web

A3S Web

A3S Web is a local-first AI workspace. Work is the single default workbench at #home; file management, Office editing, coding, and AI execution are scenes inside that workbench rather than separate products or conversations. The standalone Knowledge destination follows Work, while installed A3S Use packages can add reviewed workbench views for vertical capabilities such as Research and Finance.

Product implementation is governed by the super-app plan, product blueprint, functional specification, component specification, product architecture, design system, domain model, core user journeys, and TUI-to-Web product mapping. The signed package, contribution, sandbox, messaging, and review contracts are defined in A3S Web Plugin System. The separate, local knowledge-base library, editor, and external-import lifecycle are defined in A3S Web Knowledge, with the independent worker handoff in the Knowledge Compilation Protocol. The task-first Work entry point and its staged evolution are defined in the Work AI-Native Home plan. The persistent artifact surface, reload lifecycle, deep-link contract, and security boundary are defined in A3S Work Live Preview. Native Rust WeChat/Weixin iLink remote management is defined by the product and architecture, tracked in the development plan, and guarded by the security and operations contract.

Current product surface

• A VS Code-style Activity Bar with built-in Work and Knowledge entries. Work is the only task entry and always opens the canonical #home route; coding is entered by opening a code or text file inside Work. Knowledge sits immediately below Work, followed by enabled package contributions ordered from the live A3S Use registry. Memory, the verified plugin Market (A3S Use release bundles plus signed TUF sources), and Settings share the pinned bottom section. Remote messaging channels live behind one Settings → Channels entry rather than occupying top-level product slots. The Channels page uses one compact channel list and one detail surface; Feishu currently displays only “Coming soon”. Research and Finance are not hardcoded entries.
• A standalone, local-only Knowledge page lets users create, search, and pin personal knowledge bases; import Obsidian Vaults and other local folders; and open each base in an Obsidian-inspired file-tree, Markdown editing, and live reading workspace. Work selections can create a source-backed knowledge base without compiling it; users then compile explicitly or enable a per-base smart automatic policy with stability, quiet-window, minimum-interval, retry, and bulk-change safeguards. Only successfully promoted wiki/ generations are searchable. Import copies content into the managed workspace without changing the source folder. Market remains focused on signed Plugins and their verified Sources.
• A single Settings → Channels page with a left-aligned WeChat and Feishu list, without a second framed workspace around the selected channel. WeChat is backed only by the native Rust WeixinModule in A3S Boot. #settings/channels/weixin is the only WeChat route. The Feishu channel is a “Coming soon” placeholder and performs no credential or network activity. The read-only Beta exposes truthful entitlement, binding, account, monitor, and bounded remote target state; covers mock-tested QR, verification, pause/resume, deterministic inventory pages, restart-safe opaque selection, redacted latest replies, and remove-from-this-machine flows; and keeps tokens, cursors, owner identifiers, source IDs, PIDs, and authenticated iLink URLs out of the browser. A3S Boot now constructs the native Tencent runtime when an A3S-specific iLink app identity and permitted bot type are injected at build time, through the process environment, or through local ACL. It then enables real QR binding and the supervised read-only monitor. Missing or invalid identity, unsafe storage, and unapproved hosts fail closed without creating a QR code or sending Tencent traffic. Public release injection and live Tencent validation remain gated on A3S entitlement and security review.
• A Finder-inspired A3S Work local-files workspace backed by the real filesystem: it initially follows the default workspace shown in Settings, while an explicit “Switch workspace” choice is persisted as a user override. Native folder selection, back/forward/up navigation, breadcrumbs, grid and list layouts, current-folder filtering, bounded recursive whole-workspace filename search with an explicit scope switch, result locations and partial-result warnings, generated-tree filtering that retains .a3s/kb, metadata sorting from both the toolbar and clickable list headers, mouse-marquee and checkbox multi-selection, additive range selection, a selection-aware action shelf anchored above the status bar so file positions remain stable, background/item/keyboard context menus, Cmd/Ctrl copy, cut, and collision-safe paste, operating-system reveal, and source-package knowledge-base creation from single or mixed selections. Folder creation, Office-file naming, rename, and duplicate naming stay inline in both grid and list layouts; only irreversible permanent deletion opens a confirmation dialog. Persisted local folder favorites appear in the sidebar and follow Work-initiated moves or renames. Finder-style arrow, Shift-range, Home/End, and Cmd/Ctrl navigation work in both layouts; selected