10 Lenses on Any Codebase, in 1 Command
I don't trust "looks good to me." I've shipped bugs that looked good to me. So I built something that grills a repo the way I wish someone would grill mine: the same 10 lenses, in the same order, with no mood involved and no credit given for code I happen to be proud of.
What it actually is
/repo-audit points at any repo - a local path, an owner/repo, a GitHub URL - and reverse-engineers it across 10 lenses: architecture, an honest good/bad/ugly read, infra, security, performance, code quality, tests, docs, UI/UX, and what the thing actually does. Out comes 1 clean HTML report: an architecture diagram drawn from the real wiring, an A-F grade per lens, a scorecard, and a ranked list of the top fixes. It opens the report and drops a copy in my notes. It never touches the code - read-only, always.
The lens I'm proudest of is the one with no grade: Good, Bad, Ugly. 3 columns, no letter. I took the grade off on purpose, because the second you slap a "B+" on something, people read the letter and skip the honesty. The point is the ugly column - the stuff you'd be embarrassed to explain out loud in a review.
In practice
Last month I inherited a repo with a README that described files that didn't exist. I ran /repo-audit on it. The architecture diagram showed me the real shape in about the time it took to get coffee. The docs lens flagged the stale README. Security caught a missing header. And the top-fixes list told me where to start instead of guessing. That's the whole pitch: stop guessing where to start.
The rule that keeps it honest is that every node and arrow in that architecture diagram has to trace to a real line of code. No plausible-looking fiction. If it can't find where something is wired up, it says "could not confirm" instead of drawing a clean arrow that lies. An honest gap beats an invented connection.
What I learned
The model finding issues was never the hard part. Ranking them - what's real, what's worth doing first, what's just noise - that's the product. A report that lists 40 things is wallpaper. A report that tells me the 3 that matter is a tool I actually open. So that's what I optimized for: not more findings, fewer better ones.
The skill behind this is open source. Clone the 1 folder you need from bunlongheng/audits, point it at your own work, and see what 10 lenses say about the repo you know best.