For AI agents: the complete documentation index is available at https://a3s-lab.github.io/Office/docs/en/llms.txt, the full documentation bundle is available at https://a3s-lab.github.io/Office/docs/en/llms-full.txt, and this page is available as Markdown at https://a3s-lab.github.io/Office/docs/en/editor-quality-roadmap.md.
A3S Office
Playground
  • English
  • latest

      • Editor Quality Roadmap

      This roadmap turns A3S Office from a broad editor demo into a dependable component product. Delivery is depth-first: finish the workflows users repeat every day, prove native-file compatibility, and only then expand the command surface.

      The technical ownership model, Worker/WASM boundaries, and long-term engine stages remain defined in Browser editor architecture. This document owns product priority, quality gates, and the order in which capabilities are made release-ready.

      Product Principles

      1. Depth before breadth. A complete editing workflow is more valuable than another toolbar control without reliable import, undo, save, and export.
      2. Word first. Document editing is the first quality track because it exercises layout, styles, tables, media, review, and long-document behavior together.
      3. Native files are the acceptance test. A feature is not complete until its supported semantics survive import, edit, export, and reopen in Microsoft Office and WPS.
      4. Controlled state stays predictable. One user intent emits one host update and one undo record. External content replacement must not corrupt selection, history, or editor state.
      5. Public APIs describe stable product contracts. Internal controllers are not exported merely to make an integration possible.
      6. Documentation ships with the capability. Every public Prop, event, content field, and Extension boundary must be documented and tested.

      Release Gates Shared by Every Editor

      These are foundations, not a separate feature-expansion phase:

      • No browser-native prompt, confirm, context menu, or select control in a primary workflow.
      • Shared dialog, popover, menu, ribbon, status, save, loading, empty, and error states use the same placement and interaction rules.
      • Keyboard-only users can reach and leave every primary surface. Focus returns to the invoking control when a temporary surface closes.
      • React, Vue, and Web Component adapters expose the same supported behavior, with framework-specific names documented where they differ.
      • Controlled updates, undo, redo, import, export, and read-only preview have focused tests.
      • Desktop and compact visual checks cover normal, empty, loading, error, menu, dialog, and long-content states.
      • Performance budgets are fixture-based. A release cannot replace measured evidence with a qualitative claim.

      Current shared-shell evidence includes one modal focus boundary for the compact Office sidebar, Presentation slide drawer, AI assistant, and editor dialogs. Each surface receives focus on entry, wraps forward and reverse Tab navigation, isolates obscured content with inert, lets only the top overlay handle Escape, and restores the current semantic invoker even when React remounts that control. The Office shell keeps persistent desktop navigation and the temporary compact drawer as separate states. Crossing into the compact breakpoint therefore closes the drawer in the same render without obscuring the workspace or moving its active focus, while returning to desktop restores the user's persistent sidebar preference. Compact Word navigation, find, layout, citations, changes, and comments task panes now use the same modal boundary while retaining persistent side-pane behavior on desktop. Phone, 768 px, and desktop browser checks keep overlay and persistent-pane behavior distinct. Shared Office color palettes now replace their dense desktop grid with eight larger columns on phones; spatial keyboard movement follows the rendered column count across theme and standard-color sections. Shared editor context menus remain pointer- or selection-anchored on desktop, but become safe-area-aware bottom action sheets below 640 px. Their actions are at least 44 px tall, long menus scroll inside a bounded surface, Escape returns focus to the editing target, and deterministic Spreadsheet evidence captures the phone layout, accessibility tree, console, and page errors.

      Priority 1: Word-Compatible Document Editing

      Document work is delivered as complete vertical slices in the order below. New ribbon commands wait unless they are required by the current slice.

      1. DOCX Round-Trip Foundation

      • Move package relationships, styles, numbering, themes, comments, headers, footers, notes, fields, and media toward loss-preserving browser state.
      • Preserve unsupported package parts and attributes instead of regenerating them from visible HTML.
      • Build small, reviewable fixtures for every supported semantic and include documents produced by both Microsoft Word and WPS.
      • Report unsupported or normalized content explicitly before export.

      Exit evidence:

      • Import-edit-export-reopen passes in Word and WPS for the declared fixture matrix.
      • Unedited supported files retain package relationships and unsupported parts.
      • A failed conversion does not overwrite the source file or emit a fidelity-success claim.

