Copilot in VS Code now reports token usage per model and installs portable agent plugins

GitHub published its monthly roundup for Copilot in Visual Studio Code on August 31, 2026. It covers releases 1.132 through 1.135, shipped over the course of the month. The Agents window changed most: chats can be grouped, one session reaches more than one window, and plugins built to the Agent Plugins 1.0 standard run in other compatible clients rather than only in VS Code.

The Agents window

Related chats can be arranged in horizontal or vertical groups, and the layout comes back when you return to the session. Typing /btw opens a side conversation that shares the primary chat's context and prompt cache while the primary chat keeps running. A prompt timeline control in the transcript gutter jumps to a specific prompt and shows the file changes that came with it.

A session is no longer tied to one window or one application. The Agent Host connects several VS Code windows to the same agent session, and the Sessions list picks up recent Copilot or Claude sessions started in other applications. Inside a Claude session you can switch between models from an Anthropic subscription and models from a Copilot subscription at any point.

• Agent plugins written to the Agent Plugins 1.0 standard install in VS Code and in other compatible agent clients.
• An experimental setting opens the Agents window without a GitHub sign-in when Claude is configured with an API key.
• The experimental /rubber-duck command asks a second model to surface missed details and edge cases.
• Session details and editors open in a single side pane, with diffs that adapt to the width available.

Chat and review

Long conversations get two fixes. Text search now runs across the whole chat transcript, with case matching, whole words, and regular expressions. Chat sticky scroll pins the current prompt in place, so a response never drifts away from the question that produced it.

The response footer reports token usage. Hovering over it shows input, cached input, and output tokens for each model used in that turn, which puts the cost of a single exchange where the exchange happened. Rendered Markdown diffs can be reviewed in an experimental hybrid editor, and the breadcrumb bar switches between a regular editor and a diff editor.

Integrated browser and dictation

The integrated browser stops being a preview pane. You can select and annotate several HTML elements on a page, and the agent works through that feedback in one batch. Local HTML files reload on their own when they change on disk, and the browser can be set as the default editor for HTML files through the editor associations setting.

Dictation runs an on-device model that handles several languages and can detect which one you are speaking. Cleanup instructions can be set at user or workspace level, so the transcript follows the terminology of the project. In the terminal, shell-aware cleanup keeps command syntax rather than writing out spoken punctuation.