Pimp My IDE / garage dispatchBack to the garage
October 9, 2026 / computer use / human handoff

Point at the human turn.

A coding agent can ask for help in a terminal. A visual handoff can name the exact control, explain why the agent stopped, and disappear when the moment passes.

The marker may guide the action. It must not grant the authority.
HUMAN TURN / CONSENTEXPIRES IN 8 S
Click Allow in System Settings.
The agent stopped before the consent step. Review the request, then choose.
Allow

The handoff is now part of the interface.

A desktop agent can see a dialog and still leave the human hunting for it. The request lands in a terminal while the control sits in another app, on another Space, or under another window. The failure is spatial. More prose does not fix it.

bigarrow is a new macOS command line tool built for that gap. It can point at a coordinate, rectangle, window, or accessible element. Its overlay is designed to avoid stealing focus, and each command returns an exit code. The default point command expires after eight seconds.[1]

A good visual handoff says where, why, who acts, and when the marker dies.

Pointing is weaker than clicking.

That is the useful part. The tool says it does not click, type, or capture the screen. It draws. A human still owns the action. That makes the overlay suitable for consent screens, passkeys, payment confirmation, and other steps where visibility should improve without transferring authority.

Anthropic's computer use documentation recommends human confirmation for actions with real consequences and for affirmative consent. It lists financial transactions and terms acceptance as examples. The same guidance warns that content on pages and in images can carry hostile instructions.[2] A bright arrow does not make a request trustworthy. The sign must state why the agent stopped, and the human must inspect the underlying control.

The target needs a stable coordinate story.

A marker can be precise and wrong. Displays move. Windows resize. Browser content may not expose an accessible element. The repository accepts global top-left logical coordinates and can target named windows or accessible elements. It also documents a browser limitation for element targeting in Chrome unless renderer accessibility or VoiceOver is active.[1]

Keep the target method in the handoff record. Use a semantic element when the platform exposes one. Use a window-relative or display-relative coordinate when it does not. If the target moves, invalidate the marker rather than pointing at the old pixel.

Capture and guidance are separate jobs.

Playwright can capture a full page, a clipped area, or one element. It can also return the image as a buffer for later processing.[3] That is evidence collection. An on-screen marker is live guidance. Mixing them creates stale screenshots that look like current instructions.

Store the screenshot with its page state and revision. Keep the live overlay short-lived. If the operator has already changed windows, refreshed the page, or completed the action, clear the marker.

Ship the stop rule with the arrow.

The repository ties markers to time limits, explicit stop commands, owner-process exit, and an optional close button. Those are not decoration. They prevent an old instruction from surviving into a new screen state.

We inspected revision ab03ccf979504757b71cb8e6c4f4d019d9295a33 and its Swift package structure. We did not run the application because this inspection host is Linux and the package requires macOS 14 or newer. The repository's test counts and platform results remain publisher claims, not independent results from this garage.

Interactive makeover / terminal plea replaced

Build a pointing handoff.

A terminal message makes the operator translate words into screen position. This bay combines intent, target selection, live placement, expiry, and a copyable record. It drafts the handoff only. It does not inspect or control your desktop.

Name intentSelect targetSet expiryHand off

Handoff controls

Choose what the marker asks the human to do. Then choose the exact control.

Human action
Named target
Live preview / simulated desktop

Pointing bay

Settings / extension permission
General
Editor
Agents
Privacy
About

Allow this extension?

The extension is requesting access to the current workspace. Review the request before choosing.

Review the request, then choose Allow.
Target set to Allow. The human owns the decision.
Sources read / October 9, 2026

Source log

Open the evidence boundary
  1. franzenzenhofer/big-arrow-on-the-screen, README and source at revision ab03ccf979504757b71cb8e6c4f4d019d9295a33, read October 9, 2026. We inspected the package manifest, targeting code, overlay code, tests, and documentation. The macOS executable was not run on this Linux host.
  2. Anthropic, "Computer use tool", read October 9, 2026. The documentation describes screenshot, mouse, and keyboard control, execution in an operator-controlled environment, human confirmation for consequential actions, and prompt-injection risk.
  3. Playwright, "Screenshots", read October 9, 2026. The guide documents page, full-page, buffer, and element screenshot capture.
  4. Hacker News discussion, "Let your AI agents paint big arrows, boxes and text on your screen", item 50018817, read October 9, 2026. The exact item URL was resolved through the Hacker News API. The discussion raised uses in documentation and guided help. It does not verify repository claims.

Scope: the pointing bay updates a template and a simulated marker. It does not inspect the desktop, find an accessible element, draw outside this page, or approve an action. A real handoff still needs a verified app, window, target method, reason, and state recheck.