Browser Ops Skill

Browser Operations Guide

Tool Inventory

Loading: All are deferred tools — call ToolSearch(names: "OpenUrl,BrowserSnapshot,BrowserAct,BrowserWait,BrowserDownloadImage") to activate their schemas before invoking.

Core Mental Model

Browser operations = observe-act-verify loop. After every action you must snapshot to confirm state, because webpages are stateful — a click may trigger navigation, a popup, or an AJAX load, and you cannot predict the outcome. Blindly chaining actions is the single most common failure mode.

OpenUrl → BrowserSnapshot → analyze → BrowserAct → BrowserWait → BrowserSnapshot → ...

OpenUrl Opening Modes

• headless — First choice for pure automation. No UI; suited for scraping, background agents, and batch operations.
• tab (default) — Use when the user needs to see/operate the page. Embedded panel.
• window — Use when the user needs a standalone window for deep interaction.

Rule of thumb: user doesn't need to see it → headless; user needs to see it → tab; user wants a standalone window → window.

BrowserSnapshot Details

A single BrowserSnapshot call returns:

• Page info: URL, title, readyState
• Full text: body.innerText (truncated at 32KB)
• Interactive element list: up to 120 clickable/inputtable elements with their selectors
• iframe contents: text and elements from same-origin iframes
• Screenshot: captures the full page by default (fullPage), saved to the session asset directory
• rawHtmlPath: disk path to the complete outerHTML

What if text exceeds 32KB? Use Read/Grep on rawHtmlPath to fetch the full DOM — don't re-snapshot repeatedly.

Screenshot control: fullPage: false captures only the current viewport.

Three Core Workflows

Workflow 1: Information Extraction

OpenUrl → BrowserWait { type: "load" } → BrowserSnapshot

Extract the required information directly from the returned text.

Pagination scraping: BrowserSnapshot → BrowserAct { action: "click-text", text: "Next" } → BrowserWait { type: "networkidle" } → loop. Snapshot on every page.

Workflow 2: Form Filling and Login

1. OpenUrl → BrowserSnapshot to inspect form structure and selectors
2. For each field: BrowserAct { action: "fill", selector: "...", text: "..." }
3. Submit: BrowserAct { action: "click-css", selector: "button[type=submit]" }
4. BrowserWait { type: "urlIncludes", url: "/success" } to confirm success
5. BrowserSnapshot for final confirmation

fill vs type: fill atomically clears and inputs — best for forms; type appends characters at the current focus — best for search boxes.

Login caveats: Cookies persist for the session, so after a single login subsequent requests carry auth automatically. Do not snapshot password fields.

Workflow 3: Screenshots and Image Downloads

• BrowserSnapshot — snapshot + full-page screenshot (fullPage by default)
• BrowserSnapshot { fullPage: false } — viewport only
• BrowserDownloadImage { selector: ".product-image" } — download an image from the page

Screenshots are a debugging superpower: when text doesn't reveal the issue, the screenshot often does at a glance.

Selector Selection Strategy

Pick selectors from the element list returned by BrowserSnapshot — don't guess on your own.

Priority: #id > [data-testid] > input[name] > .class > click-text

Wait Strategies

• load: Use after traditional page navigation.
• networkidle: First choice for SPA/AJAX pages — safest but slowest.
• urlIncludes: Wait for redirect after form submission.
• textIncludes: Wait for asynchronously loaded content.
• timeout: Last resort.

Error Diagnosis

Element not found? → BrowserSnapshot to confirm page state → may have navigated elsewhere → may be below the viewport (scroll and retry) → may be inside an iframe (cross-iframe not supported)

Page load timeout? → BrowserSnapshot to see how far loading got; the content may already be sufficient.

Action has no effect? → Page not fully loaded (add BrowserWait) → selector matched the wrong element (check via snapshot) → a popup is blocking (close it first)

SPA content empty? → Snapshot again after BrowserWait { type: "networkidle" } → if still empty, wait on specific content with textIncludes

CAPTCHAs and Anti-Bot

When you encounter a CAPTCHA, 403/429, or an anti-bot page, stop immediately and notify the user — do not retry blindly.

Iron Rules

1. BrowserSnapshot after every action to verify state
2. Snapshot before acting — don't guess selectors
3. Wait for the page to be ready before acting
4. When text isn't enough, look at the screenshot
5. Don't expose sensitive info (don't snapshot after filling a password)
6. Stop immediately on anti-bot encounters