Skip to content
Ghost in the Droid

Differential A11y — What Changed On Screen

Android Supported iOS Supported

After every UI action, Ghost appends a short differential accessibility report to the tool result — the elements that appeared and disappeared since the previous action. Instead of re-reading the entire screen tree to figure out what its tap did, the agent gets the delta for free.

A11y diff (since last action):
+ 'Wi-Fi' (android.widget.TextView)
+ 'Bluetooth' (android.widget.TextView)
- 'Searching…' (android.widget.TextView)

Up to 6 added and 6 removed entries are listed; the rest collapse to + …N more new / - …N more gone. Element identity is bucketed to a 40 px grid so sub-pixel jitter doesn’t register as a change. The first action after launch has no prior state (empty diff), and an action that changes nothing is silent — only a real delta is appended.

Kill-switch

Section titled “Kill-switch”

Differential a11y is on by default. It’s fail-open: any error, or the kill-switch being off, yields an empty string, so the normal post-action screen-tree behavior is never disturbed.

Terminal window
# Disable it (kill-switch)
export A11Y_DIFF_ENABLED=false

Cost

Section titled “Cost”

One extra interactive-elements dump per UI action, cached per device. Read-only tools (screenshots, OCR) don’t update the baseline, keeping the diff strictly action-to-action. Works on the normalized element shape shared by Android and iOS.

Section titled “Related”