Side quest: a button that would have lied

Two machines that are not blade servers were about to join this lab, and the web page built to show and control its hardware had quietly assumed, in dozens of places, that everything it would ever display lives in a numbered slot inside one shared chassis. Neither new machine does. Finding out how deep that assumption ran, before building around it further, turned into its own small project.

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 →

How deep the assumption actually went, measured rather than guessed

Rather than guess at how entangled the chassis-shaped assumption was, the actual code was searched for it. One specific word describing a slot inside the chassis turned up dozens of times, spread across the interface’s data model, its hardware driver, its background worker, and the page template itself. Every one of those was a place that would need to already agree on what a machine, other than a chassis blade, even looks like before it could display one correctly.

Three separate ideas where there had been one

The fix was to stop letting one idea carry three different jobs at once, and give each its own name instead.

A leasable, powerable unit is its own idea, independent of whether it happens to live in a shared chassis or sit on a shelf by itself. A shared enclosure is a second, separate idea — something with its own state, capable of taking every machine inside it down at once if it loses power itself, which a standalone machine simply has no equivalent of. And what a given power connection can actually do is a third idea, declared explicitly rather than assumed: some can gracefully ask an operating system to shut down first, some can only cut power outright, and some support nothing controllable at all beyond simply showing that a machine is there.

That last case — a machine the interface can see but not actually switch on or off itself — was not treated as an edge case to work around later. It is a normal, supported state from the start, because the alternative was a system that could only either lie about a capability that isn’t there or make the machine invisible entirely rather than admit it is present but out of its reach.

What this changes about the page itself

Once what a given connection can actually do is declared rather than assumed, the interface can be honest about it before anyone clicks anything. A control for an action a particular machine’s connection cannot perform is not shown at all, rather than shown and left to fail after being pressed. A button that only reveals it does not work once someone has already clicked it has told them nothing they could have used a moment earlier, when it would have actually mattered.

The same discipline that already existed, applied again

One rule already in place elsewhere in this system got reapplied here rather than reinvented: nothing about who is allowed to touch a piece of hardware may ever be decided by what kind of hardware it happens to be. The specific list of machines nobody may ever touch stays exactly that — an explicit list of which machines, never a general rule keyed off a category. A rule that instead says “hardware of this kind may always be reached” quietly stops applying the moment a new kind of hardware shows up that nobody had in mind when the rule was written, and by then it has already been treated as safe by default.

Off because it was told to be, and off because it broke, are not the same fact

A machine that was deliberately powered down on purpose must not be able to trigger the same alert as one that unexpectedly went silent. This lab’s own hardware already made the need for that distinction concrete before the new machines even arrived: most of its own equipment sits deliberately idle at any given moment, and treating “not currently running” as inherently alarming would have meant a constant, correctly ignored alarm — which is exactly the condition under which a real failure later gets missed inside the noise. The fix was to record why a machine is off as a fact of its own, sitting next to whether it is off, rather than trying to infer intent after the fact from a state that looks identical either way.

What generalizes

A capability should be declared and checked, never inferred from what category of thing a device happens to be. “This kind of hardware supports this action” is a fact that holds only until the one exception that breaks it shows up — which, on real hardware, eventually happens. Ask what a specific instance can actually do, not what its category usually implies.

The right moment to tell someone their choice will not work is before they make it, not after they have already committed to it. A control that fails only once pressed has extracted an action from someone before giving them the information that would have changed their mind.