Quick answer: VS Code 1.135 can show recent Copilot, Claude Code, and Codex sessions created outside the editor, continue supported sessions inside VS Code, and run a Copilot session through the new Agent Host. To try it, update to VS Code 1.135, open the Agents window, choose Copilot as the Session Target, and use /rubber-duck when you want an experimental second-model review. External-session visibility is controlled by chat.agentSessions.showExternal.[1][4]
- In VS Code, run Help → Check for Updates and confirm version 1.135 or newer.
- Open the Agents window from the title bar.
- Create a session and select Copilot under Session Target.
- Use a Git worktree for parallel or risky edits; use the active folder only when the task needs your current uncommitted files.
- After the first implementation, enter
/rubber-duckfor a second opinion, then review every proposed change and test result yourself.
What changed in VS Code 1.135?
Released on August 26, 2026, VS Code 1.135 focuses on portable agent sessions, an experimental Rubber Duck review, a simpler Agents window, and per-model token details for chat turns. Microsoft says the release is rolling out gradually, so Check for Updates is the fastest way to request it.[1]
GitHub’s August Copilot roundup also groups VS Code versions 1.132 through 1.135 into one larger agent-workflow update. That roundup covers side-by-side chats, /btw side conversations, external sessions, Agent Host, integrated-browser feedback, transcript search, and improved on-device dictation.[2]
1. Update and confirm the release
- Open VS Code.
- Choose Help → Check for Updates on Windows or Linux. On macOS, use Code → Check for Updates.
- Restart if prompted.
- Open Help → About and confirm that the version is 1.135 or later.
If an option below is missing, first confirm your version, sign-in state, Copilot entitlement, organization policy, and whether the feature is marked experimental. Microsoft notes that the Agent Host is enabled in current Stable and Insiders builds, but specific harnesses and capabilities still have their own prerequisites.[3][5]
2. Start a session on the Agent Host
- Open a project folder.
- Select Open in Agents in the VS Code title bar.
- Select New.
- Choose the workspace or repository.
- Open Session Target and select Copilot.
- Choose the model, permission level, and isolation option shown for your account.
- Send a focused task with a clear acceptance check, for example: “Add input validation, run the existing tests, and show the diff without committing.”
The Copilot harness runs locally on the Agent Host and uses the Copilot SDK. Because the host owns the session rather than one editor window, the same session can remain available from another VS Code window or the Agents window.[5]
Test session persistence safely
- Start a harmless task, such as reading a repository and proposing a test plan.
- Close the project folder but keep VS Code itself running.
- Reopen the folder.
- Return to the session in the Agents window and confirm that its state is present.
For local sessions, closing a folder is not the same as quitting VS Code: Microsoft states that VS Code must remain running because it manages the local Agent Host.[3]
3. Show Copilot, Claude Code, or Codex sessions from other apps
VS Code’s session manager can discover supported local sessions created by Copilot CLI, the GitHub Copilot app, Claude Code, and Codex. Open the Sessions list filter, select External, and choose None, Recent, Last 24 Hours, Last 7 Days, or All.[4]
You can configure the same behavior in settings:
// settings.json
{
"chat.agentSessions.showExternal": "recent"
}
Important: use the Settings UI if the value above is rejected, because accepted setting values can change during a rollout. The release notes state that two recent external sessions are shown by default in the 1.135 experience, while the current session-management documentation describes external sessions as hidden by default. That difference suggests rollout or documentation-state variation, so the live filter in your installed build is authoritative.[1][4]
What “continue” means
Opening an external session lets you inspect its conversation. When you send a message, the Agent Host adopts the session, after which it is no longer treated as external by the filter. GitHub’s release announcement specifically says continuation uses your Copilot subscription, so seeing a transcript does not guarantee that every original provider’s exact runtime, tools, model, or billing route will continue unchanged.[1][4]
4. Use Rubber Duck for a second-model review
Rubber Duck is experimental. In a Copilot Agent Host session, enter:
/rubber-duck
Microsoft describes it as a complementary-model second opinion intended to surface missed details or edge cases.[1]
For a more useful review, ask the primary agent to complete one bounded stage first. Then run Rubber Duck with a concrete review target:
- Check this migration for backward-compatibility breaks.
- Review authentication and authorization boundaries.
- Look for tests that pass for the wrong reason.