files can be dragged into visible folders, ancestor breadcrumbs, the local root, or sidebar favorites after descendant and name-collision checks. Files and recursively read folders dragged from the operating system are copied into those same targets with bounded reads, collision-safe top-level names, and rollback of newly created roots after a failed import. New DOCX, XLSX, and PPTX files can be created directly in the current local folder from the sidebar, the folder context menu, or Cmd/Ctrl+N. Their names are entered inline in the file grid or list; Work refuses an existing name, writes through a verified sibling replacement, binds the new path, and opens the matching editor immediately. A single selection opens a persistent, resizable live-preview panel from Space, the toolbar, or its context menu instead of interrupting the file workflow with a dialog. The same panel opens from the code editor and accepts an active-workspace path or loopback development URL. Static sites reload after a debounced workspace change and expose desktop, tablet, mobile, zoom, refresh, and pause controls. Bounded text, common raster images, PDFs, and in-memory DOCX/XLSX/XLS/ODS/CSV/PPTX conversions reuse the same read-only preview surface without creating or autosaving a Work artifact. Folders require an index.html entry for site preview; unknown or oversized binaries are not read. Local deletion is explicitly labeled as permanent and requires confirmation because operating-system Trash integration is not yet available.
• An AI-native A3S Work home that starts with the complete durable task composer instead of a file grid. It preserves execution mode, model, effort, @ workspace references, / Skills, drag-and-drop context, and the real queue submission path. Implemented shortcuts create Office artifacts, open files, prepare editable data-analysis and file-organization drafts, or enter the local-files workspace. The AI Assistant stays closed on a fresh visit and opens when a home task is submitted; templates, folders, and recent files remain directly available below the task entry, while Recent, Favorites, Trash, and folder views remain focused file-management surfaces.
• A single-file detail editor for every non-binary code or text file in the selected root. It uses the same header and return flow as Office details while keeping Monaco syntax highlighting, diagnostics, semantic navigation, safe Cmd/Ctrl+S writes with external-change conflict review, and editor context actions for the AI Assistant. Markdown uses a fixed left-source/right-live- preview layout instead of a mode switch.
• A resizable right-side AI Assistant backed by the durable A3S task runtime. The home composer, file workspace, Office editors, and code editor all use the same conversation list, active session, draft, workspace, and default agent. Switching scenes never forks a hidden Code or Office conversation. At 1120 px and below the assistant becomes a full Work-pane overlay with a persistent close action instead of squeezing the file surface. File context menus can prepare ask, summarize, organize, and naming requests; document selections can prepare ask, summarize, rewrite, and translate requests; spreadsheet ranges can prepare data analysis, formula diagnosis, and cell-level change proposals; presentation slides and elements can prepare summaries, copy, layout, and speaker-note suggestions. These actions add bounded format-specific context and a draft but never auto-send. Eligible text, cell/formula, table, and speaker-note changes return through a structured proposal protocol: Work shows trusted before/after values, lets the user select individual changes, and applies only targets whose live source still matches. Layout and formatting advice remains review-only.
• A shared sidebar visual language gives the Work conversation list and Knowledge navigation the same dimensions, navigation rows, framed command glyphs, icon stroke, and interaction states. Work keeps color-coded, extension-badged file tiles where file type is meaningful.
• A compatibility-preserved A3S Work artifact library with templates, recent files, favorites, server-owned autosave, folders, move/copy/rename, recoverable trash, imported source-file recovery, version history. Folder creation and file/folder rename use the same inline naming pattern as the local workspace; moving an item to recoverable trash is immediate, while permanent deletion from trash remains confirmed. It also supports DOCX/XLSX/PPTX import and export, source-backed PDF editing through EmbedPDF and a version-matched PDFium/WASM asset bundled from the installed @embedpdf/pdfium dependency, PDF export from every native editor, and compatibility review before degraded saves or exports. A shared print-preview workflow opens from the toolbar or Cmd/Ctrl+P, shows the actual document, worksheet, notes-page, slide, or handout pagination, supports page navigation and zoom, and applies all/current/custom page ranges to browser printing and PDF export. Native document editing uses a WPS-inspired tabbed ribbon for Home, Insert, Page, References, Review, View, and Tools, with the existing commands grouped by workflow instead of placed in one scrolling row. The centered paper canvas supports page and web views plus real 50–200% visual scaling. Its status bar reports the current/total page and section, word