Side quest: no lease, no power

The interface for asking this lab’s hardware to do anything was built to serve two different callers from day one: an automated agent today, and a person clicking through a web page later. Neither gets a shortcut the other lacks. Getting there meant deciding, item by item, which real capabilities of the hardware simply do not belong in the interface at all.

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 →

Two questions, checked separately

A caller’s credential says what kind of thing it is generally allowed to attempt. It says nothing about whether that caller may act on any particular piece of hardware right now. Those are kept as two separate checks: a credential that is generally allowed to send power commands still cannot send one to a machine it does not currently hold, and holding a machine grants no ability it wasn’t otherwise scoped for either. Neither check substitutes for the other, and both have to pass.

Every request either fully succeeds or fully fails

Asking to reserve several machines at once returns either every one of them or none, never a partial set. A caller that asked for four and silently received two would go on to plan around resources it does not actually have, and would only discover that the moment it tried to use the ones that were never granted. Failing the whole request outright is the version of this that cannot be misread later.

A hold can be extended, but only up to a wall it cannot move

Extending how long a piece of hardware stays reserved is allowed, and it is capped by a limit fixed at the moment the reservation was first made — a ceiling later extensions cannot push back. Nothing can hold a piece of shared hardware forever simply by asking for a little more time in a loop.

What the interface refuses to expose at all

The sharpest decision here was not what to add, but what a real, working capability of the hardware to leave entirely out of reach of every caller, including the most trusted one.

Turning off the whole shared chassis is a genuine capability the underlying hardware has, and it is nowhere in the interface. Powering down the chassis takes every individual machine’s own management connection down with it — measured, not assumed. A caller allowed to switch the chassis off could switch off its own only way back in, with no path left to switch it back on again. An interface has to assume that mistake will eventually happen to someone, and the only fix that actually holds is never offering the button in the first place.

The same reasoning rules out low-level hardware controls generally — firmware settings, boot order, the raw commands the management controller itself understands. None of it reaches any caller. The service in front of the hardware keeps those entirely to itself, regardless of what credential is asking.

What is refused, is refused by name

A short list of specific machines can never be targeted by anyone, no matter what credential is presented — hardware that plainly does not belong to this lab at all, sitting close enough on the same network to otherwise be reachable by an ordinary command. That refusal is written as an explicit list of exactly which machines, rather than as a general rule about which network segment is off-limits.

The difference matters more than it looks. A rule built from a list of what is explicitly forbidden fails safely: whatever isn’t on the list yet is simply not reachable until it is added. A rule that instead tries to describe everything that should be allowed fails the other way the moment reality includes something nobody thought to describe — an omission there does not block the unexpected case, it quietly permits it.

What a caller sees when it asks for something it cannot have

Every refusal carries a short, stable, machine-readable reason a caller’s own code can check directly — “no active hold for this machine,” “a job is already running against it,” “this target does not exist.” The human-readable explanation alongside it is free to change its wording at any time; software is expected to match on the stable reason instead, never on the sentence next to it.

What generalizes

Size an interface to what is safe for anyone holding a valid credential to do, not to everything the underlying hardware happens to support. Some genuine capabilities are correctly left permanently unreachable through the front door, because the same interface that would let a trusted caller use them well would let any caller use them badly, and there is no credential scoped narrowly enough to fix that on its own.

A list of what is explicitly forbidden catches what a list of what is supposedly allowed cannot: the case nobody thought to write down. An allowlist is only as good as the imagination that built it. A denylist of specific, named exceptions fails toward safety when something shows up that neither list’s author anticipated.