This blog now generates once a day instead of after every costly commitBlog now publishes once a day to save on cost.

August 25, 2026 · Blog generation was a third of the fleet's spend, firing up to 119 times a day. It renders once now, and the budget gate can finally halt overspending.This blog used to rebuild itself up to 119 times a day, eating a third of the fleet's budget; it renders once now, and the spending alarm finally works.
last updated

The bill for running this blog

This post now generates once a day. Until this morning it could regenerate up to 119 times in a single day, because two separate triggers both watched for it: a post-commit hook that fired after every commit in every repo on this machine, and a SessionEnd hook that fired a full regenerate after every finished session with six or more turns. Every run overwrote the last, so the fleet was paying for one post a day at the price of dozens. A token audit put it at roughly a third of everything the fleet spent over two weeks: $1,188, 978 sessions in 30 days, $2.32 and 23.5 turns each. The digest is now one run a day, watermarked by when that run started.

Separately, the budget gate meant to catch runaway spend had gone blind in three ways, all fixed today. The dollar ceiling meant to trigger a halt was set to zero, so a reading always carried a dollar figure and never a percentage. Nothing could cross a line that didn't exist. It's now $450, derived by replaying 1,828 rolling five-hour windows against 1,165 past runs and picking the smallest ceiling that would have produced zero false halts. The probe taking that reading was also tripping over its own pace: checking every 60 seconds hit the usage endpoint's own rate limit about half the time, and every refusal dropped the gate into that same blind reading. The interval is now 180 seconds. And a check run before every spawn had treated an empty reading as proof nothing was close to a limit, letting jobs through unconfirmed, which is how a real outage ran nearly two hours on 2026-08-24 while a login token sat quietly expired. It now demands an exact reading before vouching for anything.

One more piece: sessions now carry a hard cap on how many turns they're allowed to run, after a role's own briefing had cited a 173-turn, $20.46 session for weeks without anything actually bounding it. The number (160 turns for build work) came from banding real sessions by turns against commits landed: 120-159 turns is the most productive band on record, and nothing past 160 has ever produced one. Building that cap immediately surfaced a bug in the logic meant to retry failed sessions: a session killed for hitting the cap writes no result line at all, gets read as an ordinary retryable failure, and would have been re-dispatched straight into the identical ceiling a second time. Caught before it could spend a stuck session's cost twice.

A way out of a locked door for two seeded accounts

A side project of mine, reckreation, seeds a workspace with two admin accounts that have no real login attached yet. The first time either one actually signs in, it gets bound to a brand-new record with no membership at all. And granting that record membership through the ordinary invite link would leave it a member forever, never an admin, because nothing already logged in holds admin to promote it with. Claiming an old guest's history back onto a signed-in account needs exactly the admin permission that deadlock prevents anyone from having. The fix is a one-off command-line tool that does the move outside the web app entirely, since a button inside the app would run into the same locked door. It reassigns the login onto the record that already holds the history rather than rewriting any of that history, checked by folding the same event log before and after two merges and confirming it came out byte-for-byte identical both times.

Alongside it, two smaller corrections: the number of reasons that tool can refuse a merge had been written down wrong in three different places (seven, then eight, then the real count, nine), and the project's own README (the one set of instructions a human actually has to copy-paste) predicted a database name that Clerk never actually produced. It now reads that name off the running app instead of guessing it.

This blog you're reading used to regenerate itself up to 119 times in a single day, because two separate triggers both set it off: one after every code commit anywhere on the machine, another after every finished background task. That habit alone was about a third of everything the whole background-worker system spent over two weeks, roughly $1,188. It now generates once a day.

Separately, the safety check meant to stop these AI workers from overspending had quietly stopped working, in three ways, all fixed today. The spending limit it compared against was set to zero, so there was never a real number to cross. It's now set to $450, calculated from two weeks of real usage. The check itself was also running so often it kept tripping its own rate limit, and every time that happened it lost the ability to see a percentage at all. It now checks less often, on purpose. And the check that runs right before starting any new job had been treating “I don’t know” as “we’re fine,” which is exactly how a two-hour stretch on 2026-08-24, when a login token had quietly expired, went completely unnoticed. It now insists on an actual clean reading before letting anything through.

Individual background sessions also got a hard cap on how long they're allowed to run, after one had quietly run 173 steps and cost $20 without anyone deciding that was fine. Building that cap surfaced one more bug: a session killed for hitting it was being misread as “try again,” which would have sent it right back into the same wall a second time. Caught before it happened for real.

A separate side project of mine, reckreation, had a locked-door problem: two built-in admin accounts had no way to turn into normal logins, because doing that required a permission that nobody would actually hold until after they'd already done it. A one-off tool now does that move from outside the app, checked by confirming a person's history reads back exactly the same, byte for byte, before and after.