VS Code 1.142: Copilot Agent Host Policies, Windows Sandbox Retirement and Markdown Upgrades

If you have been watching VS Code gradually evolve from a lightweight text editor into an orchestration hub for autonomous coding agents, this release marks the moment where the training wheels come off—and where IT administrators get the policy brakes they have been asking for.

Agent Workflows: The Copilot Harness and Interactive Canvases​

For the past several cycles, VS Code has experimented with how agent sessions are hosted and executed. In version 1.142, Microsoft is progressively rolling out the Copilot harness as the default for new sessions. Built on the shared Copilot SDK and running through the Agent Host, the Copilot harness lets sessions keep chugging along in the background while remaining accessible from both the standard VS Code editor window and the dedicated Agents window.

When you spin up a new chat session, the harness picker in the chat input now clearly labels Copilot when the Copilot harness is active, and chat titles explicitly display the active harness name (such as Copilot or Codex) so you do not have to go hunting through menus to figure out which runtime engine is handling your code. Responding to developer feedback that sometimes you just want to ask a question without giving an agent keys to modify your workspace, Microsoft has also reintroduced Ask as a built-in, dedicated read-only agent inside the Copilot harness.

Canvases (Preview): Moving Beyond the Chat Prompt​

Anyone who has tried to coordinate a multi-step refactor purely through a scrolling text chat knows the frustration: the context gets buried, task lists scroll out of sight, and you spend half your tokens reminding the agent what step three was.

VS Code 1.142 introduces Canvases (Preview), controlled by the chat.canvases.enabled setting. A canvas is a visual, interactive workspace that opens right beside the chat pane, allowing both you and the agent to inspect and manipulate the exact same artifact in real time. Microsoft highlights a Kanban board as a practical example: you can drag cards around or create new items manually, while the agent breaks a feature down into discrete tasks, updates ticket statuses as it writes code, or logs newly discovered bugs directly onto the board.

Before you try to launch one, keep the current preview boundaries in mind:

• Platform and Session Scope: Canvases are available in Preview in the VS Code desktop app specifically for local sessions that use the Copilot harness.
• How to Enable: Turn on chat.canvases.enabled in your settings and start a fresh session using the Copilot harness.
• Extension-Driven Architecture: Every canvas is powered by a canvas extension (which also work in the GitHub Copilot app and across both the VS Code editor and Agents windows). You can get started in two ways:
  1. Type /create-canvas directly into the chat input and describe the custom UI you want for your project; the agent skill will scaffold a canvas extension right inside your workspace and open it.
  2. Install a chat plugin that bundles a canvas extension via the GitHub Copilot app, then switch back to VS Code and prompt the agent to list and open available canvases. Once installed, canvases can also be launched in the Agents window via Add Tab (+) > Canvas.

Moving Chats Across Folders and Git Worktrees​

Have you ever started diagnosing a bug in your main repository checkout, realized the fix is going to get messy, and wished you could teleport the entire AI conversation into a clean Git worktree without starting over? Previously, switching directories or spinning up a worktree meant opening a fresh chat and painstakingly rebuilding your context from scratch.

In the Agents window, VS Code 1.142 allows you to ask the agent in plain natural language to relocate your active chat to another folder or into a brand-new Git worktree while keeping the entire conversation history intact. For instance, you can prompt:

Continue this task in a new Git worktree of the current repository before implementing it.

Click to expand...

When you issue that request, the agent finishes its current turn, migrates the chat, posts a workspace-change notice, and automatically resumes work in the new directory. All subsequent file edits and terminal commands execute in the new worktree, leaving your original checkout untouched.

Important Prerequisite Check: Do not expect this to work in every conceivable configuration just yet. Moving a chat between folders or worktrees is currently scoped to agent sessions running the Codex harness or Codex in Interactive mode, and the session must consist of a single chat and a single local folder.

Taming the Agents Window Layout​

If you live inside the Agents window, switching between different sessions used to be jarring because VS Code shared editor visibility, details, and panel visibility across all sessions. A new experimental setting, sessions.experimental.layoutScope (which requires a window reload after changing), lets you dictate how much layout state each session remembers:

(Note: Overall window geometry, pane dimensions, and sidebar visibility still remain shared across the window regardless of which option you pick.)

Additionally, you can now customize how the bottom panel sits in the Agents window by right-clicking a panel tab and selecting Align Panel. Choosing Center tucks the panel neatly beneath the chat area to free up vertical room for the side pane, while Justify stretches the panel across the full width of the window. VS Code saves this choice per profile and remembers it when you maximize or restore the panel.

Background Shell Control, Model Picker, and a Critical Windows Sandbox Change​

Stopping Runaway Background Shells​

When an agent kicks off a background command—say, a test suite that hangs or a watch script you no longer need—you no longer have to wait for it to time out. In Copilot harness sessions, a Background Shells pill sits above the chat input showing active background commands. Clicking the pill lets you select a specific shell, view its full description and untruncated scrollable output history, and click Stop Shell to terminate just that command. The control is fully keyboard-accessible and announces termination to screen readers. Keep in mind this only applies to attached shells tied to the agent's current request; detached shells running independently cannot be killed from this UI.

Redesigned Model Picker and BYOK Tweaks​

With developers juggling first-party GitHub Copilot models alongside Bring-Your-Own-Key (BYOK) custom endpoints, the model selector had grown crowded. Version 1.142 introduces a tabbed model picker that cleanly separates Copilot models from BYOK models. Hovering over a model in the list now exposes inline controls to adjust its thinking/effort level and context window size on the fly.

