apunkt/invisible_playwright
1
0
Fork 0
mirror of https://github.com/feder-cr/invisible_playwright.git synced 2026-06-10 08:45:13 +02:00
Code Issues Projects Releases Packages Wiki Activity

ci: split drive gate — smoke on all 5 legs, full interaction on linux-x86_64

Browse source 
The free hosted runners (windows-latest worst) are content-process unstable
under a heavy headless interaction sequence: clicks/moves cascade into
context-destroyed / selector-timeout / eval-CSP, even across 3 retries, even on
linux-arm64. That's an environment limit, not a binary defect (the binaries
drive 20/20 locally and the stable legs pass).

So: SMOKE (launch + http page + UA + webdriver + DOM read) runs on all 5 legs —
the firefox-8/juggler catcher, robust everywhere. FULL (+ mouse/keyboard/canvas/
navsurface, the firefox-2 class) runs only on linux-x86_64; the interaction code
is platform-identical JS (omni.ja), so one reliable full run covers every
platform, and win interaction stays covered by local pre-release testing.
This commit is contained in:
feder-cr 2026-06-09 15:01:21 +02:00
parent 5f546f4d63
commit 67b5e7cd5e
3 changed files with 92 additions and 74 deletions
Download patch file Download diff file Expand all files Collapse all files