      2. Deterministic Pagination and Text Layout

      • Resolve style inheritance, numbering, paragraph spacing, indents, tabs, bidirectional text, columns, section breaks, widow/orphan rules, and exact registered fonts before line layout.
      • Reflow from the earliest affected block and reuse stable earlier pages.
      • Keep page breaks as mapped editor decorations so layout never becomes editable document content.
      • Define font substitution and missing-glyph behavior that matches export.

      Current product evidence includes a responsive Page Setup task pane with keyboard-operated Page, Columns and Sections, and Header and Footer tabs. The default view contains only paper and margin controls; page-chrome editors mount on demand so compact and desktop workflows do not begin with one mixed, unbounded settings stack. Header and footer formatting actions are divided into balanced text-formatting and alignment/insertion rows instead of leaving orphaned controls on an accidental wrap line; a deterministic phone workflow now covers all three tabs, a live column update, modal isolation, and invoker focus restoration. Paragraph-spacing and paragraph-pagination popovers now expand to a viewport-safe 320 px surface on phones. Their numeric fields, steppers, checkboxes, and reset actions provide 44 px touch targets while preserving direct document commits and exact trigger-focus restoration. A deterministic phone workflow proves both semantic changes and captures accessibility, console, and page-error evidence. Deterministic text collection now follows CSS font weight matching while retaining exact family and normal/italic style boundaries, keeping intermediate and synthesized bold weights on the WASM layout path when a compatible registered face exists. Editing, read-only preview, and PDF composition now share the same base typography and structural content styles. Header and footer content occupies the physical page margins without shifting body flow, while empty headers remain absent and PDF output does not synthesize a filename header. Compact previews keep the same physical page geometry and use bounded scrolling instead of changing margins and text flow to fit the viewport.

      Editing and read-only preview now retain one canonical TipTap surface and the same Worker/WASM pagination result, so switching mode preserves shaped runs, automatic page breaks, table fragments, and the computed page count. Page chrome is treated as an overlay inside the physical margins rather than an additional body-height deduction. Browser PDF export finds this mounted surface by stable artifact ID, clones it without editing state, and crops physical pages from bounded batches. The three browser rendering paths therefore share one page-layout result; searchable text and vector PDF output remain separate fidelity work.

      Exit evidence:

      • Deterministic page goldens cover Latin, CJK, Arabic, Hebrew, lists, sections, headers, footers, and mixed formatting.
      • Typing and selection do not create main-thread tasks above the documented budget on long-document fixtures.
      • Pagination does not create extra undo entries or move a stable selection.

      3. Styles, Lists, and Tables

      • Make named styles, direct formatting, numbering, restart/continue behavior, nested lists, and theme values round-trip together.
      • Complete table insertion, selection, row/column/cell operations, merged cells, repeating headers, nested tables, row splitting, borders, shading, alignment, and sizing.
      • Ensure contextual commands operate on typed selections, including multi-cell and keyboard selection.

