Browser Harness

The Facebook Pages skill is designed to mine public Facebook Pages for recent posts, external URLs, and Page metadata. It is intended for read-only tasks such as harvesting content or links and is NOT for write actions like commenting, reacting, or messaging.

Requirements:

• A real Chrome browser driven by Browser Harness.
• A logged-in session is highly recommended. While Pages are public, logged-out sessions encounter aggressive 'see more' gating and login interstitials that break the scrolling loop after approximately 5 posts.

Use browser-harness for any task requiring direct browser control via CDP, such as automation, scraping, testing, or interacting with web applications.

When to use:

• Tasks requiring interaction (click, type, navigate).
• Accessing a user's logged-in session.
• Handling JS rendering or bot-protected pages.
• When a direct HTTP fetch fails or returns a shell page.

When NOT to use:

• For basic fetching of public information (use curl or a standard fetch tool instead).

Note on Domain Skills: Domain skills are disabled by default. To enable them, set the environment variable BH_DOMAIN_SKILLS=1. When enabled, if a task is site-specific, you must read the files in $BH_AGENT_WORKSPACE/domain-skills/<site>/ before attempting a new approach.

browser-harness <<'PY'
print(page_info())
PY

The DuckDuckGo Instant Answer API (https://api.duckduckgo.com) is a public, no-auth API that returns Wikipedia-sourced abstracts, infoboxes, and instant answers. It is designed for entity lookups and utility queries rather than general web search.

Key Usage Guidelines

• Do not use a browser: All requests are single http_get JSON calls.
• Recommended Parameters: Always include skip_disambig=1 (to upgrade ambiguous terms to primary articles) and no_html=1 (to strip HTML tags from results).
• Query Types: Results are categorized by a Type field:
  • A (Article): Specific entity with a full abstract.
  • D (Disambiguation): List of related topics for ambiguous terms.
  • E (Exclusive): Instant answers (calculators, converters, etc.).
  • "" (Empty): No match found.

Substack exposes a public REST API at {publication}.substack.com/api/v1/ that allows for data extraction without authentication, API keys, or a browser. This works for both native .substack.com subdomains and custom domains.

Capabilities:

• List all posts from a publication (/api/v1/posts)
• Fetch full post content by slug (/api/v1/posts/{slug})
• Fetch post comments using the integer post_id (/api/v1/post/{post_id}/comments)
• Read the RSS feed for lightweight metadata (/feed)

Limitations:

• Paywalled Content: For posts where audience is only_paid, the body_html and body_text fields return only a truncated HTML preview rather than the full article.
• Search: There is no cross-publication search API available without a logged-in session.
• Comments: The comments endpoint requires the integer post_id, not the post slug.

When interacting with elements inside Shopify iframes, you have two options:

1. Coordinate clicks (click(x, y)): These pass through iframes at the compositor level and work for simple interactions.
2. JS clicks (via js() with target_id): These are generally more reliable for routine button taps because:
   • Element text content is stable across UI redesigns.
   • Device Pixel Ratio (DPR) scaling on retina displays is handled automatically.
   • React event handlers are guaranteed to fire (whereas CDP mouse events can sometimes hit transparent layers above the button).

Data within Quora's SSR (Server-Side Rendered) payloads is double-encoded. Each .push() argument is a JavaScript string literal containing JSON. To correctly parse this data into a Python dictionary, you must perform two steps:

1. Decode the JS string escaping: Use json.loads('"' + raw + '"') to convert escaped characters (like \" to ").
2. Parse the inner JSON: Pass the resulting string into a second json.loads() call to parse the actual object.

Skipping either step will result in parse errors.

When automating browser tabs, follow these practical rules:

• Visibility: switch_tab() is often insufficient if the user expects the browser window to visibly change focus. Use cdp("Target.activateTarget", targetId=tid) to force the tab to show.
• Filtering: When calling list_tabs(), use include_chrome=False to avoid internal pages. Ignore chrome://omnibox-popup.top-chrome/ as it can appear as a fake page target.
• Validation: If a page reports dimensions of w=0 h=0, you are likely attached to the wrong target or a non-window surface.
• Dynamic UIs: For interfaces with dropdowns or modals, re-read element rectangles (rects) after opening them before attempting coordinate-based clicking.

The Infobox field is returned as a dictionary containing structured data about an entity. If no infobox is present, it returns an empty string "". Always check the type before processing.

Structure

• content: A list of dictionaries, each containing data_type, label, and value.
• meta: A list of metadata items.

Flattening the Infobox

To convert the structured content list into a flat dictionary for easier access, use the following pattern:

if isinstance(data['Infobox'], dict):
    fields = {item['label']: item['value'] for item in data['Infobox']['content']}
    # Example: fields['Founded'] == 'December 08, 2015'

if isinstance(data['Infobox'], dict):
    fields = {item['label']: item['value'] for item in data['Infobox']['content']}
    # fields['Founded'] == 'December 08, 2015'
    # fields['Products'] == 'ChatGPT, GPT-5...'

BOSS直聘 uses a hybrid messaging architecture for its chat functionality:

• Conversation list: Loaded via WebSocket (ws6.zhipin.com) upon page load (not via REST).
• Message history: Retrieved via the REST API /wapi/zpchat/geek/historyMsg.
• Real-time messages: Delivered via WebSocket push from ws6.zhipin.com.

IMPORTANT: This documentation covers read/retrieval mechanics only. Never send messages without explicit user permission.