VS Code 1.145 Stable Adds Remote Agent Previews, MCP Policy Changes and files.exclude Exceptions

The page name has been a little confusing. The same official URL has also been indexed as "Visual Studio Code 1.145 (Insiders)". Microsoft's archive listed it that way while it was the Insiders changelog, which cover the Insiders build of VS Code and continue to evolve as new features are added. The page now shows the Stable label. An update-tracking project on GitHub has also already seen the detected version (1.146.0-insider), so the Insiders channel has moved on to the next version.

Weekly releases now​

If you still think of VS Code as a monthly product, you're behind. Neowin described 1.145 as its latest weekly release, and the release notes are written in a "this week" style. For Windows admins who manage VS Code across many machines, a weekly cadence means release notes need regular reading, not a quarterly skim. Several changes this week affect managed settings and how policy works.

Section summary: 1.145 is the current Stable build. The "(Insiders)" label is left over from the pre-release changelog, and Insiders has already moved to 1.146.

Remote agents: preview your app where it runs (Preview)​

The headline feature follows the release's own pitch that remote agent sessions should "feel a bit more local." Neowin summarised it this way: when an agent session runs on a remote agent host, the integrated browser in the Agents window now loads pages through that host as it already does in remote workspaces. If an agent starts a dev server on a remote machine, you can now open it in the Agents window to preview the app, validate the agent's changes, and give it feedback.

Canvases come along too. These are the interactive visual workspaces the agent can host. Both of these features are currently in preview.

To turn them on:

1. Set workbench.browser.enableRemoteProxy to true for remote browsing.
2. For remote canvases, also set chat.canvases.enabled to true.
3. Report problems in the VS Code repository. Microsoft explicitly asks for that feedback on these previews.

Another Experimental feature uses chat.canvases.enabled: canvas pinning. Open a canvas, click Add to Sidebar above it, and it stays in the left sidebar when you move to other views.

Mixing harnesses in one session (Preview)​

Before this release, a nested session always used the same agent harness as its parent. With chat.agentHost.multipleChatProviders.enabled turned on, nested sessions can now use a different harness. Microsoft's example is asking Claude to review a plan that a Copilot chat wrote, or giving Codex the same task to run next to it, all under one main session.

What carries over and what doesn't:

• Each chat has its own model, agent and permission levels, set by its harness. Changing one chat leaves the others alone.
• A chat you move keeps its history and its harness. A forked nested session keeps its configuration.
• If a chat uses a different harness from its parent session, hovering over its tab or its row in the sessions list shows which harness it uses.

Getting a second opinion from another model is a sensible review practice. Keep in mind that each harness has its own permission level, though. A session that combines three harnesses can also have three different permission setups.

Cron comes to local automations​

Automations let agent tasks run on demand or on a schedule. In 1.145 you can select Cron to enter a five-field cron expression when the other schedule presets aren't precise enough. The dialog checks the expression and describes the schedule it produces in plain language. One limit matters: Cron schedules are not available for cloud automations.

Two smaller changes in the same area:

• When you create or edit a local automation, you can pick a custom agent from the selected workspace.
• Disabled automations move into a collapsed section, so they no longer clutter run history and active jobs.

Microsoft doesn't spell out the syntax in the release notes. The usual five fields are minute, hour, day of month, month and day of week. Microsoft says the dialog validates whatever you type.

Copilot harness, working set and sandboxing​

The Copilot harness is still the default for new sessions. Microsoft says it is built on the Copilot SDK, runs through the Agent Host, and can keep running in the background, reachable from both the editor window and the Agents window. A new Local-to-Copilot harness transition guide covers customization compatibility, session continuity, and organization rollout and verification. Microsoft is also asking organizations to file issues if something stops them from moving off the Local harness. Read that as a strong hint about where development effort is going.

The Experimental working set (chat.editor.workingSet.enabled) handles reviewing pending agent edits in Copilot harness sessions. It gets two fixes Windows developers will care about:

• Files in WSL, Dev Containers and SSH workspaces now open from the working set on the correct host.
• When several tool calls edit the same file in parallel, the edits are tracked, and the inline review controls stay available after editing finishes.

Sandboxing errors are also clearer. If your organization requires sandboxing through managed settings that disallow bypass, and the machine can't run the sandbox, the warning banner now says why terminal commands are blocked. It names the specific cause, such as a missing dependency, and links to troubleshooting steps. A GitHub pull request from Microsoft's Ross Wollman documenting these release-note entries ties the sandbox change to microsoft/vscode#344815.

Less noise in chat​

The Experimental verbosity control (chat.persistentProgressVerbosity) is in the Chat view's More Actions menu under View Mode:

You can switch modes in the middle of a chat. Only the display changes; the agent's context and work stay the same. If you'd rather have an animated input border than a persistent progress indicator, enable chat.progressBorder.enabled.

