Download Lustrew Desktop

Added

• Pick how hard Lustrew AI works in each chat, from Auto to Max. Background tasks follow it.

Fixed

• Answer a paused task right in the chat.
• Approval cards you already answered stay answered.
• "Always allow" stays with its workspace.
• Stop in the workspace terminal waits until everything has ended.
• Buttons and menus use the app's font.

Added

• Before and After in Workspaces shows each line's number, so you can find a change in your editor.
• Running a skill offers what you typed last time under each field, and one press fills it in. Passwords and other secrets are never kept.

Fixed

• A password detail in a skill is typed into a password box.

Changed

• The welcome card is shorter: five things Lustrew AI does, each with a picture, and the details in tooltips.
• A new chat says what Lustrew AI is for: "Ask Lustrew AI to do something on your PC".
• Workspaces and the status bar now call the same switch "Auto-accept edits", and on means the same thing in both.
• The command palette uses the app's own icons, clears its search, shows its keys and ticks the theme you use.
• Chat search says it looks through all chats by title.
• In the Files tab, "Show in the file tree" says where it goes, and "Wrap long lines" is a switch with a tick.
• The plus menu's workspace list shows how many folders each workspace has.
• In Workflows, "Record a task" is a plain button and "Already taught" sits under "Recordings to teach".
• The (i) buttons in Settings sit right after the words they explain.
• "Task Logs" is now "Task logs", and "Make your own" on Workspaces is "New workspace".

Fixed

• Change password ticks off each rule as you type.
• A task's tooltip in Task logs starts with its whole title, and a month's numbers say what they count.
• Task logs no longer says "nothing ran" twice.
• Long names on a Workflows recording are shortened instead of cut off.
• "..." buttons look pressed while their menu is open, and the Files tab's small buttons are easier to hit.
• What Lustrew AI learned starts with a capital letter in Settings.
• Small text across the app is easier to read.

Added

• The Files tab is now a file explorer: your workspace's folders, this chat's files and your cloud space in one tree, with tabs for the files you open.
• You can edit text and code files right in the Files tab and save them with Ctrl+S. If the file changed somewhere else meanwhile, you choose whether to keep your edit or reload theirs.
• Switch the tree between This PC and Cloud to browse, open and edit the files in your cloud space.
• Word documents, PDFs, spreadsheets, slides and pictures open with page, zoom, sheet and find controls.
• Markdown, web pages and JSON have a Preview next to their text.
• Videos and sound files play in the Files tab with full controls, and a file can cover the whole window.
• Make new files and folders, rename, duplicate, move (drag onto a folder, or cut and paste) and delete them from the Files tab. Deleting uses the Recycle Bin, or the trash in your cloud space.
• Restore files from your cloud space's trash, or empty it.
• Go to any file with Ctrl+P, and search inside your files with Ctrl+Shift+F.
• Sort the files by name, date, type or size, and drag tabs to reorder them.
• Closing Lustrew Desktop, updating it or signing out with unsaved files asks before anything is lost.

Changed

• Lustrew Desktop only runs its own code now, so nothing slipped in by a document, a web page or a reply can run inside it.
• Chat cards and widgets are off for now. One in an older chat shows as its code.
• A file Lustrew AI writes no longer pops open in the side panel on its own.
• Files comes first in the side panel, and a file's less used actions are in its More menu.
• Open files follow changes made in other apps.

Fixed

• Saving a file keeps its own encoding, line endings and Windows settings, so older files from Excel or Notepad are no longer damaged.
• A web page can no longer reach Lustrew Desktop's own local service.

Fixed

• Word and PDF files open in a protected preview, so nothing inside a document can run in Lustrew Desktop or change how it looks.
• A huge Word or PowerPoint file no longer freezes the preview.

Fixed

• The window and the taskbar now say Lustrew Desktop instead of ADA Copilot.
• The installer you download is now called lustrew-desktop-setup.

Changed

• The app is now called Lustrew Desktop, and the assistant is Lustrew AI. Your chats, files and settings carry over as they are.
• The Start menu and desktop shortcuts now say Lustrew Desktop.
• Lustrew AI still answers if you call her ADA.