count, toggleable browser proofing, live save state, view controls, and zoom controls. Document editing also includes tables, images, links, find/replace, page setup, margins, explicit page and section breaks, per-section A4/Letter orientation and margins, one-to-six equal or proportional custom-width columns, headers, footers, and page numbers; default, first-page, and even-page headers and footers support rich text, alignment, color, links, inline raster images, and independent page-number visibility. Editable footnotes and endnotes preserve common inline formatting and native DOCX note parts. Editable figure and table captions use independent numbering, expose live cross-references, and round-trip through native DOCX SEQ fields, bookmarks, and REF fields. Editable body fields cover the current page, total pages, current section, section page count, date, and time; they update in shared preview/PDF output and round-trip as native DOCX fields. An editable source library supports APA, MLA, Chicago, and IEEE citations, atomic inline citation nodes, generated bibliographies, source-tag rename propagation, and explicit missing-source states. Native DOCX import and export preserve CITATION and BIBLIOGRAPHY fields plus Word bibliography custom XML, including contributor roles, corporate authors, GUIDs, uncommon source types, and additional simple metadata. Uncommon Word style paths are retained for DOCX output and reported while Work preview uses an APA fallback. PDF files keep EmbedPDF's complete native interaction model for search, thumbnails, annotations, forms, signatures, stamps, redaction, history, printing, and modified-copy export. Saving asks PDFium for a real updated PDF binary, replaces the managed A3S source, and writes the same bytes through the verified local-file path when the artifact is bound; the header keeps original source download explicitly separate from modified PDF export. Body-text revision mode records typed, pasted, replaced, and Backspace/Delete text as attributed insertions or deletions; a review panel can locate, accept, or reject one or all changes. Pending revisions stay visible in read-only/PDF preview and round-trip through native DOCX w:ins, w:del, author/date metadata, and the trackRevisions setting. Browser-native spelling assistance is explicitly enabled on the editable document surface. Selected body text can also carry editable comment threads with replies, resolve/reopen, navigation, and deletion. Anchors, plain-text bodies, authors, dates, replies, and resolved state round-trip through native DOCX comments; read-only preview highlights anchors while print and PDF output keep the body text visually plain. Rich comment-body formatting, images, tables, and content controls normalize to plain text under an explicit compatibility warning. Structural, moved-content, formatting, table, and numbering revisions remain compatibility work. These document settings drive read-only preview and PDF pagination, including page-specific footnotes and document-end endnotes. Spreadsheet interoperability preserves common layout and style metadata, filters, frozen panes, hyperlinks, named ranges, embedded raster worksheet images, print areas, repeated print-title rows and columns, manual row and column page breaks, A3, A4, A5, Letter, Legal, or Tabloid portrait or landscape page setup, percentage or fit-to-page scaling, four page margins plus header/footer offsets, page centering, editable left/center/right print headers and footers with dynamic fields, first-page numbering, selectable page order, common validation rules, editable legacy cell comments, and editable common conditional formats including all eight core OOXML cell comparisons, explicit color-scale and data-bar thresholds, data-bar value visibility and display lengths, and standard icon sets. Print titles and manual page breaks are editable and control every spreadsheet PDF page; spreadsheet PDF export also honors paper, orientation, scaling, margins, and centering, print headers and footers, page numbers, and page order. Conditional-format priority is editable and serialized deterministically, stopIfTrue is preserved and evaluated per matched cell, and missing, invalid, or duplicate source priorities are normalized with diagnostics. Sheet protection and passwordless editable ranges have XLSX round-tripping alongside deterministic conditional-format sheet/PDF rendering; worksheet images retain their native one-cell, two-cell, or absolute positions, sizes, names, alternative text, and media parts, remain movable and resizable in the editor, and appear in read-only and PDF output. Anchored column, bar, line, pie, doughnut, area, radar, scatter, and bubble charts retain cached data and live cell references, titles, legends, names, alternative text, positions, and sizes. Doughnut hole size; standard, marker, or filled radar styles; five scatter line/marker styles; and bubble scale, negative-bubble, and area/width sizing behavior are editable. Scatter and bubble series keep independent X, Y, and bubble-size references. Column, line, and area series can be combined with editable primary or secondary axes and native multi-plot ChartML round-tripping. Legends can be placed on the right, left, top, bottom, or top-right and optionally overlay the plot area. Column, bar, line, and area charts support standard, clustered, stacked, or