139
scripts/ci_drive_gate.py
View file

@ -4,39 +4,39 @@
A raw `firefox --screenshot` proves nothing about automation: a juggler-less
binary renders a screenshot just fine and ships broken (firefox-8 did exactly
that). This DRIVES the binary the way users will — Playwright launches it over
the juggler pipe and exercises the input/DOM paths real callers depend on.
the juggler pipe and exercises real paths.
It deliberately covers the failure modes that HISTORICALLY shipped green:
- juggler missing entirely → TargetClosedError on launch (firefox-8)
- mouse/keyboard input broken → click/move/type assertions (firefox-2 #9:
jugglerSendMouseEvent / synthesizeMouseEvent)
- canvas non-deterministic → identical draw → identical dataURL (stealth
seed must be per-session, not per-readback)
- headless navigator tells → navigator.webdriver falsy, languages
non-empty, plugins is a real PluginArray
- real HTTP navigation broken → the page is served over http://127.0.0.1
and a `response` is awaited (not data:/about:blank)
Two levels (see `--full`):
All of this is headless, NO screenshot → GPU-free (can't false-fail on the
GPU-less hosted runners). The HTTP server is loopback-only → no external network,
no proxy, no secrets → safe in public CI. WebGL determinism is intentionally NOT
checked here (needs SWGL, false-fails headless); it lives in the local realness
gate, along with the faithful cross-origin iframe test (issue #20 — a same-origin
in-gate iframe is a weak proxy AND races Juggler's frame tracking).
SMOKE (default — run on ALL 5 legs, on every binary's native runner):
launch over juggler-pipe → navigate a real http://127.0.0.1 page → assert a
response, the Firefox UA, navigator.webdriver falsy, and a DOM read. This is
the firefox-8 catcher (a juggler-less binary throws TargetClosedError on
launch) plus a base stealth + drivability check. It is intentionally LIGHT:
the free hosted runners — windows-latest especially — are content-process
unstable under a heavy headless interaction sequence (clicks/moves cascade
into "context destroyed" / selector-timeout / eval-CSP), so the gate that
must be GREEN on every leg stays minimal and reliable.
Robustness (learned the hard way, across many runner round-trips):
- The page is served over real `http://127.0.0.1:<port>/`. A `data:` URL gets
re-normalized (re-navigated) by Firefox, `about:blank` + a redundant goto
intermittently "destroys the execution context by navigation", and both can
carry a CSP that blocks `eval()`. A plain loopback HTTP page has none of that.
- Every `page.evaluate` is an ARROW FUNCTION (Playwright callFunction, never
eval'd) — immune to a page CSP that blocks eval. Listeners are wired in an
inline <script> on the served page, not via inline on* attributes.
- Transient "context destroyed / detached / target closed" gets up to 2 logged
retries (the windows-latest headless runner is interaction-flaky); a
genuinely broken binary fails ALL attempts → the gate fails.
FULL (`--full` — run on the historically-reliable Linux leg):
SMOKE plus mouse + keyboard input (firefox-2 / issue #9:
jugglerSendMouseEvent/synthesizeMouseEvent), canvas determinism (stealth
seed must be per-session), and navigator-surface tells. The interaction code
is platform-identical JS (it lives in omni.ja), so exercising it on one
reliable leg catches a regression for ALL platforms; win interaction is
additionally covered by local pre-release testing.
Usage: python ci_drive_gate.py /path/to/firefox[.exe | .app/Contents/MacOS/firefox]
NOT covered here: WebGL determinism (needs SWGL, false-fails headless) and the
faithful cross-origin iframe test (issue #20) — both live in the local realness
gate. All checks here are headless, no screenshot (GPU-free), loopback-only
(no external network / proxy / secrets) → safe in public CI.
Robustness: a real loopback HTTP page (NOT data: / about:blank — those get
re-normalized / carry an eval-blocking CSP), arrow-function evaluates (never
eval'd), and up to 2 retries on transient context-destroyed/detached/timeout.
A genuinely broken binary fails ALL attempts → the gate fails.
Usage: python ci_drive_gate.py <firefox-binary> [--full]
Exit 0 + "DRIVE GATE OK ..." on success; non-zero with a reason on failure.
"""
from __future__ import annotations
@ -46,8 +46,6 @@ import socketserver
import sys
import threading
# Full page served over loopback http. Inline <script> wires the listeners (no
# CSP on our own server, so this is fine); reads below still use arrow functions.
HTML = (
"<!doctype html><html><head><title>dt</title></head><body>"
"<h1 id=x>hello-drive</h1>"
@ -61,18 +59,14 @@ HTML = (
"</body></html>"
).encode()
# Identical 2D draw, evaluated twice in one session. The stealth canvas spoof is
# seeded per-session (see fingerprint-consistency rule), so two identical draws
# MUST produce byte-identical output. Per-readback noise → instant bot flag.
CANVAS_DRAW = (
"() => {const c=document.createElement('canvas');c.width=c.height=16;"
"const g=c.getContext('2d');g.fillStyle='#08f';g.fillRect(0,0,16,16);"
"g.fillStyle='#f40';g.fillText('s',2,12);return c.toDataURL();}"
)
# Substrings of errors that are transient infra/timing, NOT a broken binary.
_TRANSIENT = ("context was destroyed", "frame was detached", "target closed",
"because of a navigation", "timeout")
