browser

name	browser
description	Drive a real Chrome window — read pages as an accessibility tree, click, type, upload, wait, and run JavaScript. Use for anything on a site that needs a logged-in session, a rendered page, or interaction — reading a page that needs JavaScript, filling a form, checking how something looks, or working through a site the way a person would.

browser

A real Chrome window, driven from the shell. It stays open between commands and keeps its logins, so you can watch what is happening and take over by hand.

cd ~/.pi/agent/skills/browser/scripts
node start.mjs

The first run installs Playwright and says so — a few seconds, once per machine, nothing to do. It also makes a profile of its own: log into whatever you need, once, by hand in that window, and it is kept from then on.

The loop

Take a snapshot, read it, act on the numbers it printed.

node nav.mjs github.com/BurntSushi/ripgrep
node snapshot.mjs          # writes a file, prints the path

The snapshot is the page's accessibility tree — every frame, nothing summarised. Read it or grep it. Interactive elements are numbered:

  [20] link "Issues"
  [39] combobox "Search Issues"

Those numbers are what the other commands take. They belong to the snapshot that printed them; after anything changes the page, take another.

Acting

node click.mjs 20                      # --double, --right, --force
node fill.mjs 39 "memory leak"         # --enter, --append, --type
node press.mjs Escape                  # --handle <n>, --times <n>
node select.mjs 12 "Option 1"          # a <select>, by visible label
node hover.mjs 8                       
node upload.mjs 5 ./report.pdf