Sites that watch for automated browsing

Choose the route

• A public page that needs no sign-in and no clicks: read it with the web search or fetch tools.
• Anything else on a social network, a marketplace or a site where the user is signed in: only the computer tools, in the Computer's own browser: launch_app({agent_browser: true, urls: [url]}). The Computer picks the installed browser (Chrome, Edge, Firefox, or a window of the user's Safari); pass bundle_id only when the user asks for one. The browser tools drive a browser in a way these sites detect, so the Computer keeps them off every site the task opened in its own browser; other sites in the same task may still use them.
• If launch_app has no agent_browser, this needs a Mac Computer: tell the user and stop.
• If a call says a computer tool does not exist, take the tool's exact name from your tool list and call it again before telling the user the tools are unavailable.
• This holds when the user says they are signed in: their sign-in for such a site is kept in the Computer's own browser. Never connect the user's everyday browser for it.
• If the site asks to sign in, ask the user to sign in once in that browser window and stop; the sign-in stays for later tasks.

Read

• Open pages by URL where the site has one, such as a search, profile or post URL: fewer clicks.
• Read with get_window_state; pass query to narrow a long page. If a read says it was cut, read again narrower instead of guessing.
• Text inside pictures: recognize_text; in a picture viewer, pages with next_key.
• A state the tree does not show, such as liked or followed: look with include_screenshot: true.
• To repeat a flow over items on one page, use run_script: scroll, expand, search inside the page and read. A script there sends at most 20 inputs and opens no tab. Many sites open a result in a new tab, so a script never clicks a link or a tab: click those as their own input. Return what the script extracted, look at the window, then continue.
• Stop scrolling when two reads add no new items. Read only what the task needs: one search per keyword, and never reopen an item already read.
• Act only in the tabs this task opened. A click that opens a new tab is the task's tab. Close each item's tab with its tab's Close button after reading it, and the task's last tab when done. If a tab the task did not open is in front, stop and tell the user.

Write

Write only when the user asked for that action in this task: publish, comment, like, follow or delete. Text to review or a request to draft is not permission.

1. Check on the page that the signed-in account is the one the user meant. If not, stop and tell the user.
2. Make sure exactly one item matches the target. If several do, ask.
3. Send once, then read the page again and confirm the result shows, such as the comment under that account.
4. If the result is unclear, read the page. Never send again blind.

If the user takes over the window, read the page and the account again before the next input.

Checks and limits

• When a page shows a captcha, a code or QR check, a sign-in wall or a too-many-requests message, call report_blocker with that kind and the page's URL at once. Then tell the user in the Room what the page asks for and that they can deal with it in that browser window, report what was done, and end the task.
• When only one item cannot be viewed, call report_blocker with content_unavailable and its URL, and skip it.
• Never pass a check yourself, switch accounts or retry around a limit.
• When the Computer refuses input with COMPUTER_INPUT_BUDGET, COMPUTER_SITE_COOLDOWN, COMPUTER_SITE_BLOCKED or COMPUTER_TAB_CHANGED, report what was done and end the task.