GitHub’s Copilot updates for the week of October 5 bring local sandboxing to general availability, add Claude Haiku 5.5 across five paid plans, and make local Ollama models discoverable in Copilot CLI. The Copilot app also gains separate accounts for licensing and repository access, while VS Code 1.141 adds tools for viewing agent sessions together and reclaiming their storage.
The weekly changelog was published on October 9, 2026. The associated VS Code release is dated October 7. Taken together, the changes put emphasis on the practical work of running agents: choosing a model, controlling access, following concurrent sessions, and cleaning up after them.
Local sandboxing reaches general availability
According to GitHub, local sandboxing is now generally available in Copilot CLI, the Copilot app, and VS Code sessions using Agent Host. It lets users limit agents’ access to files, networks, and credentials. GitHub says the capability is included with Copilot at no extra cost.
The scope matters. For VS Code, the announcement specifically names sessions using Agent Host; it does not describe availability across every possible session configuration. The VS Code 1.141 release notes separately describe terminal sandboxing that limits file and network access on Windows, macOS, and Linux.
These statements address related aspects of agent access, but their wording differs. GitHub’s Copilot announcement includes credentials among the resources users can restrict. The VS Code release highlight names files and network resources. Neither statement alone supplies a detailed account of how each restriction is enforced across every environment.
For engineering managers, the useful distinction is between the availability of access controls and evidence of their effectiveness. The supplied announcements establish the former. They do not include independent security testing or a detailed comparison of sandbox behavior across operating systems. The practical implication is to evaluate the documented controls against the access a particular workflow needs, without treating general availability as a security guarantee.
More model choice in Copilot and the CLI
GitHub says Claude Haiku 5.5 is available to Copilot Pro, Pro+, Max, Business, and Enterprise users. This is an availability announcement: the evidence does not provide comparative coding benchmarks, latency measurements, or usage allowances for the model. It therefore supports an expanded model choice, rather than a conclusion about which model will perform best for a given task.
The CLI change concerns discovering local models. Users can invoke /model to find supported models from a running local Ollama instance alongside their configured models and the cloud models provided by GitHub Copilot.
Two qualifications are central to that description. Ollama must already be running locally, and GitHub refers to supported models. The announcement does not establish that every model available through Ollama is compatible, nor does it provide a compatibility list or setup procedure.
For developers working with local and cloud options, the value is a shared place to discover available models within the existing CLI workflow. Model discovery by itself does not establish the privacy properties of an entire session or demonstrate a performance advantage. Those questions remain outside the evidence supplied with this release.
Separate accounts for the Copilot license and repositories
The Copilot app now allows users to use one GitHub account for their Copilot license and another for repository access. GitHub gives the example of using an enterprise-provided Copilot license while accessing repositories through another account.
That directly addresses workflows in which the account carrying the license differs from the account used for code access. It also makes the product boundary clear: this change is announced for the Copilot app. The release does not establish equivalent account behavior across every Copilot interface or describe changes to repository permissions.
For a developer assessing the update, the relevant question is whether those two account roles need to be separate. For a team assessing deployment, the announcement leaves the detailed account setup and organizational policy requirements unspecified.
VS Code adds session visibility and worktree cleanup
VS Code 1.141 lets users arrange agent sessions side by side in a grid within the Agents window. GitHub describes this as a way to compare results or follow multiple tasks. Its immediate benefit is visibility: related conversations can remain in view together instead of requiring repeated switching between them.
The release notes also explain the architecture behind Copilot harness sessions. The harness is powered by the Copilot SDK and runs in a dedicated agent host process based on the Agent Host Protocol (AHP). According to the notes, this allows the same agent session to be accessed from multiple VS Code windows. Users select the Copilot harness from the harness picker in the chat input; the notes say it may already be the default selection.
Storage management gets a separate control. GitHub calls it “Open worktree cleanup” in chat, while the VS Code notes provide the command Chat: Open Worktree Cleanup. It shows how much disk space inactive agent session worktrees use and allows users to choose which to remove.
The release notes describe several safeguards and options:
- Users can filter sessions by inactivity and review each worktree’s size before selecting it for cleanup.
- Automatic cleanup can be configured for sessions with merged pull requests.
- Active, running, needs-input, and pinned sessions are protected from cleanup.
- Cleaning up marks selected sessions as done and deletes their worktrees. Restoring a session later recreates its worktree.
Those details make cleanup more consequential than simply hiding a conversation. The operation removes the selected worktree, with session restoration providing a path to recreate it. The evidence does not quantify expected storage savings, so the value depends on the inactive worktrees present in a user’s environment.
The strongest operational thread through this release is control over the agent workflow. Sandboxing addresses resource access; separate accounts address licensing and repository identity; model discovery addresses selection; and the VS Code changes address supervision and storage. For developers and engineering managers, these are concrete capabilities to assess against existing workflows. The announcements provide no measured basis for claiming faster development or better code quality.
