Scrape an SPA that changes URL via history API

A single-page app does not load a new document when you move between views. It rewrites the address bar with history.pushState, swaps some DOM, and fetches the new view’s data over XHR. The URL changes, but nothing that Playwright calls a navigation ever happens. If your scraper waits for a load or a navigation event, it waits forever.

This page is the pattern that works: wait on route and DOM markers instead of load events, capture the XHR that actually carries the view’s data, and re-query the page after every route change because the old handles are now pointing at a view that no longer exists.

Why a load event never fires in a single-page app

history.pushState fires none of the events Playwright waits on for a navigation: no load, no domcontentloaded, no framenavigated, and no networkidle transition. pushState is a JavaScript call that edits the session history and the address bar in place - the document object is the same one it was a moment ago, and no request is tied to the URL change, so there is nothing for networkidle to settle on either.

Compare that with a classic multi-page site: click a link, and the browser tears down the document, makes a request, and builds a new one. Playwright has an event for every step of that: load, domcontentloaded, a framenavigated, and the networkidle state settles once the requests stop. Concretely, all four of these hang until they time out on a pushState route:

# every one of these waits for an event a pushState app never emits
page.wait_for_load_state("load")
page.wait_for_load_state("networkidle")
with page.expect_navigation():
    page.click("a[href='/app/items']")
page.on("framenavigated", handler)   # never called for a pushState route

Here is what actually happens on each signal, classic navigation versus a pushState route:

Signal	Classic multi-page navigation	`history.pushState` route
`load`	Fires	Never fires
`domcontentloaded`	Fires	Never fires
`framenavigated`	Fires	Never fires
`networkidle`	Settles once requests stop	No transition (the URL change makes no request)
`page.url`	Updates	Updates immediately
The DOM	A brand-new document	Same document, one subtree swapped

The URL did change. page.url will read the new path immediately after the click. What did not change is anything Playwright treats as a navigation, which is exactly the thing your wait was hanging on. The general question of what to wait for when the usual load states lie has its own page; an SPA route is the sharpest case of it.

The two things that genuinely change on a pushState route are the URL and the DOM. Wait on those directly.

Playwright’s wait_for_url accepts a glob or regex and resolves when page.url matches, whether the change came from a real navigation or from pushState. Pair it with a wait_for_selector for a marker the new view renders, so you are certain the view is present and not merely that the address bar moved ahead of the render:

from invisible_playwright import InvisiblePlaywright

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com/app")

    # the initial document load is a real navigation, so goto is enough here.
    # confirm the app has mounted before touching it.
    page.wait_for_selector("[data-view='home']")

    # clicking a nav link swaps the view via pushState. wait for the URL to
    # move AND for the destination view's marker to render.
    page.click("a[href='/app/items']")
    page.wait_for_url("**/app/items")
    page.wait_for_selector("[data-view='items'] .item")

The browser object here is a real Playwright Browser, so wait_for_url, wait_for_selector, click and everything else behave exactly as documented upstream. The only difference from plain Playwright is the two-line launch and a fingerprint underneath.

If the app renders its route marker before the data arrives (a skeleton or a spinner), wait for a marker that only the loaded state shows, or use wait_for_function to poll for a condition the framework sets once the view is ready:

# wait for the item list to actually contain rows, not just exist empty
page.wait_for_function("() => document.querySelectorAll('[data-view=\"items\"] .item').length > 0")

Capture the XHR that carries the view’s data

The DOM is the rendered result. The clean data is in the XHR the view made to fetch itself, and reading that response is far more reliable than scraping rendered HTML that the framework may re-render underneath you.

page.expect_response opens a wait for a matching response and lets you trigger it in the same block. Match on the API path and, if you want to be strict, on the request’s resource type so a same-URL document does not satisfy it:

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com/app")
    page.wait_for_selector("[data-view='home']")

    # arm the response wait BEFORE the click that causes the fetch,
    # or the response can arrive before you started listening
    with page.expect_response(
        lambda r: "/api/items" in r.url and r.request.resource_type == "xhr"
    ) as resp_info:
        page.click("a[href='/app/items']")

    page.wait_for_url("**/app/items")
    response = resp_info.value
    data = response.json()          # the view's data, already structured
    print(len(data["items"]), "items on this route")

Arm the listener before the action that triggers the request. If you click first and call expect_response after, the response can land in the gap and you wait for one that already happened. Capturing XHR and JSON API responses covers matching, buffering several responses per route, and reading bodies that are not JSON.

Re-query after every pushState

This is the mistake that produces the strangest bugs. A pushState route replaces the view’s DOM but keeps the same page and the same execution context, so any element handle, ElementHandle, or evaluated reference you captured on the old view is now attached to nodes that have been removed. It does not error loudly. It returns stale or empty results, which reads like a scraping bug rather than a lifecycle one.

