November 14, 2026

New

• Terminals the AI keeps: the AI can keep a named terminal for a long job (a dev server, a watcher), run one command after another ends, run the same command again after a fix, and read any part of a command's output without running it again. It answers a command that asks a question instead of getting stuck, and it is told when a command it didn't wait for ends
• Take over an AI terminal: Open live terminal on a command shows the AI's terminal in the right panel with a bar that says who has it, and Take over, Hand back and Stop. The AI is told when you type or take control, and leaves the terminal alone until you hand it back
• Command cards, redrawn: one row per command with its state (Queued, Running, Needs your answer, Finished, Failed, Stopped, Timed out), how long it took, and Stop, Run again and Copy output. Output keeps its colours, links in it open, and long output has Show all lines
• Terminals list above the composer: each chat lists its running commands with state, elapsed time or the server's address, Open live terminal and Stop. Commands, cards and the list survive a window reload
• The AI picks its shell on Windows: the AI is told which shells are installed (Windows PowerShell, PowerShell 7, Command Prompt, Git Bash, WSL) and chooses one per command, each with its own terminal, so a command written for bash no longer runs in PowerShell
• Pick your shell: in the agents view's right panel, the Terminal row and the + menu list your installed shells; pick one for a new terminal, or pin one as the default. A reopened tab comes back with its shell
• Spell check: text fields underline misspelled words. Right-click one for the words you most likely meant, and Add to dictionary for names and terms. The code editor, terminal and search boxes are left alone
• Auto-correct (optional): Settings > General > Auto-correct fixes clear typos as you type ("heiio this is nicrt" becomes "hello this is nice"). Backspace right after a correction puts back what you typed and leaves that word alone from then on. Real words, code, file names and mentions are never changed. Off by default; English only
• Arabic and right to left: Arabic text uses its own typeface across the app, text fields follow the direction of what you type, and a plan written in Arabic is laid out right to left
• Import sign-ins from another browser (macOS): the built-in browser can bring cookies and history from Chrome, Arc, Dia, Brave, Edge, Vivaldi, Opera, Chromium, Firefox and Zen, for all sites, the current site, or the ones you pick. It reads those browsers' data on your Mac (macOS asks for your Keychain) and nothing leaves the computer
• Send sign-ins to your cloud computer: sign Chrome on the cloud computer in to your sites, from the Velork browser or another browser on your Mac, for all sites or chosen ones
• Subagents, redrawn: each subagent is a row with its task, a live line of what it is doing, elapsed time and status; agents started together show as one card. Opening one slides its chat over the conversation with a back button and Stop
• Plans, redrawn: the plan card shows the title, the why and numbered steps that tick off while it is built. A bar above the composer shows Ready to build with Open plan and Build, then the current step while building
• Comment on a plan: select any text in the plan, press Comment and tell the AI what to change. Notes can be sent together and show their status
• Debug mode, rebuilt: a bar above the composer follows the investigation (Reproduce the issue, Investigating, Check the fix, Fixed) with the live log count, and buttons for your turn
• Mode switch asked in the chat: when the AI wants to change mode it asks on a card with the reason, and waits for you; no more automatic switch after 15 seconds
• Review changes setting: Settings > General > Review changes shows each answer's changed files in a card with Keep all, Undo all and per-file Keep and Undo. Off by default: changes are kept when the turn ends and no card is shown. The separate review bar above the composer is gone
• Editor settings in the agents view: Editor settings and Keyboard shortcuts open as tabs in the right panel instead of leaving for the IDE
• Tables in answers: columns size to their content, numbers align, and right-click copies a cell, a row, a column or the whole table
• Logo design skill: the AI follows a logo method when a design needs a logo (a mark and wordmark built from paths, checked at small sizes)

Improvements

• Changed files, easier to read: diffs are syntax-coloured, an edit marks the exact words that changed, and a new file shows as clean code instead of a wall of green
• Agents view editor defaults: files in the right panel open at 14px with wrapped lines and no minimap. Anything you set yourself in Editor settings still wins
• Canvas opens when the AI designs: the canvas tab comes forward when the AI starts changing a design, once per turn
• Faster @ menu: file results come from an index built once and narrow as you type
• Cloud computer tab: a More menu with Stop Cloud Computer and Send Sign-ins to Chrome
• Clear Site Data clears the whole site: it removes the parent domain's cookies too, so you are really signed out
• More provider logos: many more AI providers show their own logo in the model list
• Large pages pan smoother: on pages with tens of thousands of layers, panning no longer redraws everything when a text settles its size
• AI design output: new frames start with no fill and no clipping, padding or gap on a frame turns on auto layout, and the AI is told when text overflows its box
• Follow the system theme: with it on, dark is always Velork Dark and light is always Velork Light; picking another theme turns it off so your pick sticks
• Shorter settings: the long explanations on the Permissions and Cloud computer pages are cut to one line each

Fixes

• Windows: commands run as typed: on Windows PowerShell the first letter of a command could be lost ("npm" ran as "pm"), a failed command was reported as a success, and an idle prompt was mistaken for a question. All three are fixed, and scripts such as npm no longer fail on the execution policy
• Windows: files: the AI's file tools handle drive-letter paths, Windows line endings and files saved by PowerShell in UTF-16, and keep a file's encoding when they edit it
• Windows: sign-in and tools: the Claude sign-in link opens whole instead of cut off, and a notice offers Open Link and Copy Link if the browser didn't open. Pasting a Windows path in the built-in browser opens the file
• Velork AI credits stay visible: connecting your own AI no longer hides Velork AI and your remaining credits from the model list while you still have credits
• Images sent to Codex: an image attached to a Codex chat no longer fails the request
• Open file stays in the agents view: opening a file from the changed-files card, and Edit as Markdown on a plan, open in the right panel instead of jumping to the IDE
• Instructions save: when your instructions are saved on this computer but could not sync to your account, it says so instead of "Failed to save"
• Text stuck at its first width: centred or right-aligned text made in Velork resizes to its content again
• Text styles keep bold and strikethrough: applying a text style no longer wipes the styling of individual ranges
• Scaled instances: an instance placed at a smaller scale takes the right room in auto layout, and no longer freezes after one resize
• Answers hidden while a server ran: text the AI wrote after starting a dev server showed only once the server stopped; it shows right away
• Claude errors no longer vanish: a usage limit or an outage at the end of a turn is shown, instead of the answer silently stopping
• "Tools unavailable" after a restart: a resumed Claude chat no longer claims Velork's design tools are disconnected
• No "save workspace?" prompt: closing a window whose chats added project folders no longer asks to save an untitled workspace
• Debug logs go to the right place: debug mode's logs are kept per chat, so they no longer land in another window's project