Agent approvals

jork

Bring the next or previous pending request into focus.

aorEnter

Approve the request in focus. Enter approves while focus is on the card, not on a button or field.

d

Deny the request in focus.

EnterorEscape

In an argument field, Enter saves and Escape restores the value (Command or Control with Enter in a multiline field). Focus returns to the value.

Tab

Moves through the Rules button, the editable values, Deny, Always allow and Approve, the queue rows and the Resolved toggle.

Arrow keys, Escape

Inside the Always allow menu and the Rules popover, move between items and close, with focus back on the trigger.

• The request in focus is a group named by its title, inside a region that takes focus after each decision, so the shortcuts keep working when the buttons that were pressed leave.
• Shortcuts answer anywhere in the card, and after a click in it left focus on the page, but stay quiet in fields, the menu and the popover, so typing an edited value never approves anything.
• Risk is a shield icon in its status color plus the word Low, Medium or High (with risk added for screen readers). Outcomes in the card and the log are an icon plus a word: Approved, Denied, Allowed by rule, Not sent.
• Every decision, edit, new rule, removed rule and failure is announced in a polite status region, and the demo says that nothing reached a real agent.
• Editable values are buttons named with their label, value and Edit; edited rows add a hidden "edited" to the label and a named Restore button.
• Queue rows are buttons with the agent, the summary and the risk in their name. The Resolved toggle carries aria-expanded and aria-controls.
• Pointer focus draws no ring; keyboard focus shows Arc's shared focus ring.

• Bringing a queued request into focus is one crossfade of the request (the old one fades out in 100ms, the new one fades in over 200ms from 3px) while the card's height follows on a critically damped spring and the row closes its gap. Nothing bounces.
• After a decision the next request slides up: it rises 14px into place on the same spring as its row collapses below.
• Always allow sends a small rule chip from the menu up into the Rules button on an ease-in-out path, then the button's count rolls and a faint accent wash fades out of it. Covered rows swap their risk for Allowed by rule with one crossfade and collapse a moment later.
• Sending shows a calm text shimmer that stops when the agent answers. The pending, rules and resolved counts roll with tabular numerals.
• Reduced motion drops travel, the chip flight and the shimmer: requests, rows and labels change in place with every state, focus move and announcement kept.

• One card up to 640px wide that fills narrower containers; it holds without horizontal scroll from 320px up.
• Below a 540px container the shortcut hints and the agent names in the queue hide, since touch is the likely input there.
• Below 440px the model name and the word risk drop out of the meta line, argument labels sit above their values, the Enter key cap leaves the Approve button and queue rows keep the marks, the summary and a short status.
• Long titles wrap, while row summaries and log entries truncate with the full text still in their accessible names or titles.

• Only the request in focus renders its arguments and actions; queued requests are one row each.
• Timers for simulated delivery, settling and rule matches are tracked and cleared on unmount, and async callbacks read the latest state from a ref instead of re-subscribing.
• The relative times share one 15 second clock through useSyncExternalStore, so the card re-renders a few times a minute at most.
• Height changes are measured with one ResizeObserver per animated slot and animate only that slot.

• Write title as the call itself in a few words and summary as a short row label. Backticks in title render code in the mono face.
• Use real agent and model names and set provider so the right mark shows (anthropic, openai or google); any other value shows a neutral bot mark.
• Give each permission a stable tool id (vercel.deploy, gmail.send). Rules match on it, so several requests for the same permission are covered by one rule.
• Keep the reason to one line in the agent's voice and put anything a person might change into args. Mark values that must not change with editable: false.
• onDecision receives the edited argument values. Apply those, not the originals, and keep edited in your audit trail.
• Always allow both approves the request in focus and creates the rule; requests the rule then covers arrive through onDecision with outcome "allowed" and the rule id.
• Return a rejected promise from onDecision to show the error state; the Error's message is shown to the person.

The full library index for assistants is at /llms.txt.