My rate limiter had been silently broken. Three times.

My rate limiter on electricaltoolbox.com has failed open before. Silently. Three separate times, spread out enough that nobody noticed a pattern until this week.

Here is how it breaks. The limiter reads and writes a Cloudflare KV binding by name. If that binding is ever misconfigured or gets renamed, the check that is supposed to block abuse just stops running. No error. No log line. Every request succeeds like the limiter was never there.

I had two tests already. A unit test that always passed, because it always handed the limiter a working fake binding. And a synthetic check on the live endpoint that also always passed, because it only ever sent one request, and one request never trips a rate limit whether the limiter is doing anything or not.

Neither test could have caught the exact failure that had already happened three times. Both proved the algorithm was correct. Neither proved the algorithm was running.

So this week I wrote a different check. It sends the real configured limit’s worth of requests to the live endpoint, one after another, then sends one more. If the first batch does not all succeed, something upstream broke. If the extra one is not rejected, the limiter is failing open right now, in production, and nothing else would tell me.

Ran it against the real site: ten requests through clean, the eleventh came back rejected with a 429. Wired it into a daily check that pages me if it ever comes back wrong.

The lesson generalizes past this one script. A test that always hands your code a working dependency will never catch the day that dependency is missing. If you want to know a safety check actually works, you have to test the check, not just the logic behind it.

One release worth flagging from this week: Anthropic’s newest flagship model held pricing steady at $5 and $25 per million tokens and added a new top agentic tier for coding and research, with five effort levels you can set per task instead of picking one mode for everything. For anyone running an AI pipeline daily, that is a real lever. Most tasks do not need the model thinking hardest about every step, and paying for that anyway is a cost most people never measure.

electricaltoolbox.com also shipped four new calculators and two supporting guides this week, plus a footer fix that had three tools filed under the wrong category.

Grab the free Claude Code course: https://course.jaredhebb.com