2 November: the agent can work inside Windows VMs, and new VMs answer SSH sooner
The chat agent can now run PowerShell commands and read or write files inside your Windows VM over SSH. New VMs reach SSH in minutes, and a prepared machine makes it faster still.
If you use Windows VMs, the agent can now work in them, and new VMs become usable sooner. Without VMs nothing changes. No action needed.
Until 2 November the agent could create a Windows VM and then do nothing with it. That gap is closed.
Text version
Agent tools vm_exec, vm_write_file, vm_read_file over SSH. First start measured on test VMs. Network fixes: DNS for guests, port forwards removed on delete, worker choice without an image.
What the agent can do
The agent has three tools for the VM selected in the chat: vm_exec for PowerShell, vm_write_file and vm_read_file. Each VM gets its own key, and its host key is pinned on first contact. One held connection per VM answers a command in about 8 ms at the median, measured on test VMs, instead of about 2 seconds.
Starting sooner
On 2 November we measured test VMs: SSH answered after about 345 seconds, where before it often did not answer at all. Each worker now keeps one prepared Windows VM (4 GB RAM per worker, approved). A new VM takes a prepared one when it is ready, so SSH answers at once. Another size takes about a minute longer. If anything fails, the normal path runs.
Fixes behind it
Guests got DNS servers they could not reach, so they could not fetch OpenSSH, Tailscale or updates. They now get DNS from the worker. Deleting a VM left its port forward behind, which made the next VM on that port unreachable. That is fixed, and the forwards of running VMs were set again. A VM created without an image now picks a worker that has the image.
These numbers come from our own tests on test VMs. Your first start can differ.