percent-stacked layouts as applicable; column and bar charts expose gap width and series overlap, while line charts expose smoothing. Positive and negative stacks accumulate independently, percent stacks normalize each category, and stacked areas retain their lower and upper boundaries across editor, preview, PDF, AI Assistant, and native ChartML paths. Each series also exposes a portable solid sRGB fill with transparency, line color, width and solid/dash/dot/dash-dot pattern, plus applicable marker symbol, size, fill, and outline controls. Series appearance uses the same editor, preview, PDF, AI Assistant, and native ChartML paths. Bottom, left, top, and right axis titles can contain direct text or live worksheet references. Category and value axes expose editable normal or reverse order, four label positions, and four major-tick styles; category axes also expose an automatic or fixed label interval. Value axes additionally expose editable minimum and maximum bounds, major units, major-gridline visibility, number formats, and source-format linking. These settings drive editor, preview, PDF, AI Assistant, and native ChartML output. Each supported series can contain multiple editable linear, exponential, logarithmic, polynomial, power, or moving-average trendlines with forecasts, a fixed intercept, and optional equation or R² labels. Per-series data labels can combine values, category names, series names, pie/doughnut percentages, and bubble sizes with an editable separator and nine native positions. Per-series error bars support positive, negative, or bidirectional fixed values, percentages, standard deviation, standard error, and custom worksheet ranges, with optional end caps; scatter and bubble charts additionally support independent X error bars. Axis titles, trendlines, data labels, and error bars survive native XLSX round-trips and render in the editor, read-only preview, PDF output, and AI Assistant context. Charts can also be created from the current cell selection, moved, resized, deleted, and regenerated as native XLSX chart parts. Unsupported combination families, axis-title layouts, logarithmic scales, invalid axis orientation, label placement, intervals, bounds, or units, minor ticks, minor gridlines or units, display units, custom crossings, trendline parameters, malformed error bars, manual or per-entry legend layouts, mixed per-series smoothing, invalid plot spacing, per-point label overrides, theme colors, gradients, patterns, effects, custom dashes, and remaining advanced chart behavior are reported separately instead of being silently flattened; ordinary formulas retain exact future-function prefixes and cached Excel error types; legacy array, dynamic-array, and what-if data-table ranges retain native XLSX grouping, input references, and cached results. Shared formulas expand to equivalent per-cell formulas with an explicit diagnostic instead of retaining fragile package-level compression. Workbook automatic, automatic-except-data-tables, and manual calculation modes, full-recalculation flags, iterative limits, convergence thresholds, and full-precision behavior round-trip through native calcPr. A formula-and-calculation panel reports external and structured references, unsupported or volatile functions, cached errors, shared-formula normalization, and spill conflicts, and can explicitly recalculate the current selection or the full workbook. Formula settings and grouped-range context are also included in bounded AI Assistant selection context, while cached error text remains visible in preview and PDF output. Conflicted grouped ranges safely export as ordinary formulas rather than generating an invalid XLSX; worksheet-range pivot tables can be created from the current selection; assign row, column, and multiple value fields; choose from 11 deterministic aggregations; configure captions, styles, refresh-on-open, and grand totals; and add multiple single-selection report filters whose value is either all or one text, numeric, Boolean, or blank source item. Refresh writes filter summaries and aggregates into a protected report range. Pivot definitions and active filters are included in bounded AI Assistant context and materialized before preview/PDF output. Supported pivots round-trip as native XLSX pivot-table, page-field, cache-definition, and cache-record parts. External or named sources, grouping, calculated fields or items, multi-selection report filters, label or value filters, values-on-rows layouts, slicers, timelines, and data-model features keep their visible cached cells and receive explicit compatibility diagnostics; unsupported x14 data-bar direction, axis, border, gradient, and negative-color controls are reported explicitly; presentation editing includes rich text, images, tables, charts, layering, links, speaker notes, and editable traditional slide comments with author, date, and position preservation. A full-fidelity in-app clipboard supports element or whole-slide copy, cut, paste, and duplicate from the toolbar or standard Cmd/Ctrl shortcuts; pasted slides regenerate slide, element, and comment identities. Native presentation charts can be inserted and edited in place with titles, categories, multiple series, column, bar, line, area, pie, doughnut, radar, scatter, and bubble types. Scatter charts expose five line/marker styles; bubble charts expose per-series