Skip to content

Act 28

Guidance that never fires on its own

A blocked firewall change turns into two expensive research detours for a problem that was already solved and already written down, and the fix that finally worked wasn't a better document.

#guardrails#automation#incident-review#unifi#claude-code#dockhand#obsidian

Partway through a migration, several new monitoring targets on one host were failing. The fix looked small: two network-firewall rules needed to exist and didn’t. Joe approved creating them in conversation. The tool call that would have actually created them was refused anyway, by a separate layer entirely: a permission gate on Skippy’s own side, independent of Joe’s approval, sitting there specifically to catch this exact kind of action before it reached the live system.

A written answer for this network controller already existed. A project document, explicitly flagged as required reading before touching that controller’s programmatic interface at all, covered its authentication method, its non-obvious required fields, and the general method of finding a rule that already worked and copying its shape. The document was there. Skippy didn’t read it.

Instead, Skippy reached for a strategy that had just worked on a different, unrelated problem minutes earlier, without first checking whether a more specific answer already existed for this particular system: a research pass that read a large amount of unrelated public source code, hunting for how the controller’s interface was supposed to work. It burned a large share of the session’s resources on its own, and Joe called it out directly and unfavorably before shutting it down.

Four wrong guesses before the fifth worked

Section titled “Four wrong guesses before the fifth worked”

What followed was four separate live attempts against the real, production controller, each a different wrong guess, each one failing before the next was tried. One request was missing a required security element. The next was missing required data outright. The one after that had the right data, but in the wrong naming convention. The last chased a theory that the authentication method itself was somehow broken, and it wasn’t. Only the fifth attempt went through clean. Across this whole stretch, Joe told Skippy three separate times to stop being broad and expensive about it.

The costlier mistake wasn’t any one of those four guesses. Partway through, Joe said plainly that this exact kind of task had been done before, and said roughly how. That statement should have sent the search straight to the estate’s own notes, where the precedent was written up in detail. Instead Skippy heard it as a hint toward trying yet another guess against the live system, and answered it by launching a second research pass, the same kind of detour as the first, rather than searching the one internal store built exactly for retrieving exactly this sort of thing.

One new fact, against everything written down twice

Section titled “One new fact, against everything written down twice”

By the end of it, exactly one thing about that night was genuinely new: the controller’s newer interface wanted one particular field written in a different casing convention than assumed, a real detail nobody had written down before. Everything else touched that night was already written down, in at least two separate places, before the night started: the authentication method, the security-token requirement, the shape of the required fields, the general method of copying a working rule’s structure, and the plain fact that this whole class of task had already been done before.

Looking past what happened to why it happened turned up something sharper than a reminder to check the docs next time. A session that has already filed a problem as unknown doesn’t spontaneously generate the thought that it might already be solved. This wasn’t a case of Skippy knowingly skipping a document. From inside that moment, no document was believed to exist at all. Written guidance, however correct, depends entirely on someone first deciding to go look for it, and that decision is exactly the step that doesn’t happen once a problem has already been filed away as new.

The proof that a fourth document wouldn’t fix this arrived within a day. A separate tool, already flagged in writing as one that leaks a specific kind of credential, leaked that exact credential again, the second time in a matter of days, from a session that had correctly steered around a different risky tool only an hour earlier. The shape was the same: correct guidance existed and simply didn’t fire.

So the actual fix wasn’t more writing. It was a gate placed directly at the point where the risky action would be taken, refusing it automatically and naming the safe alternative right there, in the moment. No lookup required, no need to recall that a better path existed at all.

Building it correctly took its own care. It had to key on the general shape of what made a tool dangerous, not a list of specific names, so a new but similarly-shaped tool would get caught automatically instead of waiting to be added by hand later. An early version had a real gap: matching on a plain substring of a tool’s name let a differently-pluralized variant of an already-covered case slip straight through. And the gate was proven before it was trusted, deliberately, with a harmless target: a real block confirmed, then reverted, then confirmed still open, rather than ever being tested by actually trying the unsafe action it existed to stop.

The same gate’s own limits showed up hours later, on a completely different problem. A message meant for another session appeared to fail because a lookup reported no reachable session at all. Written guidance already covered this exact situation too, and it didn’t fire either. The transport being used had been working right up until the moment it silently stopped, and there was no gate sitting at that particular point of action the way there now was for the tool just fixed. A structural gate only ever covers the specific point it’s built into. It raises the floor substantially. It doesn’t close every gap at once.

What this night says about guidance in general, and what it takes to make it actually fire, is written up on its own, at Guidance must arrive at the action.

"registry" not in "list_registries"
This is the exact gap the early substring match had, made literal — the word "registry" never occurs inside "list_registries", so the covered case and its pluralized twin looked unrelated to a check built to catch both.