and content control with its id. EditWordDocument applies an atomic batch of replaceText, setText, formatText, insertParagraph, setStyle and delete operations to those ids. A DOCX nobody has opened yet (an upload or a duplicate) gets its live copy from the first tool call, seeded exactly as the editor seeds it, so nobody has to open it first. EditDocument still rejects DOCX.

Document history

On desktop, open the title's file menu (…) and choose History. This opens an overlay filling the current document block, with a read-only version preview on the left and a timeline graph plus sessions on the right. History is no longer a side-panel section and its file-menu item is hidden on mobile. Scrub the graph to preview a point in time, or select a session to see its changes. Current version returns the preview to the live version; Fork copies the selected version into a separate document. Close history or Escape returns to the mounted editor without losing its scroll position. The two columns scroll independently, and other app splits remain available.

Sessions, diffs, and blame tooltips name AI edits by who made them: an agent's edits show the agent's name, and other AI-editor writes show <user> (AI) for the user they were made for.

Side panel

Right side of a doc (toggle with Hide/Show Side Panel):

  • Actions → Ask Macro (opens a doc-scoped AI chat, see ai-chat.md), and on desktop a Copy as prompt pill for documents and tasks. Its primary button runs the last-used agent action; the Agent options caret lists Copy as prompt, MCP setup instructions, and an Open in group (Claude Code Web, Codex Desktop, Cursor, Zed). Tasks add Copy branch name. A document prompt wraps the title and markdown in <document> / <document-content> tags with no branch instructions; the toast reads Prompt copied to clipboard. The Files-view title menu (three dots beside the breadcrumb) also lists Copy as prompt directly above Download for plain documents.
  • Details → Owner, Created, Last updated.
  • Tags → Add tags (dialog). Click a tag's label to toggle and save; Shift-click keeps the picker open for multiple selections. Reopen it to remove a tag. With GraphQL Soup enabled, applying the first tag refetches only that entity's properties; later edits update the existing assignment optimistically. Verify both the side panel and list-row chip, then reload to confirm persistence. This must also work after background backfills populate more than 128 cached Soup variants—tag saves must not scan all cached pages. Properties → Add property.
  • Activity is collapsible; document statistics and ownership timestamps appear in the footer.
  • Activity lists the same glyph-rail lines as /app/component/activity (plain glyphs on a thin connector, one line each with long names truncated, compact 17h / 8d / 1mo times; consecutive edits fold into one made 3 edits line). Past four entries it shows the three newest, a View all activities toggle row (dotted connector, caret glyph), and the oldest fetched entry (usually created this) pinned last; the toggle flips to Show less once expanded. Human content edits appear after the next sync flush. Continued editing creates one Activity event until that editor has been inactive for five minutes; their next edit then starts another event. Opening a document alone creates no edit.
  • Header: Share, Copy Share Link, overflow menu — use Share to inspect or change the doc's visibility/permissions. Documents, AI chats, and folders have a Team access control (None / View / Comment / Edit) for sharing directly with the owner's team. That is independent of the team-scoped link control. Folders hide link sharing, so Team access is its own card on desktop and a Team tab on mobile, not nested in the Link card.

Known failure: "expected instance of LoroDoc"

Opening any doc can crash with a full-screen dialog expected instance of LoroDoc (console: [observability] expected instance of LoroDoc). Seen after the Vite dev server reconnects (HMR leaves two copies of the loro wasm module alive). Try Again and a normal reload do NOT fix it; a hard reload ignoring cache (navigate_page with ignoreCache: true) does.

AI can create a native workbook without an open editor using CreateDocument with fileExtension: "spreadsheet", empty fileContent, and isTask: false. Read the returned document with ReadSpreadsheet, then populate it with EditSpreadsheet; do not create a CSV as a substitute for a native workbook. Spreadsheet reads and edits run against server state even when no tab is open. An edit uses the revision from a fresh read and atomically applies a CRDT delta that is broadcast to connected collaborators. A stale revision is rejected: reread and reconsider the change instead of blindly retrying. Unsynced edits still follow normal CRDT collaboration semantics when they reconnect.

HTML preview tabs

HTML documents show Render and Code inset tabs in the top bar in both standalone desktop blocks and Drive detail views. Switch to Code to inspect or edit the source, then back to Render to see the current preview. Other code file types do not show these tabs.

Large-document undo checks

In a disposable Markdown document, change several paragraphs in one edit, then undo and redo. Verify the text, paragraph count, and a second peer agree after each operation; wait for saving to finish and reload to check persistence.

Undo timing has a backend-free browser regression over a 4,000-paragraph document (about the length of Moby-Dick): undoing typing and undoing the deletion of a thousand paragraphs each stay within a time budget. bunx playwright test --config src/lib/core/component/LexicalMarkdown/collaboration/browser-test/playwright.config.ts from apps/web. In traces, editor.undo / editor.redo carry document.id and editor.node_count; the editor.history.loro child covers Loro's undo and the mirror read, and editor.history.apply lasts until the reconciled editor update commits.

Task mention status

Task @mention pills and @mention menu rows display the same circular completion/status icon as task rows; pills do not repeat it after the title. The icon updates in place as preview data changes; unknown status uses an empty circle, and subtype-only previews use their completion flag. See tasks for states and the browser regression.

Sharing labels describe their audience: Only people with access, Anyone with the link, and Team members with the link. Permission choices say Can view, Can comment, and Can edit. Copying a restricted link still does not grant access. The existing immediately opening dialog and its deferred loading remain unchanged.

Drive's existing New controls open a wide menu with icons and short descriptions for creation and uploads, while preserving the current destination and feature gates.

Opening a row clears its unread state immediately; direct details mark read when ready. See reading items for failure, keyboard, mobile, and email inbox checks.