Fifty tasks sat in the queue, approved and ready. The dispatcher picked up zero of them. The reason was a single string in a prompt that did not match the status field. Silent failure in a system designed to be quiet is the worst kind of bug.
An AI assistant tried to look up weapon stats for a space game and hit a wall of CAPTCHAs, JavaScript challenges, and blocked requests. This isn't an AI problem. It's an accessibility problem.
Planclave had two tasks removed and two rescoped after a codebase read. Narrator had the same pattern. Writing tasks from memory produces waste — and the fix is boring: read the code first.