VS Code 1.130 dropped July 22 with three changes that are bigger than their release notes billing. Microsoft decoupled AI agent sessions from editor windows, shipped a new AI-gated permission model that cuts approval interruptions, and upgraded its entire repository to TypeScript 7. These are not quality-of-life tweaks. Together they signal that VS Code is treating AI agents as persistent infrastructure — not plugins you tolerate between real work.
Agent Host Protocol: Sessions No Longer Die When You Close a Window #
The headline change in 1.130 is the Agent Host Protocol (AHP). Agent sessions now run in their own dedicated process, separate from the editor. Close a window and the session keeps running. Open another window and connect to the same live session. Start a task in your main project view, switch to a split workspace to review a diff, and the agent never loses context.
The architecture is straightforward: a host process runs locally (via message port IPC) or remotely (over WebSocket). VS Code provides adapters for Copilot, Claude, and Codex. The protocol uses JSON-RPC with immutable state and pure reducers, so dropped connections trigger clean reconnection — either replaying missed actions or restoring a fresh state snapshot. You can run a standalone host with code agent host
and expose it remotely with the --tunnel
flag.
The deeper signal here is the AHP spec itself. It is an open, agent-agnostic protocol — not a proprietary API. Third-party agents can implement it. Microsoft is positioning VS Code as an orchestration runtime, not just an editor with AI bolted on.
Enable it in settings: "chat.agentHost.enabled": true
. The setting is managed at the organization level if you are on Copilot Business or Enterprise.
Assisted Tool Approvals: Letting the Model Judge Risk #
If you have run any agentic workflow longer than about ten steps, you know the approval interrupt loop. The agent wants to run a tool. You click approve. It wants to run another. You click approve again. Fifty tool calls and you have spent twenty minutes clicking through prompts that all looked low-risk anyway.
VS Code 1.130 introduces assisted permissions, where the language model evaluates the risk of each tool call before surfacing it for approval. Low-risk operations run automatically. Genuinely risky ones — file deletions, shell commands with broad scope, network requests to new hosts — still require a click. The net effect is far fewer interruptions without eliminating oversight on actions that warrant it.
Enable it: "chat.assistedPermissions.enabled": true
. This feature requires the agent host to be active, so it depends on chat.agentHost.enabled
.
The obvious question is whether you should trust the model to assess its own tool-call risk. The answer is: more than a blanket allowlist. An LLM evaluating tool calls has access to full session context — it knows whether the agent is in a sandboxed test environment or editing production config. A static allowlist knows none of that. You should still review what gets auto-approved in the agent logs, but the model reasoning about risk in context is a meaningful improvement over binary on/off.
TypeScript 7 Is Now in Production at Microsoft #
TypeScript 7.0 shipped on July 8 — a Go-based compiler rewrite delivering 8x to 12x build speedups over TS6. If you have been waiting for a signal that TS7 is production-ready before migrating your repo, VS Code 1.130 is that signal. Microsoft upgraded both its own codebase and the bundled TypeScript extension to the TS7 release version. The biggest TypeScript repository on earth just moved.
The migration path is clean: if your project compiles on TypeScript 6.0, it compiles identically on 7.0 with no code changes. The speedup shows up immediately in build times, code completion, diagnostics, and find-all-references. One caveat: the stable programmatic API for tools built on top of TypeScript arrives in 7.1, so if you are maintaining a TypeScript language tool rather than using TypeScript as a user, wait for the next minor release.
Worktree Support for Claude and Codex #
Before 1.130, Git worktree isolation in the Agents window was Copilot-only. It now extends to Claude and Codex sessions running on the agent host. You can run parallel agents on separate feature branches within the same workspace without context collisions — Agent A on feature/auth-refactor
, Agent B on feature/payments-v2
, each isolated in its own worktree.
The Agents window also gets compact diffs (less gutter whitespace, better header alignment), live file-level insertion and deletion counts, and a cleaner single-line layout for quick chats so they do not visually compete with full project sessions.
Smaller Additions Worth Noting #
Chat timestamps and elapsed time are now accessible on hover, controlled by "chat.verbose": true
. Copilot Business and Enterprise users can see aggregate AI credit usage in the Copilot status menu without configuring user-level budgets. Terminal diff output now handles Git mnemonic prefixes (i/
, w/
, 1/
, 2/
) as clickable file links.
The Settings to Enable Now #
{
"chat.agentHost.enabled": true,
"chat.assistedPermissions.enabled": true,
"chat.verbose": true
}
The full release notes are on the VS Code site. The agent host architecture docs are worth reading before enabling it in a shared environment.
VS Code 1.130 is not a maintenance release. The agent host decouples sessions from windows, assisted permissions shift risk evaluation to where context actually lives, and TypeScript 7 bundled in the editor is a clear production endorsement. The trajectory is obvious: VS Code is becoming the runtime layer for agentic development, and it is building the infrastructure to support that seriously.