Ninety-six commits, seven and a half weeks, one person. I built the whole thing, then audited it like someone else had.

I trade, and I kept six tabs open to do it: prices on one site, the economic calendar on another, a journal in a spreadsheet, a Pomodoro timer on my phone, the forum somewhere else again. None of them talked to each other, and every one of them cost a deliberate act of navigation.

Meanwhile the one page a browser guarantees you will see dozens of times a day — the new tab — was blank. That asymmetry is the whole idea. I was not trying to build a trading platform; I was trying to put the five things I already checked onto the surface I already opened.

The audience made it sharper. Persian-speaking traders get English-first tools with a right-to-left afterthought bolted on, and the data they actually care about — the Tehran Stock Exchange, the free-market dollar rate, the coin price — is exactly the data those tools do not carry. Nobody was going to build this for them, so the product had to be Persian-first in its layout, its calendar, its typography and its data sources, not translated into Persian at the end.

Everything below runs in a single page with no build step: plain HTML, CSS and JavaScript, nine independent scripts, no npm, no bundler, no ES modules. Nothing is a framework component; every widget owns its own `localStorage` key and fails on its own without taking the page down.

That constraint was a decision, not an accident. A new tab page has to paint before the user has finished pressing the keyboard shortcut. Anything that needs a bundle to boot has already lost.

Ninety-six commits between 27 June and 18 August 2026, and the shape of that history is the honest part. On day one I built a modular architecture with twenty-one feature modules. Seven weeks later I deleted all twenty-one, because I found they had never been loaded — the page ran on a completely different set of files the whole time. I also built and removed an embedded Telegram sidebar, a widget edit mode with drag-and-resize, and a forex news ribbon. Roughly a fifth of what I wrote never shipped.

Working with Claude Code made me fast, and fast is exactly how you end up with a twenty-one-module architecture nobody loads. So in the last stretch I changed what I was asking for. Instead of more features I asked for an audit, and I gave it one rule: the code is the source of truth, not the documentation, and not what I remembered deciding. Every document in the repo got rewritten from a direct reading of the live code, with that rule printed at the top.

The audit found the dead module tree, four domains in `host_permissions` with no caller left, a permission requested with zero call sites, and twelve of sixteen fetch calls with no timeout — able to hang in a loading state forever, which on an unstable Iranian connection is a normal afternoon. Then I ran seven QA tests against a real browser through Puppeteer, not by reading the source: network conditions emulated over CDP, all eleven storage keys deliberately corrupted, both market-data hosts blocked at the network layer. Four passed, two failed, one passed partially, and the tools are checked into the repo so the results can be reproduced rather than believed.