Added

• A new chat home that asks what's on your mind and offers ideas to start from: try again on a task that didn't finish, run a skill you taught, see what changed in your workspace, or begin a piece of research, writing or a task on your PC.
• Live chips on the home show a task that needs you or is running, and how many tasks ran today.
• Settings is organised by what you came to do: Setup shows at a glance whether this PC, the Chrome extension and the app are ready, then Preferences, Account and Your data.
• You can now turn the daily workspace summary off, switch to a compact window and sign out from Settings.

Changed

• Workflows, What ADA knows and Workspaces show counts, progress and who does each step as marks, bars and chips instead of sentences.
• A skill's page shows which steps ADA can do as a gauge, a run's result as a bar, and each step you do yourself with a lock or a person.
• Teaching ADA from a recording walks through Pick, Details and Name, with the values you typed shown as chips.
• The chat list's workspace filter is a proper menu with chat counts, and shows clearly when a filter is on.
• Every page uses the full width of the window, with a second column on wide screens.
• Every tip opens from the keyboard as well as the mouse, and a citation shows its source's title.

Added

• A new look: the Lustrew AI logo, and an indigo sky with glass panels across the app.
• Task Logs shows every task ADA ran on its own, what it made and whether it worked, and opens any task to its answer and each step.
• A new "What ADA knows" page, organised by topic, where you tell ADA things and see what it remembers.
• Workspaces now leads with what ADA changed, and you can undo everything one task changed in one go.
• Workflows you taught ADA have their own page, split into named parts.
• The chat header shows ADA's plan for the task and how many tokens the chat has used.

Changed

• Status reads at a glance as marks, meters and chips instead of sentences, with the details a hover away.
• Pages use the full width of the window.
• Stopping ADA mid-reply keeps what it had already said, in the chat and after a reload.
• A password you type into a chat or a request is never saved; if a task needs it again, ADA asks you.
• Opening a file from ADA never runs a program or script.

Fixed

• Chat bubbles show when each message was sent, not when the chat was opened.
• A long or quiet reply no longer freezes or drops the chat.
• An expired sign-in refreshes by itself instead of leaving the chat list empty.
• Deleting all chats removes every one, and a failed update check no longer says you are up to date.
• Find and search work, and every chat can be reached with the keyboard.
• Reduced motion stops animations without making other changes jump.

Added

• Excel files: ADA can now build multi-sheet workbooks, add, rename, and remove sheets in place, write working formulas (including ones that reference another sheet), and format cells like currency or bold.
• Hand-off and report-back: for a job that takes a while, ADA can run it in the background and then continue on her own when it finishes, telling you the result without you having to ask.

Fixed

• Reading a spreadsheet back now shows the formula in a cell that has not been calculated yet, instead of a blank one that looked empty.

Changed

• Task Logs has a cleaner, calmer look that matches the rest of the app, with a tidy list of tasks you can scan at a glance.
• Opening a task now shows a simple top-to-bottom view of what you asked, what it produced, and each step it ran, with a switch to the full graph view when you want it.

Added

• ADA now shows her thinking above an answer when she reasons one through. It stays folded away by default, so open it only if you want to see how she got there.
• Files ADA makes in the cloud show at the top of the Outputs panel, and you can save any of them to your computer with one click.
• The Workspaces page has a new Cloud tab for browsing, reading, and downloading the files ADA keeps in the cloud.
• The terminal panel shows work running in the cloud next to work on your PC, so you can tell where each one is, watch its output, and stop it.
• When a step runs in the cloud instead of on your PC, a small cloud icon marks it.

Changed

• A fresh look, with a friendlier typeface and a calmer background.
• The steps ADA runs, like commands, are quieter now: each shows what it did on one line, and you can open one to see its full output.
• The typing indicator feels a little more alive.

Fixed

• Text is easier to read in light mode.
• The message box no longer grows tall on its own when it is empty.
• The list of changes ADA made to your files loads again, so you can review and undo them.

Changed

• Removed the separate Activity panel; ADA already tells you what she's doing in her replies, so the screen is calmer.

Added

• If you send a message while ADA is still replying, a small note shows it's queued and will send next, and you can cancel it.

