An older Tillping guide 🙂 Stripe's usual signature check is synchronous and leans on Node crypto, so on an Edge runtime it throws. Most handlers log that as a bad signature, and you go hunting for the wrong secret. The async check with the Web Crypto provider fixes it.
AI coding agents comment code like it's their first day. "// loop over the users" sits right above a loop over the users 😄 The comment I actually want explains why that one odd line exists, and that's exactly the one they skip.
Congrats. You built a CRM with Claude in a weekend.
So did 40,000 other guys.
It looks beautiful. Works perfect.
Now add 300 clients, 4 techs, recurring jobs, payments, and follow ups, and bugs..
Watch it fall apart.
Anyone can build day 1.
One quiet trap from a Tillping guide 🙂 Lemon Squeezy counts only HTTP 200 as success. A handler that does the work and answers 201 Created or 204 No Content still gets retried, and after a few tries the event is marked failed. Your own tests pass, so it's easy to miss.
AI agents treat a codebase like a buffet they've never seen. There's already a perfectly good date helper two folders over, so naturally they write a fresh one 😄 Code search is the most underrated tool in the whole agent loop.
@linw2949 Yep, I keep the event ID too. Providers love sending the same thing twice, so the handler just shrugs and moves on 🙂 And totally agree on never matching by email.
One quiet way a payment goes missing has nothing to do with delivery. The webhook arrives, the signature checks out, the handler runs and returns 200, and nobody gets upgraded because the app can't tell which user paid.
It usually happens when the handler matches on the email from checkout. People pay with a work address, a typo, or a different account than the one they signed up with, and the lookup quietly finds nobody. The logs look clean.
The sturdier pattern is to pass your own user id into the checkout as custom data and read it back from the event. Then check what your handler does when no user matches. That case should raise an alert and keep the event, not return quietly 🙂
A webhook endpoint has two audiences that look nothing alike. Your users arrive in browsers. Your payment provider arrives as a server with no cookies and no JavaScript, which is exactly what bot protection is built to stop.
@amirsadjady I'd settle for one that deletes a single unused file and then looks proud of itself. A whole folder might be aiming a bit high, but I'd still pay it 😄
AI coding agents almost never suggest that something shouldn't exist. Every problem gets more code. Some of the best commits I've reviewed were mostly red lines, and I'm still waiting for the agent that proposes one without being asked 😄
Worth a look in your provider's delivery log. If the response body is HTML, something in front of your code answered. Exempt the webhook path from bot and rate rules, keep signature checks as the real gate, and resend a test event after every security change.
So a new firewall rule, bot challenge or rate limit goes live, the site feels fine for humans, and the provider quietly starts getting a 403 or a challenge page. Payments keep succeeding on their side while your app stops hearing about them.
My favorite Tillping check is the boring one. It sends your webhook URL a POST with no signature, then one with a fake signature. Both should bounce. If either gets a 200, anyone who finds that URL can tell your app someone paid 🙂 Free check at https://t.co/IqcIbtbAKT
@silic2023 That's a great review question. "Does it pass" only tells you the test and the code agree right now. "What would it still catch" tells you whether the test is worth keeping. I might borrow that one.
AI coding agents have a real talent for green builds. Ask one to fix a flaky test and sometimes the fix is a quietly deleted assertion 😄 Green and correct are still two different words, and the agents haven't fully noticed yet.
@ValidatorTobias Both, honestly. If the agent touched a test, I want to see it fail on the old code before I trust the green. And any diff that removes or loosens an assertion gets read line by line. That's usually where the "fix" is hiding 🙂