sizes, scale, negative bubble visibility, and area/width sizing. These settings, doughnut hole sizes, and radar styles survive PPTX round-trips and share the editor, thumbnail, playback, and PDF preview path. Legend visibility and five-position placement share that path, including optional plot-area overlay. Column, bar, line, and area charts expose applicable standard, clustered, stacked, and percent-stacked grouping. Column and bar charts add editable category gap and series overlap, while line charts add smoothing. Positive and negative stacks accumulate independently and percentage stacks normalize per category. Each series also exposes portable solid sRGB fill and transparency, line color, width and four dash patterns, plus applicable marker symbol, size, fill, and outline controls. Layout and series appearance use the same editor canvas, thumbnail, playback/PDF, AI Assistant, and native PPTX ChartML paths. Primary horizontal and vertical chart axes expose titles, reverse order, four label positions, four major-tick styles, category-label intervals, value bounds, major units, major-gridline visibility, and number formats. The same axis model drives the editor canvas, playback/PDF rendering, AI Assistant context, and native PPTX round-trips; secondary, date, logarithmic, minor-unit, display-unit, and custom-crossing axis features receive explicit compatibility diagnostics. Chart-level data labels can combine values, category names, series names, pie or doughnut percentages, and bubble sizes with an editable separator and common family-specific positions; the same labels appear in the editor, playback/PDF rendering, AI Assistant context, and native PPTX round-trip. Supported Cartesian series also expose multiple linear, exponential, logarithmic, polynomial, power, or moving-average trendlines with names, order or period, forward and backward forecasts, fixed intercepts, equations, and R² labels. Per-series error bars support positive, negative, or bidirectional fixed values, percentages, standard deviation, standard error, and custom numeric values with optional end caps; scatter and bubble charts additionally support independent X error bars. Imported custom-error formulas are retained even when no worksheet is available. Trendlines and error bars share the editor, canvas, playback/PDF, AI Assistant, and native PPTX ChartML round-trip paths; unsupported parameters, duplicate directions, manual trendline labels, manual legend layouts, invalid plot settings, theme or effect-based series formatting, and per-point styling receive explicit, feature-specific compatibility diagnostics. A review strip locates, edits, and deletes comments across slides; review markers stay out of slide-show and PDF output. Modern PowerPoint threaded replies, mentions, and resolved state remain explicit compatibility work. Fade, push, wipe, split, and cut slide transitions support click or timed advance, PPTX round-tripping, and playback. Slide masters and layouts are editable through a design panel with layout application, inherited backgrounds and artwork, and title/content placeholders. Imported PPTX files retain native layout assignments, placeholder geometry and basic text-style inheritance, including layouts not currently used by a slide; PPTX export emits native layout parts again. Thumbnails, playback, presenter view, PDF slides, notes pages, and handouts all render the same inherited design layers. Presentation preview also includes a presenter view with the current and next slide, speaker notes, elapsed-time controls, and navigation; PDF output can produce full slides, notes pages, or two-, three-, and six-slide handouts; that layout is selected and reviewed inside the shared print preview. Existing browser-only files migrate automatically, while IndexedDB remains a compatibility fallback. Opening DOCX, XLSX, or PPTX from the filesystem binds one managed Work copy to its source path on this device. Reopening an unchanged path returns to that copy instead of creating duplicates. If the disk bytes changed, Work opens a new bound copy and keeps the previous copy as unbound recovery state. Cmd/Ctrl+S checks for outside changes before an explicit save-back; Save As and conflict review avoid silent replacement, and successful writes use a verified sibling replacement. Platform-independent conditional atomic replacement and native file watching remain completion work. The complete local-workspace and Office-handler targets, including remaining interoperability, collaboration, and fidelity requirements, are tracked in the Work Office contract.
• A WorkBuddy-aligned global Settings dialog that preserves and dims the current product; Help and Shortcuts is a searchable dialog tab rather than a separate full-screen page.
• A compact unified conversation list with on-demand search. All historical sessions remain visible; new sessions use the default agent, and rename and delete stay inline in the affected row.
• A dedicated Memory product at #memory with complete-store retrieval, plain-language search and filters, a lazy-loaded 3D relationship view, a chronological list, and focused memory/entity detail. Internal scores, paths, raw fields, and identifiers are not shown. A separate Learning tab shows only items that need review by default and supports save, ignore, reconsider, update, and recoverable version rollback.