For developers connecting custom BYOK models that rely on a stateless Responses API, VS Code 1.142 adds a targeted fix: setting "statefulResponses": false on the model's entry inside chatLanguageModels.json forces VS Code to transmit the full conversation history with every request.

Windows Sandboxing Deprecated for Local Harness​

For Windows users and security-conscious developers, there is a vital architectural change buried in the release notes: Windows sandboxing for terminal commands in Local agent sessions is deprecated and no longer supported. If you require sandboxed terminal command execution on Windows, you must now run your sessions through an Agent Host session rather than the Local harness.

Enterprise Governance: Enforcing Agent Host and Granular Permissions​

That Windows sandboxing shift ties directly into a massive governance upgrade for enterprise IT administrators. In prior builds, enforcing runtime sandboxing or strict tool permissions was complicated if developers stayed on the Local harness.

In VS Code 1.142, whenever an organization configures a nonempty permissions object or sets sandbox.enabled to true in unified managed settings, VS Code now requires the Copilot harness in Agent Host for full chat sessions.

Here is what happens on the client machine when those policies are active:

• Automatic Chat Migration: When a developer opens an existing, unarchived Local chat, VS Code automatically migrates it to the Copilot harness, preserving the full conversation transcript and any unsent draft (including attachments) without automatically sending the draft. The original Local chat is archived only after the new session successfully opens; if the user switches away mid-migration, the move aborts cleanly without altering the original chat.
• Restricted Harnesses: While these policies are enforced, Claude, Codex, and terminal chat are disabled. Editor inline chat remains accessible for scoped read and edit actions. (An empty permissions object or legacy policy alone will not trigger this migration.)
• Temporary Admin Escape Hatch: If your organization needs breathing room before moving everyone to Copilot Agent Host, admins can set a temporary override in unified managed settings:
  
  Code:

  {
    "vscode": {
      "temporaryDefaultToLocalHarness": true
    }
  }

  Security Warning for Admins: Enabling temporaryDefaultToLocalHarness stops automatic migration and keeps the Local harness available, but it also overrides the routing required by managed sandbox and permission settings. Because the Local harness does not enforce those runtime-only rules, turning this flag on means Local sessions will run without those sandbox and permission protections (though enforcement inside Copilot Agent Host sessions remains intact).

Separating "Assisted" from "Allow All" Permissions​

Enterprise admins no longer have to choose between making developers approve every single breath an agent takes and giving the agent a blank check. Administrators can now allow Assisted permissions mode while blocking Allow All (disableBypassPermissionsMode), using settings like:

{
  "permissions": {
    "disableBypassPermissionsMode": "disable",
    "disableAssistedPermissionsMode": false
  }
}

A few crucial policy rules to note for 1.142 deployments:

1. permissions.defaultMode only sets the initial starting mode; it does not restrict which modes a developer can manually switch to in the UI.
2. The legacy ChatToolsAutoApprove: false policy still blocks both Allow All and Assisted unless you configure explicit managed mode restrictions.
3. Current Release Compatibility Caveat: In VS Code 1.142, Assisted Permissions is not compatible with granular permissions.allow, permissions.ask, or permissions.deny rules. If your organization uses any of those three rules, Microsoft advises setting "disableAssistedPermissionsMode": true for now; side-by-side support for both is slated for the next release.
4. On Windows systems, numeric enterprise policies now properly support both DWORD and QWORD registry values.

Markdown, Editor Polish, and Python Environment Fixes​

Even if you never touch an AI agent, VS Code 1.142 brings quality-of-life improvements to daily coding and documentation workflows:

• Markdown Editor Overhaul: Editing Markdown now renders each source line on its own separate visual line so document structure is much easier to track (read-only mode still renders formatted Markdown normally). Better yet, the Markdown editor finally achieves parity with the standard text editor for several core actions: you can paste clipboard images directly into the Markdown editor, press F2 (Rename Symbol) to rename file links in place across the entire document, navigate document headings using the Outline view and breadcrumbs, and click a hover Copy button on code blocks without accidentally mangling the snippet.
• Merged Document Highlights: By default, when multiple extensions offer document occurrence highlights, VS Code only displays highlights from the highest-priority provider. Flipping on the new editor.occurrencesHighlightFromAllProviders setting merges highlights from all active providers.
• Python Environments Extension Cleanup: Python developers using modern tooling get a bundle of fixes. Python installations managed by uv now show up properly under Global, and if uv is required during setup but missing, VS Code offers to install it and proceeds straight to Python version selection without forcing a manual PATH refresh. Automatic environment selection now properly prioritizes a project's own virtual environment over nested ones, background refreshes stop overwriting your manual interpreter picks, and ${workspaceFolder} or relative paths in python.defaultInterpreterPath resolve reliably.
• GitHub Pull Requests (v0.170.0): The companion GitHub Pull Requests extension now enables stacked pull request support by default and allows updating conflict-free, locally checked-out pull requests directly.
• Under-the-Hood Memory Leak Fixes: Community contributor Simon Siefke landed nearly a dozen targeted memory-leak fixes across document paste edits, native context menu dismissal, output channels, debug adapters, terminal disposal subscriptions, and extension host language model providers—a welcome cleanup for developers who leave VS Code open for weeks at a time.