AI Bridge now connects to the model you actually chose, changes provider without a restart, and an ending connection no longer takes the app down with it. FIXED: THE BRIDGE ALWAYS USED OLLAMA'S PORT • Whichever provider was selected, every workspace forwarded its AI Bridge traffic to port 11434. If you run MLX or llama-server on 8080, requests from the workspace arrived at a port with nothing listening on it. The bridge now forwards to the port the selected provider actually serves. • "Default Provider for New Workspaces" in Settings changed nothing: every workspace was created for Ollama regardless of it. It now decides what a new workspace starts with. • The Review screen of the New Workspace assistant named Ollama for every workspace. It now names the provider the workspace will really use. FIXED: CHOOSING AND CHANGING THE PROVIDER • Show Advanced Settings in the New Workspace assistant now offers Host Provider beside the AI Bridge checkbox. For a disposable workspace, deleted as soon as it stops, that is the only moment it can be set. • Each workspace's AI Bridge tab has the same control, showing the address on your host that its traffic is forwarded to, so a workspace created earlier is no longer tied to the provider it was made with. • Changing it applies while the workspace keeps running — no restart, so even a disposable one can be switched. Re-run the service command under "In the Linux Workspace" to move the workspace's own forwarder to the new port; those commands now reload and restart that service instead of leaving the old one running. • Endpoint Details and the environment variables follow the provider you pick. "Enable AI Bridge by Default" in Settings was read by nothing; picking the AI Sandbox profile now turns the bridge on when it is set. FIXED: UNEXPECTED QUIT DURING AI BRIDGE TRAFFIC • A bridge connection that closed while data was still in flight could terminate Velo Workspaces, taking every running workspace with it. In everyday AI agent use it was rare: a thousand consecutive streamed completions never triggered it, because an orderly response leaves nothing in flight when the connection closes. It took a cancelled transfer — a generation stopped mid-stream, a client timing out — or one saturating the channel, such as a throughput benchmark, which triggered it within seconds. Fixed. • A long streaming response is no longer cut short when the program reading it pauses. A slow reader is treated as slow, not as a broken connection. FIXED: TYPED VALUES IN ADVANCED SETTINGS • A CPU, memory or storage figure typed with the cursor still in the box was discarded when you pressed Continue, and the workspace was created with the previous number. Typed figures now count as they are typed — in the assistant, and in a workspace's Resources tab and custom profiles. LOWER LATENCY • On its connection to the inference server the bridge no longer lets the system hold small packets back waiting for an acknowledgement, and the suggested Linux forwarder command does the same, so each piece of a streamed response is passed on as it arrives. CLEARER DIAGNOSTICS • The log now names both the channel port a workspace dials and the host port its traffic is forwarded to, so a mismatch between the two is visible at a glance. • Every connection reports the bytes it carried in each direction and why it ended. Ordinary endings — a client that closed, a transfer that finished — are no longer written down as failures. • The bridge listener is installed before the workspace starts, so its status can no longer be reported before it exists. Thank you to everyone who sent logs and reproduction steps.