      Current implementation evidence includes the responsive paragraph-style gallery and a typed list slice with disc/circle/square bullets, five common ordered formats, restart/continue/start commands, keyboard-operated galleries, single-transaction undo behavior, and DOCX round trips for the declared list styles. Compact ribbon overflow now reserves both navigation edges so tool width does not change after paging. Phone list galleries stay within the viewport and provide 44 px clear, restart, start-value, stepper, and commit controls. Closing a list command restores the TipTap selection as well as DOM focus, so reopening the numbering gallery retains its active style and start settings. A deterministic phone workflow covers two ribbon advances, square bullets, upper-Roman numbering, a custom start value, editor-focus recovery, and Escape-to-invoker restoration. Tables now expose separate Design and Layout contextual tabs, five keyboard-operated style presets, multi-cell fill, and a reusable border pen with all, outside, inside, side, inside-horizontal, and inside-vertical targets. Horizontal and vertical alignment, exact column- width and row-height fields, equal row/column distribution, content/window autofit, row/column operations, merge/split, repeat-header, and non-splitting row controls are also available. Common cell shading, vertical alignment, explicit per-edge borders, fixed grid widths, layout mode, and explicit row heights round-trip through editable HTML and DOCX. Table geometry now models the layout algorithm, auto/percentage/pixel preferred width, alignment, indent, table-level cell margins, and cell-level margin overrides independently. tblLayout, tblW, jc, tblInd, tblCellMar, tcMar, tblGrid, and tcW round-trip without treating grid data as the preferred table width. A real fixture proves that a centered 62.5% table with a first-cell margin override keeps matching edit and preview geometry. Table insertion or removal does not reinterpret an imported autofit table as fixed merely because it has grid widths. The Layout ribbon reports the rendered column width for autofit tables and labels noncanonical preferred widths as Current width instead of misreporting them as fixed. The Table Properties dialog now owns four coherent tabs: Table, Row, Column, and Cell. It combines auto, percentage, or centimeter preferred width, left/center/right placement, and left indent with selected-row height and rule, page splitting, repeated headings, current-column width, selected-cell vertical alignment, and four-edge margins. The complete validated draft commits in one TipTap transaction and one undo record; Cancel and Escape preserve the last committed state and restore the exact ribbon invoker. Values the user did not edit retain their exact imported pixels and partial inheritance instead of round-tripping through the two-decimal centimeter display. The deterministic styled-table workflow checks truthful 62.5% centered import values, all four tabs and touch-sized controls at 390 px, an atomic change to 72% left placement with a 0.5 cm indent, a 1.2 cm exact row, disabled page splitting and repeated headings, a 4 cm column, centered cell content, a 0.4 cm left cell margin, cancellation safety, exact untouched 8 px and 16 px margin edges, and matching preview geometry with clean browser diagnostics. Table-level outer and inside DOCX borders resolve to cell edges on import without flattening mixed styles. Direct table and cell theme colors now resolve through the package theme, including themeTint, themeShade, themeFillTint, and themeFillShade. The resolved colors remain identical in edit and preview and export as stable RGB values, instead of falling back to stale OOXML color attributes. DOCX table styles now resolve the default or referenced table style through a bounded, cycle-safe basedOn chain. tblLook flags or bitmasks, row and column band sizes, grid spans, and row grid offsets select whole-table, banded row or column, first/last row or column, and corner-cell conditions in Word precedence order. Conditional fills, per-edge borders, fonts, colors, bold, italic, underline, and strikethrough enter the same edit and preview model before direct table, cell, paragraph, and run formatting. Conditional paragraph alignment, direction, indents, spacing and line rules, pagination rules, and tab stops follow that same precedence. Theme-derived style colors are materialized as stable RGB on export. A real conditional-style DOCX fixture is covered by deterministic A3S Test evidence for header, body-band, last-row, and paragraph-layout presentation with a clean browser diagnostic run. Splittable rows taller than a complete continuation-page body now fragment at measured paragraph boundaries instead of overflowing as one atomic row. The same layout path repeats leading heading rows on each continuation page. A real DOCX fixture with one 120-paragraph body row deterministically produces three physical pages; A3S Test verifies the page-two and page-three cell breaks and repeated headings, keyboard access to the final marker and following paragraph, and an error-free browser run. Desktop insertion keeps an 8 × 10 keyboard matrix without duplicate cell semantics, while phone insertion uses focused row and column controls with 44 px touch targets. A deterministic phone workflow now proves 3 × 3 insertion, row extension, compact-ribbon deletion, and editor-focus recovery. Native Word list identities, arbitrary multilevel numbering, less-common conditional table- style paragraph properties outside the supported layout set, loss-preserving semantic theme references, percentage-width column authoring, nested tables, and more complex merged-cell multi-page row cases remain part of this priority rather than being treated as complete.

      Exit evidence:

      • Formatting and table operations have command, interaction, undo, DOCX, and visual tests.
      • Large tables paginate without duplicate content, selection loss, or quadratic reflow.

      4. Media, Fields, and Page Objects

      • Complete inline and floating images, wrapping, anchors, captions, links, bookmarks, cross-references, page numbers, dates, fields, footnotes, and endnotes.
      • Keep object identity stable across move, copy, delete, undo, and DOCX relationship updates.
      • Make image and object dialogs validate size, placement, alternative text, and destructive replacement before commit.

      Current implementation evidence includes one project-owned file materialization boundary shared by Word body images, header/footer images, Presentation images, and imported Office/PDF files. File inputs can reset immediately for same-file reselection without invalidating an in-flight read. Unit coverage verifies single-read ownership and the A3S Test picture fixture verifies that a relative-path PNG reaches the Word canvas and opens its contextual picture toolbar without a false read error. The phone picture workflow now exercises one Picture Properties dialog for centimeter width and height, per-image aspect-ratio locking, square wrapping, alignment, text distance, and alternative text. The complete draft commits as one undo entry, untouched imported dimensions keep their exact source values, the selected image remains active, and apply, cancel, and Escape restore the exact Picture Properties command so the next keyboard action cannot mutate document content. Caption order and cross-reference validity now share one live transaction graph: deleting or reordering a caption renumbers surviving targets, updates every linked field, and renders dangling references as an explicit missing state. Caption numbers are included in the accessible caption name, and a deterministic phone workflow proves dialog focus, insertion, deletion, and the missing-reference transition.

