Ghosthands Jev Fast Lane
An optional browser-control fast lane uses Jev to choose operations and UI targets.
Added to Jevfast
How it uses Jev
In the optional experimental Jev lane, Ghosthands sends an indexed table of current screen controls as state and asks Jev to choose an operation, with additional target Choice questions when matching controls exist. The planner applies the selected action through its computer-control loop. It can use a small text model to write literal text only when the playbook has no matching candidate.
What Jev decides
Screen operation selection (illustrative request shape)Example answers · not a recorded Jev response · Source ↗
Question 1 · operation
YOUR APP
INSTRUCTION
Choose the next operation toward the goal using the goal, playbook, safety rules, and current-step context.
STATE
Current indexed screen state from the settled screen and action history. The operation question receives the goal, scrubbed playbook, safety rules, and optional current-step context. Criteria may be pruned when operations are excluded or no suitable control exists.
JEV · CHOICE
- CLICK
- TYPE_TEXT
- SELECT
- KEY_ENTER
- SCROLL_DOWN
- SCROLL_UP
- WAIT
- VERIFY_STOP
- DONE
App workflow
- Read and settle the screen
Ghosthands waits for a stable screen and creates Jev state from the indexed controls and action history.
- Build and validate questions
Build the operation Choice and any supported target Choice heads from the current controls, then validate that each answer is a permitted choice with probabilities.
- Act and guard the loop
Map the selected operation and target to a real mouse or keyboard plan. Additional guards handle rejected DONE decisions, repeated scrolls, dead clicks, and uncertain targets.
Why it is interesting
The fast lane turns a live screen into a compact control table, so Jev chooses among named operations and concrete element indices without a screenshot-grounding call. The repository explicitly describes the lane as experimental and keeps its vision lane as the default.