Microsoft also lists performance work for long Copilot harness chats. Existing reasoning titles are reused, and terminal output in restored tool calls is now rendered only when you expand it, instead of all at once.

Two attachment changes:

• Inline attachments (chat.inlineAttachments.enabled, Experimental) put attachment pills at the cursor inside the prompt. They're on by default in Insiders and off by default in Stable.
• Add Context (+) > Files now opens the standard file dialog, so you can attach files from anywhere on your PC, even in a new chat with no workspace. In cloud sessions the chooser lists only images.

The security details matter here. In Copilot harness sessions with sandboxing on, the agent can read attached paths outside the workspace for the rest of that session. It can't write to them, and the access isn't saved to settings. With Manual permissions, the agent no longer asks before reading an attached file. Attachments are not added to an organization's managed list of readable paths.

Enterprise: MCP allowlists now take precedence over registry-only access​

This is the change admins should read twice. Before, if an organization limited MCP to registry servers and defined a managed allowlist, every server still had to be in the registry. Now the enterprise-managed allowedMcpServers list takes priority:

• A server that matches the list can run even if it isn't in the registry. A registry server that doesn't match is blocked.
• An empty list ([]) blocks all servers. If the list is absent, the registry restriction works as before.
• User and workspace allowlists can't add entries to the enterprise list.
• deniedMcpServers and turning MCP off still override the allowlist.
• Microsoft recommends serverUrl and serverCommand rules, because a serverName rule matches a label the server's own author chooses.

The scope is narrow: this applies only to MCP servers that VS Code starts, not to servers managed by the Agent Host.

Also on the admin side, plugins required by managed settings such as enabledPlugins no longer show an Uninstall button that always failed. The action is now disabled, with an explanation on hover and keyboard focus.

Two policy fixes stand out. First, a failed GitHub Copilot policy refresh, caused by a network error or rate limiting for example, used to drop a registry-only MCP restriction until the next successful refresh. VS Code now keeps the last known policy. Second, users who had picked Claude or Codex before an organization disabled it used to get a read-only, input-less New Chat. New chats now fall back to the default harness, and any unsent prompt carries over. The first fix closes a real gap: a policy that disappears during a network blip isn't much of a control.

Editor fixes: files.exclude now supports ! exceptions​

You can now put ! in front of a pattern in files.exclude to keep part of an excluded folder visible in Explorer, Search and Quick Open, even if another pattern or .gitignore excludes it:

"files.exclude": {
  "node_modules": true,
  "!node_modules/@types": true
}

Exceptions in search.exclude apply only to Search and Quick Open. Exception patterns never hide files. Deep patterns like !**/node_modules/@types work at any depth but can slow searches down.

In Mermaid diagrams, edges no longer cover subgraph titles, and the default vscode theme now outlines subgraphs instead of filling them.

Extensions and API​

• GitHub Pull Requests 0.176.0: the overview timeline shows more events, including label and title changes, force pushes and merge-queue updates. Description templates support ${issueTitle}, ${issueNumber} and ${issueNumberLabel}. When you stack pull requests, you're prompted to disable auto-merge on any PR in the stack that has it enabled.
• Timeline API finalized: Marketplace extensions can register providers with workspace.registerTimelineProvider or workspace.registerWorkspaceTimelineProvider. Setting TimelineItem.resourceUri makes Compare Selected work across entries.

Tools removed from the Local harness: check your prompt files​

Three agent tools are removed from Local agent sessions:

1. #installExtension: agents now add recommended extensions to .vscode/extensions.json for you to review and install.
2. #runCommand: built-in workflows no longer depend on it. Network diagnostics now ask you to run Developer: Chat Diagnostics and share the report.
3. #newWorkspace (formerly #new) and /new: to start a project, open an empty folder and describe what you want. The Welcome page's Generate New Workspace link is also gone.

Remove these references, plus project-setup-info-local and project-setup-info-context7, from your custom agents and prompt files. The settings github.copilot.chat.newWorkspaceCreation.enabled, github.copilot.chat.newWorkspace.useContext7 and github.copilot.chat.useProjectTemplates no longer exist. Notebook creation is unchanged, and no upcoming deprecations are listed.

And Blobby​

Microsoft also shipped a mini-game. Type /vscode-pet in chat. Then, on the Agents window's new-session screen, choose Play Jump Game from Blobby's context menu. Hold Space to charge a jump and use the arrow keys to aim. Your high score is saved on the device. NTCompatible said Microsoft frames the release around five features including, oddly enough, a Blobby jump game.

Bottom line​

For individual developers, 1.145 mostly makes remote agent work and long chats easier to manage. For Windows shops running VS Code fleets, the parts that matter are the new MCP allowlist precedence, the fix that keeps registry-only policy in place when a refresh fails, and the tools removed from the Local harness, which could quietly break existing prompt files. Most of the new features are still Preview or Experimental, so test them before rolling them out. If you already have VS Code installed, use Check for Updates to get 1.145.