• A continuous task conversation in the resizable AI Assistant. Home, files, Office, and code scenes share its active session, draft, queue, model, effort, and workspace; compact widths use an overlay with a persistent close action.
• Natural-language task composition with a searchable recent-workspace picker, native local-folder selection, a lazy color-coded @ workspace tree, searchable highlighted Skill suggestions, safe file/folder drop import, and inline /goal control.
• A calm turn-based execution document with restored Skill/file context, stable A3S response headers, local copy and continue-edit actions, lifecycle-aware Markdown reasoning, reader-controlled stream following, and CJK-aware, typographically tuned Streamdown/Shiki rendering. Newly streamed words fade in without replaying animation on existing content, and incoming Markdown boundaries stay intact so paragraphs, lists, and tables settle as their delimiters arrive. The transcript follows a growing response with interruptible eased scrolling, stops when the reader moves away, and respects reduced-motion preferences. Headings, lists, task lists, quotations, tables, links, images, footnotes, inline code, and line-numbered code blocks share the same document treatment. Model-generated tables with a short delimiter row are repaired outside fenced code before rendering. Its semantic tool timeline merges live events, output, HITL decisions, and recovery into one execution block, with TUI-aligned command, JSON-argument, and generic tool-call syntax highlighting across persisted Web API aliases, working-directory context, live output metrics, copy actions, and a compact tail preview after completion. Successful file edits render an inline unified Diff with added/deleted counts, old/new line numbers, and full red/green row backgrounds.
• Visible task execution mode with a mode-specific icon, provider-tabbed model selection, an independent Effort slider with English values and Chinese guidance, task goal timing, context usage, manual context compaction, and an upper-right task-runtime panel that appears only for a real planning or subagent lifecycle, then shows only the checklist, completion, elapsed time, and parallel work actually published by the runtime. Ordinary task creation, analysis, and streaming never manufacture progress UI. First submission locks its controls and shows creating or queueing feedback before the initial API round trip completes. Dedicated Use workers are identified by their observed standard MCP routes (Use · Browser, Use · Office, or multiple deduplicated routes), while their evidence uses readable domain actions instead of raw MCP tool names.
• A3S Web shares the TUI execution boundaries: Default uses the active workspace's host command runner behind the shared Rust guardrail, silently admits only commands proven read-only, and asks before unproven host work. Plan denies Bash, while Auto admits only the proven read-only subset and denies other host commands without opening HITL.
• A bounded startup transition with explicit loading, version-mismatch, disconnected, retry, and technical-detail states. Older local services that lack the model-catalog route fall back to their configured Provider models instead of blocking the workspace.
• A Work-native code and text scene with a compact lazy explorer, color-coded file glyphs, Cmd/Ctrl+P quick open, lazy Monaco editing, independent dirty tabs, Simplified Chinese code-navigation menus, keyboard save and guarded close, external-change conflict protection, Markdown source/live preview, and saved-file Code Intelligence for symbols, navigation, and diagnostics. Quick open and AI artifact links delegate to Work's file handler, so Office files and source files enter the correct mounted scene.
• A larger, deliberately simple local configuration center for A3S OS account and endpoint, runtime-detected Claude Code, Codex, and WorkBuddy account status and refresh, appearance, models and Providers, Agent execution and queues, session storage and memory policy, search and headless browsing, document/OCR parsing, MCP transports, and OAuth. The dedicated connector setup has been removed; custom endpoints remain regular MCP servers. Model and Provider settings use one list/detail flow with the default model, connection, model catalog, and advanced limits kept in a clear order. Settings also covers updates, service information, and searchable Help. Categories load and save independently; effect labels distinguish new-task changes from restart-required changes, and secrets remain masked. The expanded dialog keeps the default model and common Provider connection fields prominent, while runtime overrides, model limits, headers, and session passthrough stay in disclosures. Editable Provider names and model IDs retain focus and disclosure state.
• A source-grouped task model switcher that combines configured Provider models with valid local Claude Code, Codex, and WorkBuddy account models through the same runtime discovery and client routing used by the TUI. Account credentials remain in the CLI, while Settings reports only connection state and available model counts.
• Browser-native command (Cmd/Ctrl+K) and file (Cmd/Ctrl+P) palettes for existing pages, contextual actions, and bounded workspace navigation.
• Wide and compact desktop layouts, plus system, light, and dark themes.

