Visual Studio Code 1.143: Copilot Agent Default, Compare Models and 96GB Local AI GPU Requirement

The story was submitted as "1.143 (Insiders)", and that label needs a quick explanation. The 1.143 notes were first published while the build was still in the Insiders channel. Daily.dev covered it as an Insiders build that adds a stream of Agent Host and Copilot chat improvements. Microsoft's release notes archive now lists Visual Studio Code 1.144 (Insiders) above a plain "Visual Studio Code 1.143" entry. That means 1.143 has moved into the numbered release line, and Insiders users have already moved on to 1.144. The official page is marked last updated October 19, 2026, and lists 1.143.0 downloads for Windows, macOS and Linux. If you already have VS Code installed, Microsoft says to use Check for Updates.

Copilot harness is now the default​

The headline change is that the Copilot harness is now the default for new sessions in VS Code. Microsoft describes it as being built on the Copilot SDK, a shared foundation that helps align behavior, capabilities, and improvements across Copilot products.

What this means in practice:

• Background sessions: you can keep Copilot harness sessions running in the background and access them from both the VS Code editor window and the Agents window. These sessions run through the Agent Host.
• Latency fixes: Microsoft says user reports helped us reduce delays caused by MCP servers that don't finish loading their tools.
• Scope: The change applies to new sessions. Organization policy still decides what managed users can pick, as covered below.

NT Compatible frames the release as the point where Microsoft finally stopped treating the Copilot harness as a suggestion and made it the rule. That's a fair reading of where things are heading. The release notes themselves say that the old Local harness is being retired.

Section summary: Copilot is now the default runtime for new agent sessions, and the rest of 1.143 builds tooling around it.

Reviewing and undoing agent edits​

People who found it awkward to approve or reject agent edits in the editor get partial relief:

• File-level working set (Experimental): Turn on chat.editor.workingSet.enabled. Selecting Keep on a file, or in the heading to keep all files, marks the changes as reviewed and removes them from the working set. Keep does not revert anything. The edits stay on disk, and a file comes back to the list if the agent edits it again. Microsoft says file-level Undo and per-change Keep/Undo are planned for later releases, landing in Insiders first. Neither is available yet.
• More ways to discard in the Agents window: Before this release, Discard Changes only worked in the Uncommitted Changes changeset. It now also works in three more changesets, and each one discards something different:

Check which tab you're on before you click. Branch Changes is the heavy option. It can delete files the branch created, which matters if you mixed hand-written work into an agent branch.

Compare Models: several attempts at one prompt (Preview)​

Microsoft's highlights describe it like this: Run the same task with multiple models in the Agents window, then compare and combine their solutions. It only works in the Agents window.

1. Enable sessions.chat.compareAgents.enabled.
2. In the model picker, select Compare Models.
3. Pick two or more models. You can mix Copilot, Claude and Codex harnesses and set different reasoning levels.
4. Enter one prompt. Every attempt starts at the same time, each in its own session, and you can open each attempt's worktree or changes at any point.
5. Optionally, pick a Judge model. Its scorecard shows time and token usage, a rationale you can expand, validation evidence, and links to each attempt.
6. Select Synthesize to build on the recommended attempt using ideas from the others. Synthesis never starts on its own.

Running several frontier models on one prompt will use tokens quickly. The Judge's token figures are worth checking before this becomes a habit.

Cleaning up worktrees (Experimental)​

Agents often work in Git worktrees, and over time those can take up a lot of disk space. Run Chat: Open Worktree Cleanup from the Command Palette to:

• see the size of each inactive worktree
• filter sessions by how long they've been inactive
• select worktrees to remove

Cleaning up marks the selected sessions as done and deletes their worktrees. You can restore a session later, which recreates its worktree. Active, running, needs-input and pinned sessions are protected from cleanup. The cleanup editor can also remove sessions with merged pull requests automatically. If you set chat.agentSessions.sessionStorageCleanupSuggestion.enabled, the Agents window will tell you when cleanup is worth doing.

On-device models: MAI Code 1.1 Flash (Experimental)​

This is the most hardware-dependent feature in the release. The model picker gets an On This Device tab, where you can download MAI Code 1.1 Flash and use it in Copilot harness sessions. VS Code downloads and manages the model. Responses generated locally are labeled On-device and don't use AI credits. If you turn on Use in Auto on the model's details page, Auto can send suitable requests to the local model and fall back to the cloud.

