gnome-jarvis/vendor/agent-harness/agent-workspace/skills/browser/SKILL.md

name, description

name	description
browser	Drive the headed Jarvis Chromium window with the browser tool. Use for cookie walls, forms, logins, leftover bot checks, and any extra clicks after search or fetch.

Browser

Search and fetch already run in this same Playwright Chromium window. Call browser when the page needs a click, a form, a cookie banner, or a leftover challenge. Do not use cu_observe, cu_click, curl, or wget for websites.

Choose a tool

1. web_search / google_search / wiki_search / hn_search / code_search — find public links.
2. fetch_page / web_fetch — read one public http(s) page (JavaScript runs).
3. browser — drive that window. Cookies persist across these tools.

browser actions

Call with action plus the fields for that action. Snapshot refs are strings like "1".

action	Fields	Result
navigate	url (public http or https)	Opens the page and returns a snapshot
snapshot	none	Current url, title, numbered refs, short aria text
click	ref from the last snapshot (or selector / text)	Clicks, then a fresh snapshot
type	ref + text; optional submit true	Fills the field; submit presses Enter
press	key (Enter, Tab, Escape, Control+l, …)	Key, then a snapshot
scroll	dy pixels (optional dx)	Scrolls, then a snapshot
wait	ms (optional text to wait for)	Waits, then a snapshot

Private, loopback, and metadata hosts are blocked before Chromium starts.

Loop

1. navigate or snapshot so you have fresh refs.
2. Pick the ref whose name matches the control (Accept, Next, email, search box).
3. click or type with that ref.
4. Read the new snapshot. Refs from earlier snapshots are stale.
5. Repeat until the page is usable, then fetch_page on the current url if you need the article text.

Cookie walls and challenges

• Cookie banner: snapshot, then click the Accept / Agree / I understand ref.
• challenge: true or a Cloudflare / “just a moment” page: tell the user to finish the prompt in the visible Jarvis browser window. Do not guess. Then snapshot or fetch_page again.
• Login that needs a password: stop and ask the user. Never type credentials unless they just provided them for this site.

Do not

• Speak refs, selectors, or tool JSON.
• Call browser for a page you can already read with fetch_page.
• Use computer-use tools on the Chromium window.
• Keep searching the same query instead of opening a hit.