# Browser > 用异步 Python 调用 QwenPaw 内置 Browser SDK 驱动真实浏览器。完整参考在下方；上下文压缩后请重新加载此 browser skill。 - Skill: `agentscope-ai/browser-2` (Agent Skill) - Install (CLI): `npx skillmds@latest add agentscope-ai/browser-2` - Raw SKILL.md: https://api.skillmd.com/api/skills/agentscope-ai/browser-2/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: agentscope-ai (https://skillmd.com/u/agentscope-ai) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/agentscope-ai/browser-2 --- # 浏览器 保持工作纪律：先感知当前页面，再通过已列出的 API 动作，最后重新感知后才可声明成功。 只能陈述本轮实际观察到的事实。卡住时的合格交付 = 说明卡在哪一步 + 已亲眼验证的部分结果； 不要为了给出完整答案而补全你没有看到的内容。 尊重人工边界：登录、验证码、2FA 或任何必须由人完成的步骤，调用 `await browser.handoff(...)` 后停止，绝不自动化这些流程。 浏览器完不成时，不要拿其他渠道（如 web_search）的数据顶替并仍说成浏览器结果—— 如实写明每个数据的来源。 这是 QwenPaw 内置的 Browser SDK，不是 Playwright。它是封闭的 API 表面： 未列出的方法不存在。完整参考在下方；若上下文中不再保留，请用 Skill 工具重新加载 此 browser skill。 QwenPaw Browser SDK — complete reference. This is QwenPaw's OWN internal SDK and this is the ENTIRE API; these are all the entrypoints. The SDK is already in scope as Browser — call the methods below directly. Write async Python. Work in a loop: perceive → act → verify. # Copy this shape: browser = await Browser.connect() # connect once; reused all session page = await browser.open("https://example.com") # open a page obs = await page.snapshot() # PERCEIVE — page text is obs.text if len(obs.text) < 6000: print(obs.text) else: # Large page: read selectively instead of dumping everything. lines = [line for line in obs.text.splitlines() if "keyword" in line] print(f"{len(obs.text)} chars total; {len(lines)} matching lines:") print("\n".join(lines[:80])) # For a focused count, use: await page.snapshot(query="keyword") await page.get_by_role("textbox", name="Search").fill("laptop") # ACT await page.get_by_role("button", name="Search").click() # ACT obs = await page.snapshot() # VERIFY — re-perceive to confirm print("Verified; inspect obs.text with the selective pattern above.") Session state: this is a stateful session — variables you assign (browser, page) persist across calls, so connect once and reuse them. If a call reports the session was reset, re-run await Browser.connect(). Chrome backend caveat: with backend=chrome you operate inside the user's real browser. A session is a tab-ownership group — tabs are isolated per session, but identity (cookies, logins, storage) is shared with the user's profile and with every other session. Do not rely on session-level identity isolation on this backend. browser (orchestration): await Browser.connect(*, identity: "auto"|"user"|"avatar"|"guest" = "auto") -> browser Connect as an identity: user, avatar, guest, or auto. ``auto`` picks ``user`` when Chrome is connected, otherwise ``guest``. An unavailable explicit identity raises instead of substituting. await browser.open(url: str | None = None) -> page Open a page at ``url`` and return it. Reuses this session's active page when one exists; otherwise a new page is created. Pages are released when the response cycle ends; start each cycle by calling ``open(url)`` again. await browser.pages() -> list of page ref (.id, .url, .title, .active) List open pages with URL, title, and active-state details. await browser.switch_page(page: page ref (.id, .url, .title, .active)) -> none Make the given page ref active for later operations. await browser.close_page(page: page ref (.id, .url, .title, .active)) -> none Close the given page ref in this session. await browser.session_status() -> session status (.owner, .variant, .context, .connected) Report the owner, variant, context, and connected state. await browser.handoff(reason: str, instructions: str = "") -> a result dict Hand a step back to a human (captcha, login, 2FA). Pass a short reason and instructions; the run stops on this signal — never automate these flows. The active cycle-scoped page is retained for one extra response cycle after the handoff. await browser.present(url: str | None = None) -> page Open a page retained for the chat lifetime. await browser.close() -> none Close this session's browser and release its context. page (operation): await page.goto(url: str) -> a result dict Navigate this page to ``url`` and return raw navigation facts. await page.go_back() -> a result dict Navigate back to the previous page in history. await page.go_forward() -> a result dict Navigate forward to the next page in history. await page.reload() -> a result dict Reload the current page. await page.keep() -> none Retain this page across response cycles for the current chat. await page.wait_for_load_state(state: str = "load", *, timeout: float | None = None) -> none Wait until the page reaches the requested load state. ``networkidle`` semantics depend on the backend: the Playwright backend waits for true network quiescence, while