Setup:

1. Install a llama.cpp build that supports the model.
2. Point chat.copilot.localModel.llamaServerPath at its llama-server executable.
3. Enable chat.copilot.localModel.enabled.

The requirement that rules most people out: GitHub sign-in plus an NVIDIA GPU with 96 GB or more of GPU memory. Microsoft gives the Surface Laptop Ultra with NVIDIA RTX Spark and up to 128 GB of unified memory as an example machine. A typical gaming laptop or workstation card won't qualify. Warp2Search notes that some of these features require specific hardware.

Sandbox certificate trust on Windows​

This part matters most to Windows administrators. Agent sandboxing uses a credential proxy. Sandboxed tools get placeholder credentials, and the proxy swaps in the real ones when it forwards requests to approved destinations. For HTTPS, that means tools have to trust the proxy's TLS certificate. Copilot ships a certificate bundle for tools that can use one. Tools that rely on the operating system's trust store can throw certificate errors instead.

• When trust is missing, VS Code shows a dismissible notice: Sandbox trusted CA is not installed, with an Install button.
• On Windows, installing adds the locally generated CA's public certificate to the machine-wide trust store. On macOS it goes into the user's login Keychain. Either way, it can require OS authorization.
• Sandbox: Manage Trusted CA can install, rotate or remove trust for the agent host of the last-focused chat, local or remote. Each action asks for confirmation and names the target host.
• Remove takes away OS trust but keeps the stored certificate and private key.
• Installing is optional. Domain filtering doesn't need the CA, because the local proxy can tunnel HTTPS without decrypting it. Missing trust doesn't stop chat from working.

Adding a locally generated root CA to the Windows machine-wide store is a real security decision. Security teams should review it the same way they review any interception certificate. Also note that removing trust leaves the private key on disk.

A separate change: if managed policy allows sandbox bypass, users can turn off Sandboxing for terminal for a single session from the permissions menu. Administrators who block bypass will see that toggle locked on.

Enterprise controls​

• telemetry.headersProvider: a managed setting that runs a helper program to generate and refresh HTTP headers for OpenTelemetry exports, such as short-lived collector tokens. It applies only to Copilot Agent Host sessions and must be set in enterprise-managed settings, not user settings.
• Managed model defaults: New Copilot harness chats use the organization's default model even when a user's chat.defaultModel says otherwise. Users can still switch models within a chat.
• Local harness retirement: When policy requires Copilot, Local still appears in the picker but is disabled, with a Learn why link. Copilot is required to enforce non-empty managed permissions.deny rules or sandbox.enabled: true. Setting vscode.temporaryAllowLocalHarness: false blocks Local, Claude and Codex. Setting vscode.temporaryDefaultToLocalHarness: true keeps Local available. Microsoft warns that Local cannot enforce runtime-only permission and sandbox controls.
• Managed plugins install automatically. Precedence runs from managed settings, to trusted repository settings, to user choices. Team assignments only apply when nothing else has been set.

NT Compatible counts these among the release's rules that can force Copilot while disabling Local, Claude, and Codex.

Faster large chats and other improvements​

• Large chats: Chat history now loads in small slices. In Microsoft's macOS tests on one roughly 530 MB chat, a cold open fell from 35.6 s to 9.0 s, and the longest freeze dropped from 28.3 s to 0.6 s. These are Microsoft's own medians, not a guarantee for every setup.
• Diagnostics: The editor now highlights up to 5,000 diagnostics per file, up from 500. Each extension's diagnostic collection is still capped at 1,000.
• Multi-diff file list (Experimental): multiDiffEditor.experimental.fileList adds a file list marked A/M/D/R, shown as a list or tree, on the left or right.
• Detached background shells: Copilot harness sessions now stream output from detached shells and let you stop them with Stop Shell.
• Canvases: You can open them from New Chat (+) > Canvas once canvas extensions are installed.
• Copilot CLI: The Windows installer now offers to install GitHub Copilot CLI. Organizations can turn that option off through Managed AI Settings.

Who should update, and what to check​

Individual developers will probably like the default change and the worktree cleanup. Expect to test more than you click, since many of the new features are Experimental or Preview.

Administrators should check three things before rolling out 1.143:

• whether to install the sandbox CA into the Windows machine-wide trust store
• whether your policies will force users off the Local harness
• what the new managed defaults do to user model preferences

In short, this release makes Copilot the default and leaves most of the remaining decisions about sandboxing, harnesses and models to administrators.