Fixed

• If you send another message while ADA is still replying, it now goes as a follow-up the moment she finishes, instead of starting a second reply that overlaps the first.

Added

• ADA can now run commands, builds, and scripts for you in a workspace, and you watch the live output in a terminal panel right inside the app.
• Self-contained work like builds, tests, and installs can run in a secure cloud sandbox instead of on your PC, so it keeps going even when your machine is busy or asleep.
• Long-running jobs move to Background mode and keep running while you carry on, with an indicator beside the message box.
• Show ADA a task once and it can repeat it for you later.
• Settings now shows the lessons ADA has picked up from your conversations, as cards you can open.

Changed

• The Workflows page shows its filmstrip right away, before you expand a run.

Fixed

• You can clear finished or stopped processes from the terminal panel.
• After ADA changes something on your screen, it checks its own work so you do not have to ask.

Fixed

• The Web switch now shows a spinner the instant you click it and ignores repeat clicks until it finishes, so it no longer feels unresponsive or bounces you back to Copilot.

Added

• ADA Web is ready: switch between the Copilot and the web app from the title bar, and stay signed in across the switch.

Changed

• The Copilot and Web switch now stays hidden until you sign in.

Fixed

• The buttons on the Workspaces page no longer overlap or squeeze together.
• The Tour button on the Workspaces page works again.
• Links and URLs in your own messages are now clearly visible.

Changed

• Your message bubbles have a softer, lighter blue look.

Fixed

• Dragging the window by its title bar works again.
• The tasks pill only appears when you actually have tasks.

Added

• A new title bar: ADA, with a Copilot and Web switch in the middle. ADA Web is coming soon, and the switch says so until it is ready.
• The logo now eases in on the splash screen when the app starts.

Added

• ADA can now work inside your browser from a normal conversation: open pages, read them, click buttons and links, fill whole forms in one go (text fields, radio buttons, checkboxes, dropdowns, sliders, dates and times), scroll to any part of a page, and drag items such as cards between columns. Every step shows as it happens.
• Strict limits are built in, not just promised: ADA never types into password, payment or verification-code fields (the request is refused at two independent layers), never presses submit, send, buy or sign, and never drags anything onto a target that would delete or send. Those actions stay yours.
• Interactive widgets in chat: ask for a small tool — a counter, checklist, calculator, game — and it appears as a working mini-app in the reply that remembers its state, and can be dragged out into its own floating window.
• Long code in replies now folds to a preview with a "Show all" bar, tool activity collapses into one expandable line per run, and widget windows carry a slimmer single-line header, so conversations stay compact.

Changed

• Conversations run on a stronger model. It checks pages before making claims, admits and corrects its own mistakes with evidence, and stops repeating an approach that has already failed twice — it diagnoses or tells you plainly instead of retrying forever.
• Tabs ADA opened stay open when the work is done — the page on screen is your receipt. It closes them only when you ask.
• If a page is still loading, ADA now waits and looks again instead of closing and reopening it.

Fixed

• Actions aimed at a specific browser tab can no longer land on whichever tab you happened to be viewing. If the intended tab cannot answer, the action fails honestly instead of touching the wrong page.
• Screenshots requested for a specific page are verified to actually show that page before being presented.
• Time and date fields on web pages are now visible to ADA (they were previously skipped when reading a page's controls).
• A memory glitch that silently prevented past conversations from being remembered has been fixed.

Changed

• Workspaces has been redesigned. One line at the top tells you how many folders ADA can open and how many it can change. Folders are listed by name with their location beside them, and permissions read as "Can edit" or "Read only".
• Changes ADA made are now grouped by day and read as plain rows, with Undo on each one. Temporary and test files are hidden behind a toggle so your own files are easier to find.
• Switching workspace, creating one, renaming and deleting all moved into proper dialogs that say what will happen before you confirm.
• The page now shows how much space your undo history is using.

Fixed

• ADA now answers from your own uploaded documents when you ask about them. Questions phrased like "according to my SOPs, what must never happen?" used to be answered from the open web instead of the file you uploaded.
• Uploading the same file twice no longer creates a second copy of everything in it. Anything already saved is recognised and skipped.
• After an upload, ADA tells you what it actually saved (how many documents and tips) instead of a generic confirmation, and points you to Personality when a file also describes how you want ADA to work.
• An upload that could not be saved no longer reports success.
• Your personality setting no longer shows the name of a preset you replaced with your own custom text.