The discipline: query fresh after each route change. Wrap a route transition in one helper so re-querying is not something you can forget:

def goto_route(page, link_selector, url_glob, view_marker, api_fragment):
    """Follow an in-app pushState link and return the view's XHR payload."""
    with page.expect_response(
        lambda r: api_fragment in r.url and r.request.resource_type == "xhr"
    ) as resp_info:
        page.click(link_selector)
    page.wait_for_url(url_glob)
    page.wait_for_selector(view_marker)      # fresh query, current view
    return resp_info.value.json()


with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com/app")
    page.wait_for_selector("[data-view='home']")

    payloads = []
    for path in ["items", "orders", "settings"]:
        payload = goto_route(
            page,
            link_selector=f"a[href='/app/{path}']",
            url_glob=f"**/app/{path}",
            view_marker=f"[data-view='{path}']",
            api_fragment=f"/api/{path}",
        )
        payloads.append(payload)

Prefer page.locator(...) over a stored ElementHandle across route changes: a locator re-resolves its selector every time you use it, so it naturally points at the current view, while a handle is frozen to the node it was created from.

One long-lived context, and the caveat that comes with it

Here is the stealth upside, and it is genuine. Because a pushState route is not a full navigation, one browser context holds the entire multi-view session. There is no document teardown, no new page, no fresh identity draw between views. The fingerprint that was generated when the context launched persists across every route for free.

With seed=42 you can watch this hold. Read a stable surface such as the canvas hash or a FingerprintJS visitor ID on the first view, walk twenty routes, and read it again: it is identical on every route, because it is the same context throughout. A design that spun up a fresh context per view would instead present twenty independently generated machines to any scoring endpoint, and a session that changes machine between clicks is a louder signal than any single fingerprint.

The caveat is the same coin’s other face. That same longevity means execution-context state and page state accumulate across a long session: listeners, timers, retained references, and memory the app never releases between views. Two habits keep it honest. Re-query after each route change, as above, so you are never reading a dead view. And be ready for the app itself to blow away the execution context mid-session, which surfaces as an "Execution context was destroyed" error on your very next call:

from playwright.sync_api import Error as PlaywrightError

def read_current_view(page, view_marker):
    try:
        page.wait_for_selector(view_marker)
        return page.locator(view_marker).inner_text()
    except PlaywrightError as exc:
        if "Execution context was destroyed" in str(exc):
            # the app re-created the context (a hard in-app redirect or a
            # framework remount). settle on the new one and query again.
            page.wait_for_load_state("domcontentloaded")
            page.wait_for_selector(view_marker)
            return page.locator(view_marker).inner_text()
        raise

That error is not always a bug in your code. In an SPA it is frequently the application doing a hard remount or an in-app redirect underneath you. Why “Execution context was destroyed” happens and how to tell the two causes apart separates the ordinary race from the app-triggered one, because the fix differs.

Conclusion

An SPA that routes with history.pushState breaks every wait that assumes a navigation, because there is no navigation: no document load, no navigation event, no networkidle. Wait on the URL and a DOM marker instead, capture the XHR that carries the view’s data rather than scraping re-rendered HTML, and re-query after each route so you are never holding a handle to a view that no longer exists. The long-lived context that makes all of this necessary is also what keeps one consistent identity across the whole session for free, as long as you re-query and stay ready for the app to reset the context on you.

Short answers to the questions that lead here

Why does wait_for_load_state hang on a single-page app? Because pushState changes the URL without loading a document, so load, domcontentloaded and networkidle never fire for the route change. Wait for the URL and a DOM marker instead.

How do I wait for a pushState URL change in Playwright? page.wait_for_url("**/app/items") resolves on a pushState change as well as on a real navigation. Pair it with a wait_for_selector for a marker the new view renders.

Should I scrape the rendered DOM or the XHR? The XHR, when you can. Arm page.expect_response before the click that triggers it, match on the API path, and read response.json(). It is cleaner and more stable than HTML the framework may re-render.

Why are my element handles suddenly stale after clicking a link? A route change replaces the view’s DOM while keeping the same page, so handles created on the old view point at removed nodes. Re-query after each route, and prefer page.locator(...), which re-resolves every use.

What causes “Execution context was destroyed” in an SPA? Often the app itself, doing a hard remount or in-app redirect, not the classic navigation race. Catch it, settle the page, and re-query the current view.

Does one context for the whole app help or hurt stealth? It helps: one identity persists across every route with no per-view redraw. The cost is accumulated page and context state, which is why re-querying and watching for a destroyed context are part of the pattern.

Sources

Playwright’s wait_for_url, expect_response, locator and wait_for_function APIs, read from their documented behaviour rather than from a verdict: Playwright Python Page API reference.
The history.pushState behaviour itself - it edits the session history and the address bar without asking the browser to load a document: MDN, History.pushState().
This project’s own measurements of a single context across a multi-route session: with a fixed seed, the canvas hash and visitor ID stay identical on every route.

Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. The stale-handle bug in section four cost me an afternoon before I understood it was a lifecycle problem, not a selector one.