Benchmarked patchright against a consistency scorer: 100/100 host-adjusted, the only headed stack with no webdriver deduction

[liarjsdev]

I benchmarked twelve Chromium automation stacks last week on one machine, in one sitting, with one scorer, and patchright came out at the top of the table, so I wanted to hand you the numbers rather than just say so.

Host-adjusted scores (higher is better):

stack	score	webdriver	remaining deductions
patchright, headed Chrome	100	ok	none
playwright + Chrome + args, headed	100	ok	none
rebrowser + args, headed	100	ok	none
playwright + Chrome, headed	60	bad	webdriver
rebrowser, headed	60	bad	webdriver
puppeteer, headed	54	bad	webdriver
puppeteer-extra-stealth, headed	30	ok	gpu-triad, worker-consistency, gl-ext-formfactor, webgpu-features
playwright + Chromium, headless	20	bad	webdriver, headless-ua, perm-notif, chrome-object

Method, so you can dispute the number rather than take my word for it: each stack was launched with --remote-debugging-port, then scored by attaching over CDP, with the probes run on an https:// page rather than about:blank (several checks are inert on about:blank and scoring there quietly inflates everything). Two of the deductions on every stack are this laptop's own hardware, not the stack's, so the baseline is not 100 and I report both the raw and the host-adjusted column.

Two things in there that may be worth your attention:

• patchright reaches 100 without needing the extra launch args, which is the difference between it and the two other stacks on 100.
• puppeteer-extra-stealth is the interesting failure: it clears the webdriver flag, then loses four separate checks on GPU and worker consistency. Patching the obvious tell and leaving the derived surfaces inconsistent scores worse than patching nothing, because a mismatch between what a page is told and what the worker or the GPU string says is itself a signal.

The reason I am filing this rather than just posting it somewhere: your README's detector table is a list of pass/fail verdicts, and pass/fail cannot express that last point. The scorer I used is https://liarjs.dev/ - it checks whether a browser's JavaScript story agrees with itself and with what the network layer saw, and it outputs a number plus the individual deductions. It runs as npx liarjs --cdp http://127.0.0.1:9222 --page https://example.com/ --json, so it fits next to pytest rather than next to a screenshot. If you think it earns a row in that table, that would be welcome; if you would rather not add a checker you did not choose, the numbers above are yours to use regardless.

Full write-up with the per-check deductions: https://liarjs.dev/blog/stealth-stacks-measured/

[ttlns]

@liarjsdev Please avoid opening issues with non-issue content.