Fixed

• Uploading a document now files it correctly. A single file is split into its separate parts, so an SOP becomes several documents plus any standalone tips, each listed and editable on its own, instead of the whole file being mistaken for a personality setting.
• Uploading a document can no longer change your personality setting. If a file describes how you want ADA to work, that part is offered as a suggestion rather than applied silently.
• ADA now draws on more of your saved knowledge in each reply.

Added

• You can now add documents to the knowledge base by uploading a file. Drop a PDF, Word, Excel, PowerPoint or text file onto the upload area on the Knowledge page, or click to browse. ADA reads the text out of it on your machine, so only the text is sent, and files up to 20 MB are accepted.

Fixed

• Tips and documents you save are now actually used when ADA answers you. They were being saved but never reaching it.

Fixed

• Asking what is in your workspace now lists your own documents only. ADA's internal folder, and other behind the scenes folders, no longer show up as if they were your files. You can still reach them by asking for them by name.

Fixed

• Asking for a spreadsheet "in Excel on my PC" now builds it in the app on your screen, instead of only handing back a copy to download.
• Spreadsheets are filled in completely. The whole model, every month, total and formula, goes straight into the sheet rather than a partly filled or jumbled table.
• Filling a sheet right after Excel opens now works, instead of falling back to typing the data in by hand.
• Stop works on every task. The Stop button now shows for follow up tasks and new tasks, not only the first one.
• A new task starts with a fresh list of steps instead of stacking them on top of the previous task's steps.
• "Start a new task instead" now works when you click it.
• Two tasks can no longer fight over your screen at once. Starting a new task on your PC cleanly takes over from the one before it.
• Saving a file at the end of a task no longer stops with an internal error.
• Spreadsheets shown in chat appear as a clean table instead of raw data and formatting codes.

Fixed

• The window resizes and moves properly now. Drag any edge or corner to resize it, and drag the top bar to move it around.
• Maximizing no longer covers the taskbar or sits on top of everything. It fills the screen up to the taskbar, and other windows can come in front of it.
• The document side panel now scales with the window instead of staying a fixed width.
• The task indicator no longer shows on the sign-in screen. It appears once you are signed in.

New

• Voice now does everything typing does. Speaking a request runs the same full assistant: it can research the web, use tools, and take actions, then show the results right in the chat, instead of only giving a short spoken answer. Long or detailed replies are spoken as a quick summary while the full result stays on screen.

Fixed

• Building multi-file projects works. Ask ADA to create a project and it can now make a folder and put files inside it in your workspace, instead of stopping with a permission message.
• Desktop tasks are more reliable. Steps that open or act on an app now finish cleanly instead of quietly retrying themselves.
• Switching between chats is cleaner. The unread-messages pill, Task mode, and the follow-up prompt no longer carry over from another conversation, so a new chat starts fresh.

Changed

• Spoken replies are being reworked, so the read-aloud toggle is temporarily hidden. It will return.

New

• ADA can show a document right inside the chat when it helps. Ask it to show a file (Word, spreadsheet, PDF, or slides) and it renders the real thing inline, with buttons to open the full view, save a copy, or keep it as a reusable artifact. ADA uses this for richer, research-style answers and keeps everyday replies plain.

Improved

• PDF previews now render reliably. Documents show as real pages everywhere, fixing cases where a PDF preview could appear blank.

A small follow-up release with three fixes.

Fixed

• Opening a chat now works from any page. If you switched to Task Logs or Workspaces and then clicked a conversation, it could open a blank new chat with the old task panel still showing. Now it opens the conversation you clicked and clears the panel.
• The app always closes now, even when your session has expired. Closing while signed out could show a backend error and leave the window stuck open; it now closes cleanly every time.
• Cleared the remaining em dashes from in-app messages. A few toasts, notifications, and error messages still used them; the wording now uses plain punctuation throughout.

This release makes ADA a real collaborator on your files inside chat. It can now look at what you have, show it to you, and write or edit files with your approval, one careful change at a time. It also brings the read-only "look at your computer" abilities that were prepared in 3.8.15, and fixes in-app updates and a spurious connection error when closing.

