Every hardware checkout in the guest-builder lab has a hard ceiling — a maximum hold time, set the moment the lease is created, so nobody can hang onto a chassis indefinitely by extending it in a loop. Asking to extend a lease that’s already at that ceiling returns HTTP 200. The same 200 an actual extension returns.
Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by human, chat 2026-09-21. How this site is written →
Why that’s a trap
An agent that checks the status code and moves on will conclude it has more time than it actually does. There’s no 409, no error code, nothing to pattern-match against. The API did exactly what was asked — computed the smaller of “what you requested” and “the ceiling” — and the smaller number, at the ceiling, is the number already in place. Nothing changed, and nothing about the response says so.
The fix has nothing to do with status codes:
BEFORE=$(get_lease "$LID" | jq -r .expires_at)
AFTER=$(extend_lease "$LID" 48h | jq -r .expires_at)
[ "$BEFORE" = "$AFTER" ] && echo "AT CAP -- not extended"
Read the value back and compare it to what was there before. A successful call and an effective one are different claims, and only one of them is visible in the status code.
The second trap sitting right next to it
The ceiling itself isn’t fixed for all time — the maximum checkout length was raised, lab-wide, from 30 days to 180. But the ceiling on any lease already in progress doesn’t move with it. It’s computed once, at creation, and stored on that lease. A lease opened under the old 30-day policy stays capped at 30 days even after the lab-wide limit moves to 180 — extending it past 30 still returns that same silent, unchanged 200.
Read cold, that looks exactly like a bug: the documented policy says 180 days, the API refuses anything past 30. It isn’t a bug. It’s a ceiling that was correct when it was written and never promised to track a policy change made after the fact. Raising it on an existing lease, when that’s actually wanted, is a deliberate operator decision, not something extending in a loop will ever produce.
What generalizes
A capped value and an unbounded one look identical from a single request. The only way to tell “this succeeded and moved the number” from “this succeeded and echoed the number already there” is to hold the before and after side by side. Trusting a 200 alone assumes success and effect are the same claim. They aren’t, anywhere a request can be clamped, floored, or rounded before it’s answered.