Playbooks · Coding tasks
Review Copilot's work in a nested Claude session
Written November 6, 2026
VS Code 1.145 lets one agent implement and another review in the same session. How to set it up, and what the reviewer can catch.
VS Code 1.145, released 4 November 2026, lets nested sessions use a different harness than their parent session. That is the setup for a second pair of eyes: Copilot implements a change in the main chat, and Claude reviews it in a nested chat inside the same session, without losing the session grouping. This page is a setup guide from the 1.145 release notes and the sessions documentation. No nested review was run for it, and no review results are claimed — the steps below are what the docs say, and the cost of a review is unmeasured here.
What 1.145 adds
A session groups the workspace, code changes, and conversations for a task. A nested session is a separate, interactive conversation grouped under an existing session, starting blank, without the other chats' history. Previously, every nested session inherited its parent session's agent setup — if you wanted a second opinion from another agent, you had to start a separate session and lose that grouping. In 1.145, nested sessions can use a different agent setup than their parent session. The release notes give the exact example this page is built on: ask Claude to review the plan that a Copilot chat wrote, or give Codex the same task, side by side in the same session. Each chat has its own model, agent, and permission levels, based on the agent running it, and changing them in one chat does not affect the others. When you move a chat, it keeps its history and continues with its selected agent setup. When you fork a nested session, it continues with the same configuration.
How to set it up
Enable chat.agentHost.multipleChatProviders.enabled — it is off by default. Then create a nested session from the Agents window: hover over the parent session and select the plus button for a new nested session, use its context menu, or press Cmd+T on macOS (Ctrl+T on Windows and Linux) while the session has focus. If mixed-agent chats are enabled, a picker appears above the input: choose the agent setup before the working folder, because the available workspace options depend on it. Pick the working folder next — keep the parent's folder to share it, or choose a new worktree to isolate changes — then enter the prompt.
For a review, the docs give a prompt shape worth copying: ask the agent to start a separate chat within the session to review the work, reusing the current workspace and not modifying files. The docs' own example is "Start a separate chat within this session to review the authentication flow. Reuse the current workspace and do not modify files." For a Copilot-implements, Claude-reviews run, the sequence is: let Copilot write the plan or the feature in the main chat, then open a nested chat, switch its harness to Claude, and ask it to review what Copilot produced. A chat that uses a different agent setup than its session names it when you hover over its chat tab or its row in the sessions list, so you can tell at a glance which chat is the reviewer.
What the reviewer can catch
A reviewer in a nested chat sees the same workspace, so it can read the implementation, the tests, and the surrounding code — but it starts blank, without the main chat's history. That is both the strength and the limit of the setup. It has not been steered by the conversation behind the implementation, so it reads the code as written rather than as intended, which is where a reviewer finds the gaps an implementer talks itself past: missing error handling, a test that does not cover the case it names, an assumption about an API that the code does not check. What it cannot do is judge intent it was never told — if the plan lives only in the main chat's history, the reviewer does not have it, so the review prompt needs to carry the intent: what the change was supposed to do, and what it must not break. Keep the review read-only unless you mean otherwise: the nested chat shares the parent's folder by default, and separate conversations do not isolate file changes, so a reviewer asked to "fix what you find" will edit the same files. Choose a new worktree for the nested session when the review must not touch the implementation.
Where permission modes differ
Each chat uses its own agent's model options and permission controls. Changing one chat's model or permissions does not change the other chats, and each chat retains its harness, model, and permission settings after restarting VS Code. That means the reviewer does not inherit the implementing agent's permissions: if Copilot ran with broad auto-approval and Claude's chat is set to ask before edits, the reviewer will ask where the implementer did not. Set the reviewer's permissions deliberately before sending the review prompt — read-only if the review is a check, approval-gated if it may apply fixes. The docs note that the agent setup must support multiple chats, and the parent session must not be archived; setups that support only one folder limit the nested chat to the parent's folder or to no workspace.
What it costs
The cost of a nested review is unmeasured here: no review was run, so no premium-request count is reported, and no estimate is made from vendor pricing. What the docs do say is that each chat carries its own model and its own session accounting — the context window control in the chat input shows how much context the active chat uses, including token counts and the session's total AI credits — so the review's cost is visible per chat rather than blended into the parent session. Expect the review to cost a session of its own: the reviewer reads the implementation and the surrounding code from scratch, since it starts without the main chat's history. Whether that cost is worth it depends on what the change is: a reviewer that re-reads the code is most useful where a mistake is expensive and the implementer had the least outside information — exactly the changes where a second harness earns its keep.
Limits
- No nested review was run for this page. The steps are from the release notes and the sessions documentation; nothing here reports what a reviewer actually caught.
- The premium-request cost is unmeasured. No estimate is made from vendor pricing, and the per-chat accounting the docs describe is the way to measure it on your own run.
- The reviewer starts blank. It does not see the main chat's history, so intent that lives only in the conversation behind the implementation is invisible to it unless the review prompt carries it.
- Nested chats share the parent's folder by default, and separate conversations do not isolate file changes. A reviewer can edit the implementation unless it is told not to or given a separate worktree.
- Mixed-agent nested sessions are a preview behind a setting that is off by default, and the available agent setups depend on setup requirements and organization policies.
Related playbooks
- Copilot code review goes Adaptive, checks secrets
- Decide if Copilot Code Review Max is worth it
- Copilot CLI 1.0.106: /code, /settings, GPT-6.1 Terra
- Kimi Code 2.5.3: steer a running session
Sources
- Visual Studio Code 1.145 — Release notes. mixed-harness nested sessions, the enabling setting, and the Claude-reviews-Copilot example
- Manage agent sessions in VS Code — Sessions documentation. nested sessions, harness picker, worktree isolation, and per-chat models and permissions
Back to: Coding playbooks hub · Also: Run your own agent bake-off on one real task · Run three independent lines of research at the same time in Claude Code, then combine the short summaries · Run Sonnet with an Opus advisor in Claude Code