New

• Chat can look at your computer to answer you, read-only. Ask things like "what files are in my workspace?", "read the proposal and summarize it", "search my files for the pricing", or "what did we decide about X" and ADA looks at your active workspace's files, searches inside them, recalls what it remembers, and can check what is open on screen, then answers and cites the file it read. It only ever looks: it never changes, moves, or sends anything, and you see a short "Looking through your files" note with a small result chip in the reply.
• ADA can read your files and show them in the side panel. Ask it to open or read a file and it displays the real thing: Word, PDF, PowerPoint and Excel render as formatted pages, spreadsheets and CSV files show as tables, images show as pictures ADA can actually see and describe, and markdown shows as clean formatted text instead of raw symbols.
• ADA can write and edit files, and always asks first. When it wants to create or change a file you get a compact card showing exactly what will change (a before-and-after diff), with Allow, Always allow, and Cancel. Nothing is written until you approve, and every change has an Undo.
• Edits are surgical. ADA changes only the part that needs changing and leaves the rest of the file exactly as it was, so a small edit never rewrites or reformats your whole document.
• Precise spreadsheet edits. A change to a spreadsheet targets the exact sheet and cell you mean, instead of touching the wider file.
• File management from chat. Ask ADA to rename, move, copy, create folders, delete, zip, or unzip files and folders and it does it directly, quickly and reliably, without opening File Explorer.
• Multi-step file work in one go. ADA can read a file, work out the change, and write it back as a single flow, and it keeps talking to you after each approved step so you always know what it did and what is next.
• An auto-accept toggle in the workspace chip. Leave it off (the default) to approve every edit yourself, or turn it on for a workspace where you want ADA to apply changes without stopping to ask.

Improved

• File operations are now clearly labelled and colour-coded in chat: green for reads, yellow for new files, orange for moves and updates, red for deletes, so you can see at a glance what ADA touched.
• The write card is compact: it shows the file and its location and collapses to a single tidy row once you have decided, with Undo still one click away.
• ADA's replies and the app's wording no longer use em dashes.

Fixed

• File contents no longer leak into the chat as a wall of text when ADA reads a document.
• An approved edit can no longer be applied to a file that changed underneath it. If the file was modified since ADA read it, the edit is stopped rather than silently overwriting the newer version.
• Moves and renames are now undoable, and only one live change proposal is kept per file so old cards can no longer apply by mistake.
• Failed writes now show an honest failure chip instead of appearing to succeed, and the side panel no longer opens stale content.
• Closed several data-loss and consent gaps found in a full review of the file system, so ADA never changes a file without showing you first.
• Checking for updates now works reliably. The app re-checks for new versions on a schedule while it is open, and a brief network hiccup no longer stops it from finding an update until the next restart. Both the automatic check and the "Check Now" button pick up new releases.
• Closing the app no longer flashes a "Can't reach ADA" connection error. That message was the app noticing its own shutdown, and it is now silent when you close the window.

This release makes ADA a real collaborator on your files inside chat. It can now look at what you have, show it to you, and write or edit files with your approval, one careful change at a time. It also brings the read-only "look at your computer" abilities that were prepared in 3.8.15.

New

• Chat can look at your computer to answer you, read-only. Ask things like "what files are in my workspace?", "read the proposal and summarize it", "search my files for the pricing", or "what did we decide about X" and ADA looks at your active workspace's files, searches inside them, recalls what it remembers, and can check what is open on screen, then answers and cites the file it read. It only ever looks: it never changes, moves, or sends anything, and you see a short "Looking through your files" note with a small result chip in the reply.
• ADA can read your files and show them in the side panel. Ask it to open or read a file and it displays the real thing: Word, PDF, PowerPoint and Excel render as formatted pages, spreadsheets and CSV files show as tables, images show as pictures ADA can actually see and describe, and markdown shows as clean formatted text instead of raw symbols.
• ADA can write and edit files, and always asks first. When it wants to create or change a file you get a compact card showing exactly what will change (a before-and-after diff), with Allow, Always allow, and Cancel. Nothing is written until you approve, and every change has an Undo.
• Edits are surgical. ADA changes only the part that needs changing and leaves the rest of the file exactly as it was, so a small edit never rewrites or reformats your whole document.
• Precise spreadsheet edits. A change to a spreadsheet targets the exact sheet and cell you mean, instead of touching the wider file.
• File management from chat. Ask ADA to rename, move, copy, create folders, delete, zip, or unzip files and folders and it does it directly, quickly and reliably, without opening File Explorer.
• Multi-step file work in one go. ADA can read a file, work out the change, and write it back as a single flow, and it keeps talking to you after each approved step so you always know what it did and what is next.
• An auto-accept toggle in the workspace chip. Leave it off (the default) to approve every edit yourself, or turn it on for a workspace where you want ADA to apply changes without stopping to ask.

