Pimp My IDE / Garage Dispatch
← Back to the garage
September 2, 2026 · Agent workspace architecture

More roots. One steering rack.

Coding agents are crossing repository boundaries inside one editor window. Good. But a wider workspace is not permission soup. Every root needs a name, every hook needs a home, every command needs a working directory, and every handoff needs a receipt.

The take

Multi-root support turns “the workspace” from a place into a routing decision. If the cockpit cannot show which root supplies context, hooks, execution, and review, it has added lanes without painting the road.

The workspace just became a small network.

Visual Studio Code 1.136 adds experimental multi-root support for Copilot and Claude agent sessions in the editor window. Sessions can work across all folders, preserve project URIs and working directories, and resolve projects by name. When names collide, the release notes say the agent reports the choices instead of silently picking one.

That last detail is the real upgrade. Ambiguity should become interface, not inference. A command aimed at api must not drift into the wrong folder because two projects share a cute name. The editor should surface the collision, and the operator should choose the lane.

One wide session still has narrow parts.

The same release draws a useful boundary: multi-root support is currently scoped to the editor window, while agent hooks remain scoped to a single workspace folder. If hooks appear in multiple folders, VS Code asks for one primary folder to load them from.

That is not an embarrassing limitation. It is a reminder that “can read several roots” and “can execute policy from several roots” are different capabilities. Hooks are executable wiring. Loading every root’s wiring because every root is visible would turn convenience into ambient authority.

Visibility may be plural. Ignition should stay explicit.

The competition is routing the same problem.

Zed’s current stable notes approach the cockpit from another angle: the Agent Sidebar can reorder projects in a multi-workspace, while its terminal tooling has been tightened around targeting project subdirectories. Different product, same pressure. As agents operate across larger workspace graphs, folder order, target selection, and command location stop being sidebar trivia.

The old file tree answered “where is the file?” The agent cockpit must answer four harder questions: Which roots are visible? Which root supplies executable hooks? Where will this command run? What record survives the handoff?

Garage rules for multi-root agents.

  1. Name every root uniquely. Treat duplicate project labels like duplicate service names: an error to resolve, not a hint to guess.
  2. Pin one hook root. Show it in the session header and require a deliberate change when policy moves.
  3. Put the working directory beside the action. Not buried in logs after the command has already fired.
  4. Keep review local to the change. Cross-root context can help reasoning; it should not erase ownership boundaries or separate checks.
  5. Print a dispatch receipt. Roots seen, hook source, commands and directories, changed files, checks, and unresolved ambiguity.
Interactive makeover / workspace axle rack

Lock the route.

Traditional purpose replaced: a folder list plus one vague “agent active” badge. Better version: one native, keyboardable rack mechanically links visible roots, executable hooks, command location, and a copyable dispatch receipt.

Workspace Axle Rack

Tab through the four locks. The driveline only reaches full engagement when visibility and authority are labeled separately.

Workspace routing locks

Dispatch Readout

Route unverified · 0/4 locks

The session can see a broad workspace, but nothing proves which root owns policy or execution. Keep the ignition off.

ROOTSweb → /shop/web · api → /shop/apiunlabeled
HOOKSprimary policy folderfloating
ACTIONcommand → working directoryhidden
RECEIPTfiles · checks · unresolved choicesmissing
Why it is better: the component does not pretend four checkboxes create security. It turns one overloaded workspace label into four reviewable decisions and produces the handoff artifact the folder tree cannot.

Sources read, not vibes

  1. Visual Studio Code 1.136 release notes (September 2, 2026): experimental multi-root agent sessions, project-name resolution, working-directory preservation, session hierarchy, and the single-folder boundary for agent hooks.
  2. VS Code documentation — Multi-root Workspaces: workspace files, folder naming, settings scope, tasks, launch configurations, and search behavior across roots.
  3. Zed stable release notes (1.17.2): multi-workspace project reordering in the Agent Sidebar and project-subdirectory targeting improvements in agent terminal tooling.