Chinese
Uncategorized

GitHub Copilot Cloud Sessions: Continuity Is Now an Operations Question

GitHub now connects a Copilot session across VS Code, the web, mobile, and CLI. Convenience brings environment, permission, snapshot, and billing questions.

Jacky Wang 4 min read 74 views

I often start a coding agent in a terminal on my Mac, let it run tests, and close the laptop while I do something else. The difficult question is not whether the agent can write code. It is where the work exists the next morning: a local process, a remote machine, a cloud sandbox, or a chat message that says “please continue.” GitHub’s recent Copilot changes put that operational question into the product itself.

GitHub’s October 28 release notes describe one Copilot session that can move between VS Code, the Copilot app, GitHub Mobile, github.com, and Copilot CLI. Cloud sessions run in isolated environments managed by GitHub. When a sandbox shuts down, it saves a snapshot so the session can be resumed later. The notes also say cloud sandboxes do not have a one-hour limit on total running time, and that usage billing begins in November 2026 from the shared Actions compute entitlement.

“Continue” can mean three different things

I would separate three modes. In a cloud session, closing the laptop does not stop the work because the environment is on GitHub’s infrastructure. With Agent Host, the work stays on my Mac or another machine; VS Code becomes a control surface, but the machine and Host must remain available. In an ordinary chat, there may be no durable workspace or tool state at all.

Their failure modes are different. Cloud work depends on repository permissions, network access, and snapshots. A local Host depends on sleep settings, VPNs, disk space, and process health. A chat may preserve the conversation but not the environment that produced the result. If these modes are treated as interchangeable, it becomes hard to tell whether a task failed in code or simply lost its runtime.

Start with the environment boundary

GitHub documents this command for registering the local environment from a project directory:

code agent host --github-environment

After that, I would check the branch, dependency reconstruction, allowed MCP servers, and approval behavior for every tool. GitHub says Copilot asks for approval before a tool call that requires it, and organizations can disable computer use through policy. Those defaults help, but I would still write the repository-level rules explicitly.

I would not give a background session unrestricted kubectl, production cloud credentials, and arbitrary egress. For a small SaaS project, I would begin with read-only tests and make deployment a separate, human-approved step. The agent can be aggressive about repairing a failing test without being allowed to turn a CI failure into a production change.

Shared compute makes accounting part of the workflow

Cloud sandbox billing starts in November and draws from a shared dollar-based Actions pool rather than a fixed number of runner minutes. That maps better to actual compute, but it also makes “start a long task and forget it” more expensive. A team needs to know which repository owns a session, who is responsible for it, and how much work is being resumed while it waits on an external system.

I would tag sessions with the repository, owner, and budget, then decide whether a snapshot should be retained when the task ends. A README edit should not leave a durable environment behind. A multi-hour migration may need a snapshot, because rebuilding dependencies and context from scratch wastes both time and compute.

GitHub is turning the coding agent from an editor panel into a continuously running work process. My guess is that the next useful differentiator will not be another model picker. It will be whether a session can explain where it runs, what it can access, what it cost, and who must approve its next action.

The primary references are GitHub’s Copilot in VS Code release notes and the connected cloud experience notes. I would still verify availability and the actual billing view in the organization before making this a team-wide default, because preview features and policy controls can change.

I also want a clear handoff record. When I move from phone to desktop, the session should show the last tool call, changed files, test result, and pending approvals. A pretty timeline is not enough if the next instruction can be issued against a different branch or environment. Continuity is useful only when the state is legible.

发表评论