Pimp My IDE / garage dispatch
Back to garage
September 30, 2026 | terminal history / command design

Give power tools labeled controls.

The early Bloomberg Terminal put domain commands on a custom keyboard. Agent IDEs keep adding capability. Their next job is to make frequent actions, dangerous actions, and stop paths visible without asking users to remember a chat spell.

Use chat for intent. Use named controls for repeated operations, authority changes, and recovery.

The keyboard was part of the product model.

IEEE Spectrum's history of the Bloomberg Terminal describes the 1982 Market Master as a monochrome display, custom keyboard, controller, and private network. Its color-coded function keys named market tasks. Users could request data and analysis without memorizing every command.[1]

The design grew with the job. The article records a trackball in 1990, audio hardware in 1992, multilingual keyboards by 1996, and fingerprint authentication in the early 2000s. Those additions were not decorative. Each one exposed a new mode of work or a new trust boundary.[1]

The useful lesson is not "build a weird keyboard." It is "make the operating model visible."

Agent IDEs now have the same density problem.

Visual Studio Code 1.140 adds a Copilot harness, experimental multi-folder sessions, remote delegation, a research preview for multi-model orchestration, shared worktree folders, and enterprise controls. The release notes also say the harness runs in a dedicated agent-host process and can be reached from several windows.[2]

That is a lot of machinery behind one chat input. A prompt can ask for any of it, but free-form language does not show the current folder, authority, process, or escape route. The interface still needs stable names for repeated operations. It needs separate controls for changing reach and stopping work.

Commands need discovery and dispatch proof.

VS Code's keyboard documentation already has the right pieces. Its shortcut editor lists commands with and without keybindings. Users can filter the list, inspect conflicts, and turn on dispatch logging to see which shortcut and command the editor received.[3]

Accessibility help also changes by context. It covers the editor, terminal, notebook, Chat view, and Inline Chat. Keyboard-only navigation and screen-reader support are first-class paths rather than an afterthought.[4]

An agent command deck should combine those ideas. Show a short set of named jobs. Keep the native keyboard path. Display the target and authority before execution. Record the dispatched command after execution. Put Stop where it cannot disappear behind a model response.

Do not turn every command into a permanent key.

A dense interface becomes useful when controls match repeated work. Rare commands still belong in search. High-frequency commands deserve stable placement. Dangerous commands need an explicit confirmation and a visible scope. Recovery commands need a route that does not depend on the system being healthy.

  1. Count real command use before pinning anything.
  2. Keep labels literal and stable across menus, shortcuts, logs, and receipts.
  3. Show the target workspace and authority beside the action.
  4. Log the resolved command, not only the user's words.
  5. Test conflicts, keyboard access, screen-reader output, and Stop under load.
Interactive makeover / command deck chassis

Cut four honest keys

Traditional purpose replaced: one overloaded chat box plus a buried command palette. Better version: select the job on a labeled key, close the target and authority interlocks, then copy a command-map test card. Native controls preserve keyboard and screen-reader operation.

Choose the pinned job

The labels are examples. Replace them with commands observed in your own workflow.

Command keys
Command-map sections
Command map draftKey 1 / 4
Visible operating model

Command deck chassis

FindDiscover
ActChange
CheckProve
StopCancel

Pin search only if it repeats.

Show which corpus will be searched and what kind of result will come back. A general search key should not silently edit files.

Selecting every section means the command-map structure is ready. It does not prove command frequency, shortcut safety, dispatch accuracy, accessibility, or cancellation.

Four interface jobs

One chat box cannot show all four.

01 / DISCOVER

Show what can happen

Search and command palettes expose available operations. They do not grant extra authority.

02 / DISPATCH

Show what did happen

Resolve free-form intent into a named command, target, and result that another person can inspect.

03 / CONFLICT

Show who owns the input

When shortcuts overlap, log the key event, active context, winning rule, and invoked command.

04 / ESCAPE

Keep Stop independent

A cancellation path should not require cooperation from the task, model turn, or interface that needs stopping.

Sources read

Source log and evidence boundary
  1. IEEE Spectrum, "A Brief History of the Bloomberg Terminal", published and read September 30, 2026. Allison Marsh traces the Market Master launch, dedicated network, color-coded keyboard, trackball, audio, multilingual hardware, authentication, and later terminal growth.
  2. Visual Studio Code 1.140 release notes, released and read September 30, 2026. Microsoft documents the Copilot harness, agent-host process, multi-folder sessions, remote delegation, HydraFusion research preview, shared worktree folders, and enterprise controls. Preview and experimental labels are preserved here.
  3. Visual Studio Code keyboard shortcut documentation, read September 30, 2026. The page documents command discovery, shortcut editing, conflict inspection, context rules, and keyboard dispatch troubleshooting.
  4. Visual Studio Code accessibility documentation, read September 30, 2026. The page documents keyboard navigation, screen-reader support, context-specific accessibility help, contrast controls, and terminal access.
  5. Hacker News discussion, exact item 49909583, checked September 30, 2026. It led to the IEEE article. Reader comments are not used as evidence for the historical or product claims.

Evidence boundary: this page extracts interface lessons from documented products. It did not use a Bloomberg Terminal, test VS Code 1.140, inspect a keyboard, or benchmark command use. The interactive deck creates a review template. It does not configure an editor or prove a command route.