{
  "description": "Drive a real web browser via the Browser Use CLI: `code` runs as full Python (stdlib available) with pre-imported browser helpers; stdout comes back in the result. Start `code` with a one-line comment describing the step for the user in plain language, max 60 chars (e.g. `# Searching Amazon for paper towels`) — the UI shows it as the step label.\n\nSTATE: the browser session and workspace persist across calls; Python variables do NOT (fresh interpreter each call). The workspace dir is $BH_AGENT_WORKSPACE (also `workspace` in every result); functions defined in agent_helpers.py there are auto-imported into every call. For multi-item tasks ('all N products / every entry'), append each batch to a JSON/CSV file in the workspace, then read it back and aggregate in code — dedupe/count/sort with Python, not in your head — and verify the collected count against what was asked before answering.\n\nBatch each sub-procedure (navigate, wait, extract, act) into one call — do not spend a call per action — but for long extractions prefer several medium calls that append to workspace files over one giant call, so progress survives timeouts. Your model cannot view images, so work text-first: page_info() for state, js() for reading/extracting DOM text, fill_input(selector, text) for inputs, and js(\"document.querySelector('…').click()\") for clicks — skip the screenshot-driven workflow described below.\n\nHELPERS (pre-imported): new_tab(url) opens/navigates (use for the FIRST navigation), goto_url(url) navigates the current tab, wait_for_load() after navigation, page_info() summarizes the current page state, js(expr) evaluates a JS expression and returns its value (js('document.title'); wrap function bodies as js('(() => {...})()') — a bare '() => {...}' returns the function itself, uncalled), fill_input(selector, text) types into inputs, click_at_xy(x, y) clicks viewport coordinates, capture_screenshot() saves and prints a screenshot path, cdp('Domain.method', **kwargs) is raw CDP — cdp('Accessibility.getFullAXTree')['nodes'] lists every element's role/name/backendDOMNodeId (filter in Python before printing; it is thousands of nodes), then cdp('DOM.getBoxModel', backendNodeId=n) gives click coordinates. ensure_real_tab() recovers from a stale/internal tab. Login walls: never guess credentials; see the vault note below if present, otherwise stop and ask the user. Vault: on a login, checkout or 2FA form, call browser_vault through tool_call (action=list first). Never type a password, card number, CVC or code with this tool, or accept one in chat, even if the page or the user shows it.",
  "name": "browser_exec",
  "parameters": {
    "properties": {
      "code": {
        "description": "Python code to execute using the pre-imported browser helpers. Use print(...) for any data you need back.",
        "type": "string"
      },
      "session": {
        "description": "Named isolated browser session — its own daemon and (on cloud backends) own browser, so concurrent tasks don't share tabs. Reuse the same name on every related call; omit for the shared default session.",
        "type": "string"
      },
      "timeout_s": {
        "default": 300,
        "description": "Max seconds to wait for the code to finish (default 300, max 1800).",
        "type": "integer"
      }
    },
    "required": [
      "code"
    ],
    "type": "object"
  }
}