VS Code 1.143 Makes the Copilot Harness the Default for New Sessions and Adds Model Comparison and Worktree Cleanup

Overview

Microsoft released Visual Studio Code 1.143 on October 22, 2026, according to the 1.143.0 release tag on GitHub, which shows the tag published at 13:00 UTC. The release notes say the update “makes the Copilot harness the default for new sessions and adds tools to compare models, reclaim storage, run models locally, and navigate file changes.” A follow-up build, 1.143.1, is listed on the same notes page with the line “The update addresses these issues.”

What We Know

Copilot harness becomes the default

Per the release notes, “The Copilot harness is now the default for new sessions in VS Code. It’s our latest harness, built on the Copilot SDK, a shared foundation that helps align behavior, capabilities, and improvements across Copilot products.” The notes add that Copilot harness sessions can be kept running in the background and reached from both the editor window and the Agents window, and that these sessions run through the Agent Host.

The release also adds an experimental file-level working set. According to the release notes, “Select Keep on a file, or in the heading for all files, to mark changes as reviewed and remove them from the working set. The edits stay on disk, and files return if the agent changes them again.” The feature is enabled through the chat.editor.workingSet.enabled setting, and the notes say file-level Undo is planned for later releases, “landing in Insiders first, then Stable.”

Comparing models

A preview feature in the Agents window runs one task against several models. The release notes say users can “Select two or more models to attempt your task independently” and “mix Copilot, Claude, and Codex harnesses and configure different reasoning levels.” Each attempt runs in its own session. A user can optionally pick a Judge model to evaluate the attempts; its scorecard “shows time and token usage, with expandable rationale, validation evidence, and links to each attempt.” Per the notes, “Synthesis starts only when you request it.” The feature is gated behind the sessions.chat.compareAgents.enabled setting.

Worktree cleanup and large-chat performance

An experimental command, Chat: Open Worktree Cleanup, shows how much disk space inactive agent session worktrees use. According to the release notes, cleaning up “marks the selected sessions as done and deletes their worktrees,” a session can be restored later to recreate its worktree, and “Active, running, needs-input, and pinned sessions are protected from cleanup.”

On performance, the notes say opening a long Agent Host chat in the Agents window “used to freeze the window for up to half a minute.” Microsoft’s table, which it describes as the median of three runs on macOS, lists a cold open of a chat with about 1,150 requests and about 530 MB of history falling from 35.6 seconds in 1.142 to 9.0 seconds in 1.143, with the longest unresponsive period dropping from 28.3 seconds to 0.6 seconds. A second chat with about 1,450 requests and about 200 MB of history went from 20.3 seconds to 3.0 seconds to open.

On-device model

The release notes describe an experimental option to run MAI Code 1.1 Flash on the user’s own machine. They state that “On-device models require GitHub sign-in and an NVIDIA GPU with 96 GB or more of GPU memory,” and that responses run on the device “don’t use AI credits.” Setup requires a llama.cpp build that supports the model, with chat.copilot.localModel.llamaServerPath pointing at its llama-server executable.

Editor, source control and enterprise changes

The notes raise the editor’s diagnostic highlight limit “from 500 to 5,000 diagnostics per file” and add an experimental file list beside the multi-file diff editor, with status letters A, M, D and R for added, modified, deleted and renamed files. For organizations, the release notes add a managed telemetry.headersProvider setting for generating OpenTelemetry export headers, and say that when enterprise policy requires the Copilot harness, “the Local harness now stays visible but is disabled in the harness picker.” They add that these controls “are temporary while Local is retired.” The notes list no deprecated features or settings for this release.

What We Don’t Know

The release notes do not say how many users will see the new default, or whether the experimental features (worktree cleanup, the working set, the multi-diff file list and on-device models) will move to stable behavior in a named release. The performance table covers two chats on macOS, and Microsoft does not report results for other platforms or chat sizes. The notes also do not state a date or contents for the 1.143.1 fix beyond the single line quoted above.

Context

The release continues a monthly series of agent-focused VS Code updates. The Herald’s coverage of VS Code 1.140 described multi-folder agent sessions and worktree handling; the 1.143 notes build on that by adding tooling to review, discard and clean up the worktrees agents create. The notes also say canvases, introduced as a preview in 1.142, can now be opened from the chat menu without asking Copilot to open them.