Improved

• File operations are now clearly labelled and colour-coded in chat: green for reads, yellow for new files, orange for moves and updates, red for deletes, so you can see at a glance what ADA touched.
• The write card is compact: it shows the file and its location and collapses to a single tidy row once you have decided, with Undo still one click away.
• ADA's replies and the app's wording no longer use em dashes.

Fixed

• File contents no longer leak into the chat as a wall of text when ADA reads a document.
• An approved edit can no longer be applied to a file that changed underneath it. If the file was modified since ADA read it, the edit is stopped rather than silently overwriting the newer version.
• Moves and renames are now undoable, and only one live change proposal is kept per file so old cards can no longer apply by mistake.
• Failed writes now show an honest failure chip instead of appearing to succeed, and the side panel no longer opens stale content.
• Closed several data-loss and consent gaps found in a full review of the file system, so ADA never changes a file without showing you first.

Fixed

• Light mode now applies everywhere. The live task activity panel (the step-by-step view of what ADA is doing) stayed dark even after switching to light mode; it now follows your theme like the rest of the app, with readable text and visible dividers.
• Document previews open properly again. Word, PDF, PowerPoint and Excel files that ADA creates now show a real in-app preview (a formatted page, a spreadsheet grid, a PDF, or slide cards) instead of a wall of plain text. If a live preview cannot load for any reason, you now get a clear notice with a one-click Export button rather than raw content.
• The workspace indicator at the bottom of the window updates the moment you switch workspaces from the Workspaces page, and after renaming or deleting one. Before, it only changed when you switched from the indicator's own menu and could show a stale name (which also meant tasks could run against the wrong workspace).

This release is a major upgrade to how ADA works in the browser: web tasks are far more reliable, noticeably faster, and can handle much more complex pages — plus ADA can now take and check screenshots of what it sees.

New

• Website screenshots: ADA captures and saves screenshots of the pages it browses on its own, not just pages in your connected browser — so a request like "take two screenshots of each of these sites" now saves real image files you can open. Every screenshot from a task is kept, so multi-shot requests deliver all of them.
• Screenshot verification: ADA looks at each saved screenshot to confirm it actually shows the right page, instead of just assuming the file is correct — a blank or wrong page is caught rather than passed off as done.
• Autocomplete and typeahead fields: ADA can type into a search or autocomplete box and pick the matching suggestion from the dropdown.
• Reaches inside embedded frames and modern web components: form fields and controls nested in iframes or shadow-DOM (common on component-based sites) are now visible and operable — previously ADA couldn't see them.

Improved

• More reliable in-page actions: clicks, typing, selections, checkboxes and hovers now target elements by their accessibility role instead of guessing pixel positions, so they land on the right control far more often.
• Stays on the correct tab: a task binds to the tab it started on and keeps acting there even if another tab or window takes focus.
• Much faster web and PC tasks: browser steps skip unnecessary screenshots, extra AI passes and fixed waits (pages are given just enough time to settle), multi-step on-page sequences run directly instead of taking detours, and PC steps skip redundant planning. Responses are quicker across the board on a faster underlying model.
• Handles more complex pages: longer multi-section forms and multi-step web flows are planned and completed more reliably, with a larger action budget for complex pages.
• Follow-ups that build on the last task: right after ADA finds or downloads a file, "install it" or "open it" acts on that exact file instead of searching for it again from scratch.
• Finding your files: ADA can locate files in your Downloads, Desktop and Documents even while a workspace is active, and "install the app I downloaded" finds the installer on your PC instead of looking it up online.

