Hitting a usage limit was logged as success, so 11 tickets got reassigned to meHitting a usage limit was logged as success, so 11 tickets got reassigned to me
August 21, 2026 · A usage-limit kill got logged as success in four places, falsely reassigning eleven tickets, while two unattended jobs' stale tool lists left email access open.Running out of usage credit got logged as a success, wrongly kicking eleven tickets back to me, and two unwatched jobs left email access open.
A usage-limit kill that four layers logged as success
The pipeline that runs background work here logs how a session ended, works out why it failed, sorts that failure into a class, and decides whether to retry. Every one of those four steps trusted the same field: the terminal event's subtype, treated as success whenever the session reached its own end. Across the 866 session logs from 2026-08-21, that field reads success on all but two of them. That includes the 58 sessions actually killed mid-turn between 06:00 and 08:10Z, when the account's usage window ran out and its overage was exhausted. A second field on the same event, is_error, was the one that actually separated a session that finished from one that got cut off. Nothing downstream was reading it.
Because the whole chain read success as success, eleven tickets got reassigned to me as though a person was blocking them, all inside eight minutes. One of those had already been closed with a comment blaming a coincidental restart, claiming the session had “reached its own terminal event,” and declining a retry on the theory the work might already be done. It hadn't. That session was cut off at turn 40 of 40, with 39 tool calls behind it and 370 lines across 21 files still sitting uncommitted. Matching on the wording in the log (“hit your session limit”) wouldn't have caught this cleanly either. Of 178 logs that mention it, 117 finished normally anyway, because a session can hit the limit, wait, and keep going on overage.
The fix adds the missing check. A usage-limit kill now gets its own retryable failure class, gated on the account's window having actually reset. A second safeguard that was supposed to catch a run of failures like this one didn't help either. It watches how much of the usage window is used, not whether the account is already over on overage, so it never tripped across those two hours. That one ticket is still being refused a retry today, but now for the real reason (its own changes were left uncommitted) instead of a false “it already succeeded.”
Two unwatched jobs that ignored their own list of what they couldn't do
Two jobs here run with nobody watching them as they go: a weekly chess digest, and a monthly pass over around 300 scraped web pages for a vintage-map project. Both were set up naming eleven tools they weren't allowed to use. Reading each session's actual startup line, instead of trusting that list, found 21 tools available rather than eleven excluded. Among them: one that runs shell commands directly, and another that hands the work to a fresh session carrying every tool that exists. The list had simply gone stale as new tools were added to the underlying CLI since it was written.
That kind of list only ever names built-in tools, so it never reached the Gmail, Calendar, and Drive connections running in the background regardless. Checked directly: a session carrying the vintage-map job's exact flags could see the tool that sends mail from my account. A scraped web page, pulled from a site nobody here controls, could in principle have reached it. Both jobs are rebuilt the other way around now, naming only what they're allowed to have instead of what they aren't, with a separate flag turning off those outside connections entirely. The one spawn in the fleet left deliberately unrestricted is the one that carries out a delegated ticket end to end. It has to edit a real repository and run its tests, so bounding its tools would mean bounding the work itself.
My background task system logs how each automated job ends, works out why it failed, and decides whether to retry it. All of that trusted one signal: whether a session reported “success” when it stopped. But a session can report success even when it was actually killed mid-task. That happened 58 times in one stretch, when my Claude account ran out of usage and every session running at the time got cut off. Everything downstream trusted the “success” label anyway, so eleven of my open tickets got kicked back to me as though a person needed to step in, all within eight minutes. One of them had already been closed, with a note blaming an unrelated restart and saying the work might already be done. It hadn't. That session had been cut off with real, unfinished changes still sitting on disk. I've added a check for the actual failure signal, not just the label, and given this kind of failure its own category that waits for the usage window to reset before trying again. A separate safeguard meant to catch a long run of failures like this one didn't catch it either. It watches usage, not whether the account is already over on extra capacity.
Separately, two jobs that run completely unattended, a weekly report and a monthly check of around 300 old web pages, were both set up with a rule meant to stop them from using certain tools. I checked what they could actually do instead of trusting the rule. It let through almost twice as many tools as intended, including ones that amount to running arbitrary commands. Worse, that same rule never touched a separate set of tools tied to my email, calendar, and files. One of these jobs reads scraped web pages nobody here controls. It could, in theory, have sent email from my account without anyone noticing. Both are fixed now. Instead of listing what's forbidden, they list only what's allowed, which for both is nothing, and the email, calendar, and file access is switched off outright.