The primary journey is:

Open #home → describe an outcome or choose a local folder
→ browse and select real files → prepare visible context in AI Assistant
→ execute and decide permissions → inspect evidence or proposals
→ open the resulting Office or code file in Work
→ edit, validate, and explicitly save back or Save As
→ continue the same conversation

Code Intelligence

Monaco shares the workspace-scoped native Code Intelligence runtime owned by the CLI backend. The file toolbar exposes definition, declaration, references, implementations, and Go to Symbol in Editor together with their keyboard shortcuts. The same navigation commands remain available from Monaco's context menu, whose native actions and A3S additions use the same Simplified Chinese catalog as the surrounding Web IDE; diagnostics appear as Monaco markers. The editor toolbar keeps a bounded, workspace-scoped location history: Ctrl+- returns to the previous file and caret, while Ctrl+Shift+- moves forward. Open drafts are reused rather than reread, a new navigation clears the forward branch, and file rename/delete operations reconcile stored locations. Dirty buffers remain browser-local, and the status bar says when semantic results use the saved version while also following the live cursor, selection, and Monaco model line ending. Its line ending control converts between LF and CRLF through Monaco's undo history and normal dirty-file protection. Files without a native language profile keep Monaco's local editing and outline features without exposing backend protocol errors; the document-symbol bridge always uses the mounted Monaco model's concrete language identifier, including plaintext fallback, so an unfamiliar extension cannot corrupt Monaco's provider registry. Genuine runtime failures remain visible as concise status messages.

Cmd/Ctrl+P opens a keyboard-first workspace file picker from Monaco or the surrounding task surface. It ranks exact filenames, path fragments, and fuzzy subsequences, keeps open tabs at the top before a query, ignores dependency and build outputs, and reports bounded or truncated catalogs without blocking the current draft. Selecting an already-open file reuses that tab and its unsaved content. Explorer directory entries and quick-open results share one file-type contract: common source, configuration, and lockfile formats remain editable; known binary formats use extension hints and unfamiliar extensions are sampled before either entry point labels them as binary.

The editor and diff viewer share one lazy local Monaco runtime. It retains the complete standalone editing surface and the JSON, CSS, HTML, JavaScript, and TypeScript language services, while registering tokenizers only for the file types the product opens as source text. The runtime does not import Monaco's broad package entry, so unused bundled languages and protocol clients stay out of the editor activation graph without weakening find, folding, suggestions, diagnostics, navigation, themes, workers, or diff behavior.

Explorer mutations preserve the visible tree as well as open editors. Renaming an expanded directory rebases its loaded descendants, expansion state, and retry state; deleting a path evicts its entire cached subtree immediately. Late directory responses are ignored, so a failed parent refresh cannot bring back an old name or a deleted child. New files use the backend's atomic create-only operation, so a conflicting path reports an error in place without truncating the existing file.

Text reads also return a content revision. A normal save sends that revision as its server-side precondition and performs no client-side read-before-write request. Restored legacy tabs without a revision use their last saved content as the precondition. A mismatch returns HTTP 412 without changing the file, after which the editor fetches the latest disk version and offers reload or explicit unconditional overwrite. Successful saves and both conflict resolutions update the tab's revision before the next edit.

The entire editor workspace is task-scoped rather than repository-scoped. Switching tasks snapshots tabs, dirty drafts, the active selection and view, explorer and search state, Git review state, conflicts, and load errors, then restores the destination task even when both tasks use the same repository. Within one live page, every text document has a task-isolated Monaco model URI. The mounted editor switches models instead of remounting, so each open document keeps its undo/redo history, cursor and selections, folding, and scroll state across file and task switches. An explicit search or semantic location is consumed once; later tab returns restore the document's own view rather than replaying a stale jump. A successful file or parent-directory rename rebinds the new logical path to the document's existing immutable model URI before its tab path changes, preserving that same editing session and live status. If a different file is later opened at the old path while the renamed document is still retained, it receives a distinct URI instead of sharing undo or view state. Models remain retained only while referenced by an open tab in the active task or an inactive task snapshot. Cancelling a guarded dirty close keeps the model, while confirmed close and snapshot removal dispose the orphaned model deterministically. A page refresh restores persisted drafts and tabs but starts a new in-memory Monaco undo history. Because the switch does not discard content, dirty drafts do not interrupt it; destructive close, reload, replacement, and overwrite operations still require explicit resolution. Workspace reads and mutations carry a monotonic task generation guard, so responses from a previously selected task cannot write into the current editor. Versioned browser-local snapshots are debounced during editing and flushed synchronously on pagehide, so a normal refresh restores the active task as well as inactive task snapshots. Malformed data is ignored; if browser storage cannot fit the complete caches and clean documents, the fallback snapshot preserves dirty drafts and restores other tabs with an explicit retry state.