      Exit evidence:

      • Relationship-bearing objects reopen with working targets in Word and WPS.
      • Object layout has deterministic fixtures at page and section boundaries.

      5. Review and Long-Document Quality

      • Finish anchored comments, replies, resolve/reopen, tracked insert/delete, accept/reject, citations, navigation, find/replace, outline, and references.
      • Define conflict behavior when a controlled host update changes a reviewed range.
      • Keep outlines, search results, revision review, and comment review on bounded windows, and incrementally derive page chrome for large files.

      Current implementation evidence includes a persistent Word-style navigation pane with a typed heading hierarchy, active-heading tracking, collapsible branches, keyboard traversal, and responsive left-side placement. Its search returns full-text results with section context, bounded excerpts, match highlights, and safe selection-based jumps that do not create history entries. Compact result selection closes the modal pane before restoring the exact body selection and focus. The same pane now offers a page view with real rasters cropped from the live paginated surface, physical-page and restarted-page-number labels, active-page tracking, arrow/Home/End keyboard traversal, and selection-safe jumps to each page. The current and adjacent pages enter a serialized capture queue first; the rest use bounded viewport admission, debounced source-mutation refresh, and off-screen image release. Measured text is retained only as a loading or failure fallback. Capture readiness does not depend on unrelated document fonts or animation frames that background agent tabs can suspend. The dedicated Find/Replace task pane now has deterministic phone-width coverage for query entry, match navigation, single replacement, disabled-action focus recovery, content synchronization, and modal close-to-invoker focus restoration. The compact revision pane now exposes recording controls inside its modal boundary, reports whether new changes are actually being recorded, retains focus when that state changes, and keeps the truthful active state after the last pending revision is resolved. A deterministic phone workflow covers enabling tracking, creating and accepting an insertion, empty-state recovery, and exact invoker restoration. Revision collections above 48 items now reuse the same bounded window as document navigation, mounting at most 32 contiguous rows while physical spacers preserve the complete scroll range. Arrow, PageUp/PageDown, Home, and End keep the whole review queue keyboard reachable; individual accept/reject decisions retain the same action focus on the next revision. A real DOCX fixture with 120 native OOXML insertions proves bounded mounting, first/last access, the 120-to-119 decision transition, focus continuity, spacer geometry, and an error-free browser run. Comment collections above 48 items now mount at most 32 contiguous cards while the active comment and comments with unsent replies remain pinned. Anchor geometry is collected in one linear pass, connector rendering is limited to mounted cards, and measured card heights are reused when a card leaves the window. Arrow, PageUp/PageDown, Home, and End traverse the complete collection; keyboard navigation keeps the document selection, physical page, review window, and focus synchronized. Deleting a comment moves focus and selection to its adjacent surviving comment. A real DOCX fixture with 120 native OOXML comments proves anchor import, bounded mounting, first/last access, the 120-to-119 deletion transition, page synchronization, and an error-free browser run. New comment drafts stay inside the visible review rail even for document-wide selections, and discard confirmation is limited to drafts or replies that contain written content. Cancelling a task-pane switch now returns keyboard focus to the exact unfinished comment, reply, or citation field on desktop and compact layouts, so protected content can be edited immediately. The same editing context is retained when users cancel comment deletion, citation deletion, or an internal citation switch. Citation source validation is now attached to the exact invalid tag or title field, moves keyboard focus to that field, and clears only when its value changes. A deterministic phone workflow covers validation, source persistence, citation insertion, and exact close-to-invoker focus restoration. Raster-quality page thumbnails and bounded raster admission are implemented. Page buttons now use a bounded window with physical scroll spacers and sparse current/focus pins. A deterministic 120-page DOCX workflow proves Home/End navigation, selection, focus retention, and bounded mounting. The same window model now bounds heading and search-result rows above 48 items to at most 32 contiguous rows plus sparse active, selected, and keyboard-roving pins. The real 120-page fixture contributes 120 headings and 120 matches; deterministic browser evidence proves first/last keyboard access, exact final-result selection, physical spacer geometry, and zero console or page errors. Comments now have the same bounded long-review evidence. Page-chrome descriptor derivation is incremental as well: a tail edit reuses the stable physical-page prefix, including resolved first/default/even headers, footers, page numbers, preview text, and jump targets. The deterministic 120-page DOCX workflow proves that a final-page edit reuses 118 descriptors and derives only the two-page boundary suffix while page navigation, focus, console, and page errors remain stable.