CDP-based backends (cdp, chrome) degrade to ``document.readyState == "complete"`` plus a fixed 500 ms quiet delay and do NOT track in-flight requests — content loaded by late XHR may still be missing when this returns. await page.wait_for_timeout(timeout: float) -> none Sleep unconditionally for *timeout* milliseconds (capped at 30 000). Prefer :py:meth:`locator.wait_for(state, timeout) ` when waiting for a specific DOM condition — it returns as soon as the condition is met and is both faster and more reliable than an unconditional sleep. await page.screenshot() -> a result dict Capture this page to a PNG file in the active workspace. page.get_by_role(role: str, *, name: str | None = None) -> locator Locate elements by accessible role and optional name. page.get_by_text(text: str) -> locator Locate elements by their visible text. page.get_by_label(text: str) -> locator Locate a form control by its associated label text. page.get_by_placeholder(text: str) -> locator Locate an input by its placeholder text. page.locator(selector: str) -> locator Locate elements by a CSS selector when no semantic locator fits. page.frame_locator(selector: str) -> locator Scope subsequent locators to the iframe matching ``selector``. await page.snapshot(query: str | None = None) -> observation (read .text; .match_count when you pass a query) Perceive the page and return readable content in ``.text``. Pass ``query`` to also report ``.match_count``. await page.current_surface() -> surface facts (.url, .title, .load_state) Return this page's current URL, title, and load facts. page.mouse -> coordinate/keyboard input surface (see methods below) Viewport-coordinate input surface. click(x, y) -> a result dict; verify the effect with snapshot(). page.keyboard -> coordinate/keyboard input surface (see methods below) Keyboard input surface. press(key) -> a result dict; verify the effect with snapshot(). page.get_by_* / page.locator(...) return a locator that mirrors a SUBSET of Playwright's Python locator API — the Playwright-shaped part of this SDK: compose/scope (chainable): get_by_role/get_by_text/get_by_label/ get_by_placeholder, locator(sel), filter(...), nth(i), first, last (properties) iframe scope: page.frame_locator(sel).locator(...) (one frame; no nested frames) read (await): count()->int, inner_text()->str, text_content()->str|None, all_text_contents()->list, get_attribute(name)->str|None, input_value()->str, is_visible()->bool, is_enabled()->bool act (await; returns a short evidence line — read .evidence): click(), fill(v), type(t), press(key), check(), uncheck(), set_checked(b), select_option(*v), hover(), dblclick(), scroll(), focus(), blur(), clear(), wait_for(state), screenshot(), bounding_box()->dict|None (viewport ["x"] ["y"] ["width"] ["height"]; None when the element is not visible) strict-mode uniqueness is enforced — act only when the locator resolves to exactly one element (use count() to check). element_handle / raw CDP unavailable. Backend differences (chrome/cdp vs playwright): on chrome/cdp the accessible name is a heuristic (aria-labelledby > aria-label > alt > title > text content) - container elements may match get_by_role(name=) more broadly than under playwright, so strict-mode errors are more likely there; narrow with filter(has_text=) or a more specific role. is_enabled() reflects only the disabled property, not aria-disabled. press() supports a fixed key set: printable characters, Enter, Tab, Escape, Backspace, Delete, Arrow keys, Home/End/PageUp/PageDown, and Control/Shift/Alt/Meta combos - anything else fails with guidance. type() sets the value directly and fires an input event; editors that need real per-key events may not react - prefer fill() where possible. Reading results (read these fields; the type names don't matter): snapshot() -> .text (page text), .match_count (when you pass query) current_surface() -> .url, .title, .load_state page refs -> .id, .url, .title, .active screenshot() -> result dict; read ["path"] bounding_box() -> viewport ["x"] ["y"] ["width"] ["height"]; None when the element is not visible mouse.click()/keyboard.press() -> result dict; fields depend on the backend, so verify with snapshot() actions -> .evidence (a short line saying what happened) locator reads return plain str/int/bool/list directly. If a locator fails, step DOWN one rung (don't jump): 1. semantic page.get_by_role/label/text first choice 2. css page.locator(css) role missing/unstable 3. coordinates use locator.bounding_box() first for an exact, low-cost viewport rectangle; use a screenshot to explore only when the element is absent from snapshot() For captcha/login/2FA or any human-only step: await browser.handoff(reason, instructions) and stop — never automate them.