The first release supports Rust through rust-analyzer and TypeScript/JavaScript through typescript-language-server --stdio. See the Code Intelligence guide for installation, shortcuts, saved-file behavior, and the typed local API.

Text search excludes repository metadata, dependency caches, and common build outputs by default so source matches remain useful. The Search panel can opt into those directories and reruns the current query immediately; query, scope, or workspace changes make the displayed replacement set stale until a search succeeds. Search renders at most 300 matches and probes one additional match; when that sentinel exists, the panel reports truncation and keeps replacement disabled until the user narrows the search. Search, replace, file loading, and writes continue through the existing workspace APIs. Code Intelligence returns only semantic metadata and locations, then reuses the normal file-selection flow to open a target.

Memory visualization

Open Memory from the Activity Bar, the command palette, or #memory. The surface loads every indexed memory page while requesting graph topology once, so search, filtering, timeline results, and graph projection use the whole local store rather than the first response page. The WebGL renderer is loaded only when Graph is visible. Focused mode renders the compact service projection; Panorama keeps a connected, selection-aware sample bounded at 600 nodes and 4,000 relations, while always reporting rendered and complete totals. An accessible node browser provides keyboard selection outside the canvas. Refresh failure keeps the last successful snapshot visible with an inline retry.

Learning candidates come from the CLI's LLM-structured memory evolution service; the browser does not infer them with keyword rules. Matching evidence accumulates across sessions, and a conflict-free candidate can become a local versioned asset automatically only after strict maturity thresholds. It is never published automatically. The default list contains ready candidates and saved candidates with updates. Saving, returning to the unmaterialized baseline, or restoring a learned preference or method uses the typed Evolution API. Preference instructions are injected into bounded prompt context, Skills are loaded into the registry, and the activation barrier clears only after all affected live sessions rebuild successfully. Existing versions, audit history, and recovery actions remain available in a collapsed History section without crowding the review flow.

Memory visualization is intentionally non-destructive. Manual consolidation and forgetting remain excluded until those actions have explicit review and recovery journeys. Task branching, automation authoring, and global processes are also deferred. Local knowledge management is handled by the separate Knowledge product and does not mutate Memory entries. Plugin workbench views and signed lifecycle management are supported, while callable plugin actions remain CLI/MCP/Skill surfaces.

The logo at public/logo.png preserves the authoritative A3S OS asset. Its SHA-256 is 72b94cf69a95dc6153f865c4f8742c0f67079caa876f35f8b2b5f970ea795a2d.

Stack

• Rsbuild, React 19, and Monaco Editor
• React Force Graph 3D and Three.js for the lazy Memory graph scene
• TypeScript 7
• Valtio for shared client state
• Biome 2.3.14 with A3S OS formatting conventions
• Vitest and Testing Library
• Basecoat CSS, Tailwind CSS, Streamdown, and A3S semantic design tokens

Development

Install Bun, then install the exact locked dependencies for direct app-local commands:

bun install --frozen-lockfile

Start the CLI API from the monorepo root:

cargo run --manifest-path Cargo.toml -- \
  web --host 127.0.0.1 --port 29653 --web-dir apps/web/dist/workspace

For frontend development, run the API separately and start Rsbuild from this directory:

A3S_CODE_API_ORIGIN=http://127.0.0.1:29653 bun run dev

Build the frontend consumed by a3s web:

bun run build

The production output is dist/workspace. From the monorepo root, just web checks for Bun, runs the frozen dependency install, builds this application, and starts the complete local service. It safely replaces an existing same-workspace A3S Web process on the configured port after verifying its health PID, executable, command, and port. If another application owns port 29653, the command refuses to stop it; set A3S_PORT to another port or stop that application explicitly.

Validation

bun run format:check
bun run lint:check
bun run typecheck
bun run test
bun run build

The Web API is implemented in src/api/code_web. Browser code does not access A3S OS tokens directly; the CLI owns token storage and refresh. The plugin endpoints expose only verified catalogs and content plus digest-reviewed lifecycle operations; see A3S Web Plugin System.

Deployment boundary

The default service binds to 127.0.0.1. The browser and a3s web are expected to run on the same machine. Do not expose file, session, configuration, or Git APIs on a non-loopback interface without an authenticated gateway.