v4.1.0

[4.1.0] - 2026-10-20

Added

• The server sends MCP instructions telling the model which tool to
  use when: a pre-built scraper first for known sites, then
  scrape_as_markdown / scrape_batch, and the browser tools only for
  clicks, typing or screenshots.

Changed

• Every tool sets readOnlyHint, destructiveHint and openWorldHint
  explicitly, following OpenAI's plugin rules. run_scraper and the AI
  insights tools are no longer read-only (they start a collection job),
  browser navigation and scrolling are no longer destructive, and
  scraping_browser_click_ref / scraping_browser_type_ref are
  destructive. Clients may ask for confirmation before those tools.
• Tool errors are token-redacted, stripped of stack traces and capped
  at 500 characters, including browser tool errors and upstream error
  pages.
• Tool descriptions no longer claim to unlock any page "even if it uses
  bot detection or CAPTCHA".
• scraping_browser_screenshot describes the full_page option.

Fixed

• sophie_assistant.js file was missing from the npm package.
• Passwords in URLs (user:pass@), such as the browser zone password,
  are redacted from every tool error and from the server's error logs,
  not only when the browser fails to connect.
• search_dataset accepts is_null and is_not_null filters without
  a value. They were listed as operators but rejected by the schema.
• The startup zone check times out (BASE_TIMEOUT, default 10s)
  instead of hanging the server when the API does not answer.