- Identify unhandled null, timeout, retry, and partial-failure cases.
- Compare the implementation against the acceptance criteria.
A second model is still AI output, not independent verification. Keep the final gate deterministic: inspect the diff, run the project’s tests and linters, review dependency and permission changes, and avoid granting broad access merely to remove approval prompts.
5. Pick the right harness and isolation mode
VS Code distinguishes the harness from the language model. The current documentation lists Local, Copilot, Claude, Codex, and Cloud harnesses. They differ in runtime, available tools, code access, provider-specific behavior, and authentication.[5]
| Choice | Use it when | Main caution |
|---|---|---|
| Local | You need current editor context, extension tools, diagnostics, or configured local/BYOK models. | It edits the active workspace. |
| Copilot Agent Host | You want persistent local sessions and Copilot-specific capabilities. | Check subscription, policies, tools, and approval mode. |
| Claude or Codex | You need that provider’s harness behavior or want to continue a compatible external workflow. | Authentication, billing route, permissions, and available tools can differ. |
| Cloud | The task is well scoped and should run remotely against a GitHub repository and pull request. | It cannot use all local editor context or tools. |
For code isolation, choose a new Git worktree when parallel tasks should not touch your active workspace. Choose the current folder when the agent needs uncommitted changes. Microsoft warns that worktrees start from committed Git state and do not automatically include uncommitted, untracked, or ignored files. Worktree isolation also does not itself block commands, network access, or files outside the worktree.[5]
6. Check token usage instead of guessing
Hover over the redesigned response footer to see input, cached-input, and output tokens used by each model in a chat turn.[1]
Use that readout to spot overly broad prompts, repeated context, or an expensive review loop. If a session has become noisy, start a new session for a new topic or use /compact with instructions about what must be retained. VS Code’s session documentation says automatic compaction occurs when the context window fills, and manual compaction can preserve selected decisions while reducing later context.[4]
Suggested five-minute verification checklist
- VS Code reports version 1.135 or later.
- The Agents window opens and lists the intended workspace.
- Copilot appears in Session Target and authentication succeeds.
- A test session survives closing and reopening its folder while VS Code stays running.
- The External filter shows only the history window you intend.
/rubber-duckappears only in a compatible Copilot Agent Host session.- Generated changes are reviewed in a diff and validated with real tests.
- No Full Access or bypass mode is enabled without a deliberate need.
Troubleshooting
External sessions do not appear
Confirm that the originating app is supported, the session exists locally, and the External filter is not set to None. For Copilot discovery, the current documentation says repository-associated sessions updated within the last seven days are included. Also check chat.agentSessions.showExternal in Settings.[4]
Rubber Duck is unavailable
Verify that this is a Copilot session running on the Agent Host, not merely a Local, Claude, Codex, or Cloud session. The command is experimental, so availability can vary by build, rollout, account, or policy.[1]
Codex is missing from Session Target
VS Code’s current harness guide says Codex is not listed by default. It requires either the OpenAI Codex extension for Chat view usage or the experimental Agent Host setting chat.agentHost.codexAgent.enabled; using Agent Host in Chat view additionally requires chat.editor.codex.preferAgentHost.[5]
The agent continued but lost a tool
Session continuity does not mean every harness has identical tools. The Local harness can use VS Code and extension-provided tools, while Copilot, Claude, Codex, and Cloud have their own capabilities and restrictions. Re-check Session Target, enabled tools, MCP availability, authentication, and permissions after a handoff.[5]
FAQ
Does VS Code 1.135 keep an agent running after I quit VS Code?
Not for a local Agent Host. Microsoft says closing a folder can leave the turn running, but VS Code must remain running because it manages the local host. Remote Agent Host behavior is different.[3]
Can I continue a Claude Code or Codex session in VS Code?
VS Code documents discovery for Claude Code and Codex sessions. Supported external sessions can be opened from the Sessions list; when you message one, Agent Host adopts it. Exact provider behavior, tools, and billing may differ from the original app.[4]
Is Rubber Duck a replacement for code review?
No. It is an experimental AI second opinion. Treat it as another review pass, then verify the diff, tests, security boundaries, and acceptance criteria yourself.
Can two chats edit the same project?
Yes. Multiple chats in one Agent Host session share the session workspace or worktree, and their edits are combined in session changes. Use separate worktree-isolated sessions when tasks must not modify the same files.[4]