[Bug]: Unhandled ValidationError in DispatcherConnection crashes the whole driver process on malformed CDP data (connect_over_cdp)

[Yoshinaga-iwnl]

Version

1.52.0

Steps to reproduce

Connect to a third-party CDP endpoint that reports a malformed value in a network event. Concrete case: Lightpanda <= 0.3.5 serializes response header values / suggestedFilename containing non-UTF-8 bytes (e.g. a Shift_JIS Content-Disposition filename, still common on Japanese government sites) as JSON byte arrays instead of strings (lightpanda-io/browser#2972, lightpanda-io/browser#2992).

# pip install playwright==1.52.0
# lightpanda serve --host 127.0.0.1 --port 19222   (0.3.4, https://github.com/lightpanda-io/browser)
import asyncio, socket, threading
from playwright.async_api import async_playwright

RESPONSE = (
    b'HTTP/1.1 200 OK\r\n'
    b'Content-Type: text/html\r\n'
    b'Content-Disposition: attachment; filename="\x82\xd3\x82\xe9\x82\xb3\x82\xc6.xlsx"\r\n'  # Shift_JIS bytes
    b'Content-Length: 5\r\nConnection: close\r\n\r\nhello'
)

def serve():
    s = socket.socket(); s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
    s.bind(("127.0.0.1", 18080)); s.listen(4)
    while True:
        c, _ = s.accept(); c.recv(65536); c.sendall(RESPONSE); c.close()

async def main():
    threading.Thread(target=serve, daemon=True).start()
    async with async_playwright() as pw:
        browser = await pw.chromium.connect_over_cdp("http://127.0.0.1:19222")
        page = await (await browser.new_context()).new_page()
        await page.goto("http://127.0.0.1:18080/", wait_until="domcontentloaded")

asyncio.run(main())

Expected behavior

A validation failure on data coming from the browser fails the corresponding request/event (or at worst closes that one connection). page.goto raising an error would be fine.

Actual behavior

The ValidationError thrown while building the outgoing protocol message is uncaught inside the dispatcher, and the whole Node driver process exits. Every unrelated browser/connection served by that driver dies with Connection closed while reading from the driver. In our production crawler (many independent CDP connections sharing one Python async_playwright instance) a single malformed HTTP response on one site took down the entire fleet.

ValidationError: headers[7].value: expected string, got object
    at tString (.../driver/package/lib/protocol/validatorPrimitives.js:75:9)
    ...
    at DispatcherConnection.sendCreate (.../lib/server/dispatchers/dispatcher.js:172:19)
    at new Dispatcher (.../lib/server/dispatchers/dispatcher.js:75:24)
    at new ResponseDispatcher (.../lib/server/dispatchers/networkDispatchers.js:70:5)
Node.js v22.14.0

(With Lightpanda 0.3.5 the same crash occurs via the download path: ValidationError: suggestedFilename: expected string, got object at DispatcherConnection.sendEvent -> PageDispatcher._dispatchEvent.)

Additional context

I understand third-party CDP endpoints are not officially supported targets for connect_over_cdp. Filing this because the failure mode seems disproportionate regardless of the data source: one malformed event escapes DispatcherConnection.sendCreate/sendEvent as an uncaught exception and kills the shared driver process. Catching validation errors at the dispatch boundary (dropping/sanitizing the offending event or failing the associated call) would keep the blast radius to one page/connection.

Environment: Python 3.12, playwright 1.52.0 (bundled driver, Node.js v22.14.0), reproduced on macOS arm64 and Ubuntu (mcr.microsoft.com/playwright/python:v1.52.0-noble).

[HarshDevelops]

Hi — I'd like to be assigned or explicitly approved for this issue. Honest
summary of what I have and have not yet done:

Inspected (no reproduction yet):

• Read the issue body and performed read-only source inspection. I have not yet
  run Playwright against Lightpanda 0.3.5 here; the malformed-bytes CDP repro
  environment is not yet set up. So I have not reproduced the crash.

Likely affected path:

• packages/playwright-core/src/client/connection.ts and the
  DispatcherConnection connect_over_cdp path, plus the network response
  parser feeding suggestedFilename / response header values back to
  page-side callers.

Proposed narrow fix (still subject to maintainer input — see below):

• Treat per-field ValidationError from a non-spec CDP source as a
  diagnostic log, not a driver kill. Specifically: when serializing
  outbound CDP events from the dispatcher, lossy-decode string fields
  rather than rejecting the whole event; let the upstream code already
  do this for response headers, but extend the same to suggestedFilename.

Intended regression test:

• A unit test that injects a well-formed CDP Network.responseReceived
  payload whose headers contain a Content-Disposition value as raw
  UTF-8 bytes (Shift_JIS fallback for the same character sequence), and
  asserts (a) the driver does not crash and (b) callers see the
  decoded text rather than throwing.

Open question for a maintainer: the project may prefer rejecting
non-spec CDP traffic outright instead of lossy-decoding. If that's the
intent, the narrow fix is a documentation/comment update plus a type
test, and I'll switch plans.

If you're willing to assign me, I'll then clone the tree, add the
proposed test (failing first), then propose the fix and run the
existing connect_over_cdp tests as gate.

Thanks.