      Exit evidence:

      • Review operations retain authorship, anchors, replies, and state through DOCX round trips within the declared compatibility boundary.
      • Representative 100-page and review-heavy fixtures meet interaction, memory, reflow, and save budgets.

      Priority 2: Markdown

      After the Document gates above are stable:

      • Prove lossless switching between source, split, visual, and preview modes for the supported GFM model.
      • Complete tables, task lists, links, images, fenced code, escaping, and clipboard behavior.
      • Publish Markdown Extension parse/serialize requirements and fixtures.
      • Keep source typing coalesced and visual updates responsive on large files.
      • Keep source undo/redo independent from browser-native textarea history, restore selections, rebase on host replacements, and avoid duplicate visual history records during synchronization.
      • Keep the default split view directly resizable without breaking synchronized scrolling; at phone widths, give Source or Preview the complete workspace through one explicit, touch-sized switch.
      • Keep the preview a calm reading surface with a bounded line length and one visual hierarchy across headings, prose, quotations, code, tables, and task lists.

      Exit evidence: canonical Markdown fixtures round-trip without semantic drift, custom structural Extensions have serialization tests, source/visual undo remains one coherent history, and desktop split plus phone full-pane behavior pass visual and deterministic E2E regression.

      Priority 3: Spreadsheet

      • Replace remaining dense and main-thread workbook work with an A3S-owned sparse model and virtualized viewport.
      • Expand Worker/WASM formula parity, dependency tracking, number formats, sorting, filtering, validation, charts, pivots, and print layout.
      • Keep lightweight selection statistics responsive for dense and sparse sheets.
      • Keep common number and percent presets plus decimal-place commands available from the primary ribbon while the native format engine expands.
      • Keep font family, vertical alignment, and text wrapping available from the primary ribbon and backed by the native workbook cell-style model.
      • Keep worksheet lifecycle actions in the shared Office interaction system: destructive deletion uses a safe default, invalid names remain editable with local accessible feedback, and temporary menus and dialogs restore focus.
      • Keep phone worksheet renaming focused on the active task: unrelated footer tools yield the available width, invalid feedback expands into a readable second row, and successful or cancelled edits restore the normal workbook status controls without covering the grid.
      • Keep Cmd/Ctrl+F inside the editor through the compact Spreadsheet Find bar, including displayed, raw, formula, and sparse-cell matching, deterministic navigation, repeated-shortcut refocus, and grid-focus restoration on close.
      • Keep the phone Find surface within the viewport with a 40 px input and 40 px previous, next, and close actions. The deterministic phone workflow must prove an exact cell result and return focus to the grid after Escape.
      • Keep workbook task panes as bounded desktop regions and focus-contained phone dialogs. At phone widths the ribbon, grid, and worksheet footer remain inert; Tab and Shift+Tab stay in the pane; Escape cancels a dirty draft before it closes the pane and restores the exact ribbon invoker. Deterministic browser evidence covers the modal semantics, focus order, isolation, screenshot, accessibility tree, console, and page errors.
      • Preserve direct type-to-edit after focus recovery and keep multi-cell paste, selection, copy, and undo coherent when React replays vendor state updates.
      • Preserve XLSX relationships, names, styles, formulas, comments, charts, images, and unsupported package state.

      Exit evidence: render cost depends on the viewport, recalculation depends on the dirty dependency graph, and the compatibility fixture matrix passes in Excel and WPS.

