Vestra Code 1.16

#Background runs

Type a task, open the + menu and choose Run in background. The task goes to a Vestra cloud machine and keeps going when you close Vestra Code or turn your computer off.

• Background in the sidebar lists every run on your account, with its status, log and changes, and a Stop button. The same list is in the Vestra web app, so you can check on a run from your phone.
• When a run needs your approval, it pauses and tells you, on the desktop or on your phone. Approve or deny with a note from either.
• Nothing is written into your folder until you’ve looked. A review screen lets you accept everything, pick files or single changes, edit a file first, make a branch, or discard it. Accepting is one step of Undo.
• Schedules run a task for you, such as “every weekday at 07:00, update the dependencies and run the tests”. Up to five, at most hourly, with an optional email when a run finishes or needs you.

Background runs use your plan’s cloud machine hours: 10 a month on Plus, 40 on Max 10x and 100 on Max 20x.

#Faster

The biggest change is one you won’t see. Vestra Code thinks harder when it plans and lighter on routine steps. Until now, each change of thinking level made our model server read the whole conversation again from the start: 34.6 seconds for a 7,356-token chat in our test, and minutes for long ones. It now keeps what it has already read, so the same step takes 0.26 seconds. A chat you continue the next day no longer starts from scratch either.

Prompts are smaller too. A new request now comes with a short map of the files and functions that match what you asked for, so the agent goes straight to them instead of searching first. In our test task that halved the number of requests and cut the prompt from 15,690 tokens to 4,660. Long files arrive as an outline first, and the agent reads only the parts it needs.

When Vestra is busy, you’re no longer left with a silent wait. The status line says 2nd in line · about 40 s, then Reading your prompt · 41% until the first words arrive. The estimates come from the server’s own measurements; when there isn’t enough to go on, you see your place without a time rather than a guess.

Plan first, in the + menu, answers each message with a short plan you can edit, reorder and approve before anything changes.

#Verified Fix

Verified Fix reviews your branch, your uncommitted changes or up to 40 files you pick, and looks for security problems. For up to five findings rated medium or higher, it writes a fix and a harmless exploit test that uses only marker files and canary values. Then it checks each fix on a cloud machine with the network switched off:

• the exploit test has to fail on your code as it is;
• it then has to pass with the fix;
• and your own tests have to keep passing.

Each finding ends with one of four results: Fix verified against its exploit test, Fix incomplete, Fix breaks tests or Couldn’t verify, with the reason. Verified fixes go through the usual approval cards, with Undo. One that wasn’t verified needs Apply anyway.

We built it this way because of what we saw in our own testing. Fixes that read well don’t always hold up: once they had to apply and pass tests, some broke features the app’s tests covered, and one closed the obvious path while leaving a variant open. Reading a diff doesn’t catch that. Running it does.

Verified Fix needs access to Cyber or Cyber Mini, granted through Vestra Enclave. Cyber Mini reviews by default; Second opinion asks the other model to look too. It supports Node.js and Python projects. A run takes about one to two hours and costs about £0.10 to £0.20, and you can do three a day. Open it from Ctrl+K or the shield on the Changes panel.

#Smoother

Streaming replies no longer make the window stutter. Text now appears once per screen refresh, and only the paragraph still being written is redrawn. In our test, a 1,921-word answer with lists and code dropped 1,522 frames before and 5 after. The words arrive from the server at the same speed either way; before, the window couldn’t keep up, so the answer took 30.2 seconds to appear instead of 6.4.

• Auto-scroll glides at the bottom of a chat, and stays put the moment you scroll up.
• The effort slider follows your pointer exactly and settles on a level with a soft spring. Fast now shows a warp starfield, and the Ultra stars drift slowly.
• The side panel and sidebar slide without re-wrapping the chat on every frame. With the CPU slowed four times, opening and closing the side panel six times dropped 72 frames before and 1 after.
• Menus and dialogs can be closed while they’re still opening, and long chats scroll faster.

If “Show animations” is off in Windows, nothing slides and the slider jumps straight to a level.

Motion figures come from a test that runs the app headless at 60 Hz on a machine without a GPU, comparing 1.16 with 1.15.5.

#Also in 1.16

• Memory and commands. Start a message with #, or use “Remember this”, to save a note for this project, for the team in VESTRA.md, or for every project. The agent can offer to remember things, always asks first and never saves anything that looks like a secret. Custom slash commands take arguments and can be shared through git in .vestra/commands/.
• GitHub pull requests. Push & open PR from the Changes panel, with a title and description the model can draft for you to edit. It uses the GitHub CLI or a fine-grained token, and the secret scan still runs on the push. Diffs can now be shown side by side.
• Swarm isolation. Helpers can work on their own branch or their own cloud machine, with the cost shown before they start. Their work reaches your project only after you choose Merge, Keep branch or Discard.
• Hosting. With a server attached, /deploy shows the plan first, uploads only changed files, never uploads secrets, and backs up anything it replaces. /logs follows a log live, and Changed files lists everything changed on the server, with Restore backup. Every deploy, log and restart is confirmed in a dialog that names the server. Start, stop and restart for game servers are rolling out.
• Autonomous auto-approve. An optional switch, off by default, that lets chats in Autonomous mode approve every prompt by themselves, from deploys to git pushes. You’re asked again in a system dialog after each update, and everything it approved is listed in the run’s “While you were away” report. The hard limits still apply whatever it says: the secret scan, the project-folder boundary, and the .git and symlink guards. Anything the agent reads, such as a web page or a file, could steer it into a risky action, so only use it on projects you trust.
• Security. ssh, scp, sftp and rsync to a live server follow the same LIVE rules as the hosting tools, in every mode. Hosting keys are revoked when you disconnect or sign out, signing in again is quicker, and the installer keeps its files in a private folder.

#Get it

Vestra Code is for Windows and comes with paid VestraAI plans. Sign in at ai.vestra.systems and open Vestra Code from your account menu. If you already have it, it updates itself.

Every change is listed in the changelog. Questions, or a fix that didn’t hold up? Write to [email protected] or find us on Discord.