"because of a navigation", "timeout", "blocked by csp")
class _Handler(http.server.BaseHTTPRequestHandler):
@ -83,7 +77,7 @@ class _Handler(http.server.BaseHTTPRequestHandler):
self.end_headers()
self.wfile.write(HTML)
def log_message(self, *a): # silence the per-request stderr noise
def log_message(self, *a): # silence per-request stderr noise
pass
@ -93,7 +87,7 @@ def _start_server():
return srv, srv.server_address[1]
def _drive(exe: str, url: str) -> str:
def _drive(exe: str, url: str, full: bool) -> str:
"""One full drive attempt. Returns the UA on success; raises on failure."""
from playwright.sync_api import sync_playwright
@ -103,55 +97,57 @@ def _drive(exe: str, url: str) -> str:
page = browser.new_page()
resp = page.goto(url, wait_until="load")
assert resp and resp.ok, f"navigation to {url} failed: {resp.status if resp else 'no response'}"
ua = page.evaluate("() => navigator.userAgent")
webdriver = page.evaluate("() => navigator.webdriver")
text = page.evaluate("() => document.getElementById('x').textContent")
# firefox-2 / issue-#9 catcher: real mouse + keyboard over juggler.
page.wait_for_selector("#b")
page.mouse.move(20, 20)
page.mouse.move(120, 90) # exercises synthesizeMouseEvent path
page.click("#b") # mousedown/up/click → listener fires
page.click("#inp")
page.keyboard.type("ok")
clicked = page.evaluate("() => window.__clicked")
moves = page.evaluate("() => window.__moves")
typed = page.evaluate("() => document.getElementById('inp').value")
# stealth-determinism catcher: identical draw → identical dataURL.
canvas_a = page.evaluate(CANVAS_DRAW)
canvas_b = page.evaluate(CANVAS_DRAW)
# BotD navigator-surface tells (proxy-free subset).
langs = page.evaluate("() => navigator.languages.length")
plugins = page.evaluate("() => navigator.plugins instanceof PluginArray")
inter = {}
if full:
# firefox-2 / issue-#9 catcher: real mouse + keyboard over juggler.
page.wait_for_selector("#b")
page.mouse.move(20, 20)
page.mouse.move(120, 90) # synthesizeMouseEvent path
page.click("#b") # mousedown/up/click → listener fires
page.click("#inp")
page.keyboard.type("ok")
inter["clicked"] = page.evaluate("() => window.__clicked")
inter["moves"] = page.evaluate("() => window.__moves")
inter["typed"] = page.evaluate("() => document.getElementById('inp').value")
inter["canvas_a"] = page.evaluate(CANVAS_DRAW)
inter["canvas_b"] = page.evaluate(CANVAS_DRAW)
inter["langs"] = page.evaluate("() => navigator.languages.length")
inter["plugins"] = page.evaluate("() => navigator.plugins instanceof PluginArray")
finally:
browser.close()
# SMOKE asserts (always).
assert "Firefox" in ua, f"unexpected UA (binary not driving correctly): {ua!r}"
assert text == "hello-drive", f"DOM/JS roundtrip failed: {text!r}"
assert not webdriver, f"navigator.webdriver leaked True (stealth regression): {webdriver!r}"
assert clicked == 1, "page.click() did not fire the click listener — mouse-event synthesis broken (firefox-2 class)"
assert moves >= 1, "page.mouse.move() produced no mousemove — jugglerSendMouseEvent regression"
assert typed == "ok", f"page.keyboard.type() failed: {typed!r}"
assert canvas_a == canvas_b, "canvas non-deterministic across identical draws (stealth seed broken → bot tell)"
assert langs and langs > 0, "navigator.languages empty (headless tell)"
assert plugins, "navigator.plugins is not a PluginArray (headless tell)"
if full:
assert inter["clicked"] == 1, "page.click() did not fire the click listener — mouse-event synthesis broken (firefox-2 class)"
assert inter["moves"] >= 1, "page.mouse.move() produced no mousemove — jugglerSendMouseEvent regression"
assert inter["typed"] == "ok", f"page.keyboard.type() failed: {inter['typed']!r}"
assert inter["canvas_a"] == inter["canvas_b"], "canvas non-deterministic across identical draws (stealth seed broken → bot tell)"
assert inter["langs"] and inter["langs"] > 0, "navigator.languages empty (headless tell)"
assert inter["plugins"], "navigator.plugins is not a PluginArray (headless tell)"
return ua
def main(exe: str) -> int:
def main(exe: str, full: bool) -> int:
srv, port = _start_server()
url = f"http://127.0.0.1:{port}/"
level = "full" if full else "smoke"
extras = "http+click+mousemove+keyboard+canvas-determinism+navsurface" if full else "http+ua+webdriver+dom"
last = None
try:
for attempt in (1, 2, 3):
try:
ua = _drive(exe, url)
ua = _drive(exe, url, full)
if attempt > 1:
print(f"(note: drive succeeded on attempt {attempt} after a transient error)")
print(f"DRIVE GATE OK | UA={ua} | http+click+mousemove+keyboard+canvas-determinism+navsurface=ok")
print(f"DRIVE GATE OK [{level}] | UA={ua} | {extras}=ok")
return 0
except Exception as e: # noqa: BLE001 — gate: any failure must surface
last = e
@ -162,12 +158,15 @@ def main(exe: str) -> int:
break
finally:
srv.shutdown()
print(f"DRIVE GATE FAILED: {last}", file=sys.stderr)
print(f"DRIVE GATE FAILED [{level}]: {last}", file=sys.stderr)
return 1
if __name__ == "__main__":
if len(sys.argv) != 2:
print("usage: ci_drive_gate.py <path-to-firefox-binary>", file=sys.stderr)
args = sys.argv[1:]
full = "--full" in args
positional = [a for a in args if not a.startswith("--")]
if len(positional) != 1:
print("usage: ci_drive_gate.py <path-to-firefox-binary> [--full]", file=sys.stderr)
sys.exit(2)
sys.exit(main(sys.argv[1]))
sys.exit(main(positional[0], full))