Fixed

• Checkbox and toggle actions are verified by their real checked state, and repeating a toggle no longer flips a control that is already in the state you asked for.
• A step that failed could previously derail the check at the end of a task; that no longer happens.

Improved

• Updates no longer take over the screen. When a new version is available you now see a small "Update" indicator in the status bar; click it to view the details and update when you choose, or carry on working.
• Updating shows live download progress, both in the status bar and on the update card.

Fixed

• The update action could silently do nothing in some cases. Installing an update is now reliable, shows clear progress, and surfaces a clear message if anything goes wrong.

New

• Mid-task steering: send follow-up guidance to ADA while a task is already running, and it adapts without restarting.
• Autonomous mode toggle: let ADA proceed through a task on its own, or keep it asking for confirmation.
• Mid-task pause and resume: ADA can pause to ask for help through the assistance panel, then continue from where it left off.
• Live task timer: every task shows elapsed time in hours, minutes and seconds, with per-step and final durations, in both the live trace and task history.

Improved

• Smarter document generation: a new typed document engine produces richer Word, PDF, PowerPoint and Excel files, including tables of contents, PDF outline navigation, Excel number formats, totals, charts and conditional formatting.
• More reliable file delivery: a request to save a file now always produces the file, and bare requests like "make a PDF" generate full content instead of a near-empty document.
• Clarify only when it matters: ADA asks focused questions up front when a request is ambiguous, and stays out of the way when it is not.

Fixed

• Task completion now shows a single, clean notice instead of a duplicate or truncated summary toast.

v3.6.0-beta: Major update spanning browser automation, research view, and desktop UI.

What's New

• Redesigned Memory tab with expandable items, sub-tab filtering (All/Patterns/Memories/Preferences), confidence indicators, and parallel API loading
• Fixed critical asyncpg timezone mismatch (TZDateTime TypeDecorator) that caused HTTP 500 on dialogue messages
• 20 security fixes across the copilot: subprocess injection hardening, path traversal prevention, HMAC constant-time comparison, token entropy upgrade, input validation on all API routes
• Improved browser automation grounding pipeline reliability
• Research view rendering improvements

Security

• All subprocess calls now use shell=False with executable whitelisting
• Path traversal protection on file-serving endpoints
• Constant-time token comparison (hmac.compare_digest)
• 64-byte local auth tokens (upgraded from 32)
• Screenshot and log file path validation
• Integer overflow protection on window resize API

Fixes

• Memory tab now displays patterns, memories, and preferences correctly
• Fixed early-return bug that silently killed Memory data loading
• Fixed LEARNED_TYPE_LABELS temporal dead zone in JS
• Orchestrator v1.3.6: TZDateTime across all 10 model files for PostgreSQL compatibility

Infrastructure

• Orchestrator test suite: 4905 passed, 91.44% coverage
• Copilot test suite: 457 passed
• Build system: added --channel flag for beta/stable release channels

• Fix telemetry propagation across execution pipeline
• Improve tactical executor prompts with cognitive context enrichment
• Fix job storage batch operations and state management
• Fix react executor iteration tracking consistency
• Minor stability improvements

v3.1.6 — Research rendering overhaul, device agent connection reliability

• Rewrote renderMarkdown: proper heading rendering (## / ###), asterisk list support (* items), code blocks with copy button
• Inline citations now show source favicon icons with hover tooltips showing full reference title and clickable URL
• Sources section redesigned as a clean stacked list with favicons (replaces heavy card layout)
• Fixed device agent false-negative status: connection shows Connected reliably on app startup with existing session
• Added dual-path status sync: HTTP poll + WebSocket re-push when frontend connects, eliminating race condition
• Fixed exponential backoff stalling: no-token retries use fixed 2s delay instead of growing to 30s
• VLM hallucination guards: refuse text-only vision calls, return assistance instead of fabricated screen data
• Goal extraction validation: reject garbage summaries, synthesize goals from step descriptions
• VLM phase subgoals replace tool-level checklist for clearer user-facing progress
• Screenshot attachment to execution timeline perceive_reason events