      Priority 4: Presentation

      • Complete masters, layouts, themes, placeholders, groups, tables, charts, links, notes, comments, transitions, and presenter workflows.
      • Keep exact row and column table insertion behind the shared, keyboard- accessible Office table picker and one controlled history record.
      • Keep scene object identity stable through selection, transforms, clipboard, history, import, and export.
      • Keep native slide and object context actions available independently of optional AI integrations.
      • Keep the chart inspector as a bounded desktop region and a focus-contained dialog whenever the responsive layout overlays the canvas. The ribbon, workspace, and status bar remain inert; Tab and Shift+Tab stay inside; Escape cancels a dirty field before closing; and close restores the selected chart. Deterministic phone evidence covers semantics, focus order, isolation, screenshot, accessibility tree, console, and page errors.
      • Keep comment review as a docked desktop strip and a full-editor modal on phones. The compact surface uses readable review text and 44 px primary actions, isolates the ribbon, workspace, and status bar, contains Tab and Shift+Tab, lets a dirty comment consume Escape before the panel closes, and restores the exact New Comment, View Comments, or comment-marker invoker. Deterministic phone evidence covers both invoker paths, modal semantics, focus order, isolation, screenshot, accessibility tree, console, and page errors.
      • Keep F5 and Shift+F5 as direct slideshow paths with an in-page fallback when the browser denies fullscreen.
      • Move bounded geometry, text fitting, connector routing, snapping, and thumbnail work behind Worker/WASM contracts.

      Exit evidence: complex slides stay interactive, supported objects preserve their semantics through PowerPoint and WPS round trips, and one gesture creates one controlled update.

      Priority 5: PDF

      • Keep color, opacity, and compatible stroke-width controls routed through typed annotation capabilities for both tool defaults and selected objects.
      • Add form authoring, signatures, page organization, and explicit redaction review while preserving the implemented search, history, and save paths.
      • Keep navigation, zoom, and history reachable from the responsive overflow menu whenever the primary toolbar hides those controls.
      • Retain the implemented active-page thumbnail synchronization, bounded long-document thumbnail window, and focus-contained phone page drawer. Its toolbar-owned trigger must never cover rendered PDF content.
      • Keep PDFium lifecycle, page bitmap cache, and source Blob ownership bounded.
      • Verify saved annotations and forms by reopening the emitted Blob.

      Exit evidence: large files render a bounded page window, edits survive reopen, and destructive operations always show a review step before save.

      Public Extension Plan

      The public Extension surface grows only when its context can remain stable across minor releases.

      Available now

      • DocumentEditor.extensions: additional TipTap Extensions.
      • MarkdownEditor.extensions: additional TipTap Extensions.
      • DocumentEditor.getSelectionMenuItems: a host-owned selected-text menu factory with full document snapshots and conflict-aware text-edit commands.
      • MarkdownEditor.getSelectionMenuItems: the same host-owned boundary across source and visual selections, with complete Markdown context and safe edits.
      • Both editors append host Extensions after built-ins and throw on duplicate names.
      • Vue accepts :extensions; custom elements accept the .extensions property.

      Structural TipTap Extensions own a persistence obligation. Document Nodes and Marks need DOCX import/export handling. Markdown Nodes and Marks need Markdown parse/serialize rules. An Extension that only adds shortcuts, storage, or a ProseMirror Plugin does not change the persisted content schema and is the recommended first integration.

      Stable host extension points

      • Document and Spreadsheet: fileActions and onAgentRequest.
      • Presentation: fileActions, onAgentRequest, and onStartSlideshow.
      • PDF: loadSource and onSave.

      Spreadsheet, Presentation, and PDF command contexts remain internal until each has a typed, versioned Extension context, deterministic command ordering, capability discovery, cleanup semantics, and compatibility tests. Consumers must not import internal modules.

      Playground Documentation Definition of Done

      The Playground integration guide is the public API reference for a release. For each editor it must include:

      • Every public Prop with its exact type, required state, default value, framework binding, and behavioral description.
      • React callbacks, Vue events, and Web Component events or properties.
      • The canonical content model or, for PDF, the Blob load/save lifecycle.
      • Supported Extensions, collision behavior, lifecycle guidance, persistence requirements, and a copyable highlighted example.
      • Document and Markdown selection-menu snapshots, async target lifetime, stale-selection behavior, editing commands, and React/Vue/Web Component bindings.
      • Explicit alternatives when an internal Extension API is not public.
      • Links or examples for React, Vue, and Web Components.

      The guide is tested as product code. Navigation, all five editor tabs, Props tables, content contracts, Extension examples, code highlighting, horizontal table scrolling, and compact layout must pass before release. A public API change is incomplete until the corresponding guide and tests change in the same pull request.

      Execution Order

      1. Keep the shared quality gates green while finishing the current Word vertical slice.
      2. Close that slice with unit, interaction, visual, performance, and Word/WPS round-trip evidence.
      3. Update public types and Playground documentation only after the contract is stable.
      4. Move to the next Word slice. Do not start a lower-priority editor milestone merely because its UI is easier to demonstrate.
      5. Revisit the priority order only with measured user demand, compatibility risk, and engineering evidence.