A Windows guest on the chassis gets a freshly minted, single-use Administrator password rather than one a caller supplied and the system never saw again. That password reaches the person who actually needs it as a one-time link, mailed to the lease holder, burned the moment it’s opened. The mechanism was built three weeks before it delivered a single link that actually worked.
Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →
Why minting the password mattered in the first place
Before this existed, a caller generated the Windows guest’s password, passed it in, and the builder discarded it once the job finished. That produced two running Windows guests on the lab network whose only credential existed nowhere at all — not logged, not stored, not recoverable by the person who was supposed to be using the machine. The build reported success. Minting the password server-side and handing it over deliberately, once, to the one person entitled to it, closes that gap — and it’s the same trade already made for the blades’ own host accounts, extended to the guests running on them.
What “built” turned out not to mean
The delivery link comes from a third-party one-time-secret service. The first attempts to use it — 2026-09-05 — both failed silently, logged simply as a generic delivery failure with no further detail. Nobody went back to find out why, because nothing else about the mechanism looked broken.
Three weeks later, a routine check of the account credentials turned up the actual cause, and it was two mistakes layered on top of each other. The service’s original, unauthenticated API had been retired entirely — every call to it now fails, regardless of whether credentials are supplied, which looked identical to whatever had gone wrong back in September. And the earlier assumption about why it was failing was also wrong: the working theory had been that the service stopped allowing anonymous, credential-free use, which is not true of its current version. Two different explanations for one failure, and the real one was neither.
Two smaller bugs on the way to fixing it
Getting the new account credentials onto the host surfaced a second, unrelated defect: they were piped from a Windows workstation to the Linux host over a plain text stream, which appends a carriage return Linux doesn’t expect. Both values landed one character too long, and every authentication attempt failed with a generic invalid-credentials error that gave no hint the values had been corrupted in transit. The fix was mechanical — strip the extra character — but finding it came from comparing the credential’s length against what it should have been, not from the error message, which said nothing useful at all.
Confirming the mail path worked hit a similar problem: a manual reachability check waited for a fixed number of bytes from the mail server and timed out, because the server’s own greeting was shorter than the check expected — correctly working, misread. A check that waits for a byte count rather than a line will time out against a server that is behaving exactly as intended. A real client, rather than a manual probe, sent mail successfully in well under a second.
What it looked like once every piece was actually correct
With the account ported to the service’s current API, the mail path confirmed working, and the recipient address actually set on the tenant record, the next Windows guest build completed and its access link was delivered, opened, and used. The three records that mattered — the link was generated, the notification was marked sent, and its payload was cleared afterward — all came from the system’s own database, not from a report the build job made about itself.
What generalizes
A mechanism that silently fails at the very start of its life can sit that way indefinitely, because nothing about a quiet failure looks different from a mechanism nobody has needed yet. The two 2026-09-05 failures read, at the time, as “not yet exercised.” They were actually “already broken,” and the difference was invisible until someone went looking specifically for why. The gap between a system that has never been used and a system that has already failed every time it ran looks identical from the outside, and only one of them is fine to leave alone.