CLI checking a generated web page w/ zero tests written and zero LLMs.
Drives a real browser, tests every control, and tells you what broke.
Zero false positives & 14s median runtime.
Plugin/Skill support for 15 agentic coding harnesses:
Try assay: https://t.co/XlCazPUIAv
@manish_iitg claude code since it gets the full plugin. the skill plus a stop hook, so every page gets checked at the end of the turn without asking. deepseek harness runs the same hook through its claude code bridge. the rest are just as easy to set up & smooth, they just skip the stop hook
CLI checking a generated web page w/ zero tests written and zero LLMs.
Drives a real browser, tests every control, and tells you what broke.
Zero false positives & 14s median runtime.
Plugin/Skill support for 15 agentic coding harnesses:
Try assay: https://t.co/XlCazPUIAv
@Danjoe_4 neither, there's no baseline or artifact. you run it on a web page, it works out a plan from what's rendered, clicks through everything, and flags where the page contradicts itself, like an undo that's one press behind. nothing to record first, and it never reads the code
@HaseebMir91 yes for any frontend framework, since it reads the rendered page, not the code. react, vue, svelte and plain html all look the same to it, just point it at the built folder like dist. backend is out of scope though, live apis and logins aren't tested. let me know how it goes!
@llmcodes on the benchmark, yes: 0 false alarms across 225 pages, and it fills fields and presses submit like a user would. it works on any page you have the folder for, served locally. what it won't test is a live site's backend, so a form posting to a real server isn't what it's checking
@kumar_ritikk thanks! honestly it doesn't yet. it serves the page from its folder on your machine and drives what's there, so anything behind a real login or a live backend is out of scope. for most pages you build or an agent writes, that's not an issue
@akankshabangwal not quite, it means no tests to write before you can check it. it catches the page contradicting itself, but it can't know what your app is meant to do, so for that you still want real tests. it's for the page an agent wrote 10 seconds ago that nobody's tested yet
@FuzzyKidoo no, on purpose. it doesn't know what your page is meant to do, so a fix on its word could be the wrong fix. it tells you exactly what it pressed and what happened, so you or your agent know where to look instead of re-reading the whole file. in summary it FINDS the bugs :)
@Rahulsinr no hand-written assertions, it derives them from the page. it measures the rendered controls, builds a plan, runs each case in a fresh tab and diffs pixels, text, fields and styles. it flags contradictions, like undo doing nothing then something from the same state