Day 1 - Devlog 1
Date - 20 August 2026
Worked - 4 hours and 50 minutes
Today I started “Fresp” from an empty folder. It’s a local CLI. You run a command on your laptop and it does the work there.
I split the code into folders so it’s not one giant file but structured - core (calls the others), drivers (chrome), vision (screenshots + checks), state (saves results), logger, report, types, config.
Got Node set up (package.json, tsconfig). The logger prints a timestamp plus a message and also writes logs/run.txt. State is a scorecard per page - url, passed, notes, screenshot name. Core loops the pages from a heat sheet (a list of paths).
Then Playwright. Real chrome opens, hits a tiny demo I made (home and about). I broke home on purpose with a 4000px red bar so it should fail. About is normal so it should pass. It waits 3 seconds so you can actually see the page, then takes full-page screenshots into fixtures/baselines.
Vision has one real check, not AI: overflow. It asks chrome if the page is wider than the window (scrollWidth vs clientWidth). Home = true / fail. About = false / ok. That’s the point of today, not just pretty screenshots.
It writes logs/last-run.json and an HTML report you open in the browser, so you can see pass/fail without living in the terminal.
I also wrote docs: ARCHITECTURE, VISION (full product: facts first, then AI taste, no cloud dashboard), ENGINEERING (how the AI path should work later, including prompt cache and skip-AI if there’s no key), and taste.md, which is the judgement file (school site vs portfolio vs ai slop, etc.). Taste file + a vision API later is the plan.
GitHub is atireksd11/fresp-ai-tester. I published [email protected] on npm, so in theory npx fresp-ai-tester works. I tried it from another folder and it crashed because tsx was only a devDependency, so strangers don’t get it. I started 0.1.1 and moving tsx into dependencies. Haven’t published that yet, it might work. There’s also bin/fresp.mjs so npx fresp works inside this repo (had to fix a Windows path with a space in the folder name).
What’s not done: the actual ai taste api, more facts (clip, overlap, contrast), VS Code extension, crawling the whole internet, etc.
Next session: Publish 0.1.1 so npx works anywhere. Rebuild the demo so home has several real bugs (overflow, overlap, clipped text, messy theme) and about stays clean. Add those extra fact checks in vision, expand the json so every failure is listed, and redesign the HTML report (colors, screenshots, fact list, room for ai notes). Put product type and goal on the heat sheet, then start the vision api with taste.md (skip ai if there’s no key).
