Skip to content

fleet-routine — Learn More

Plain-language background on the ideas this project uses. For the technical version with links to the actual specs and APIs, see docs/technical/references.md.

"Run this as me, whether I'm logged in or not" (the S4U login trick)

Windows lets you schedule a task two very different ways:

  • "Only run when I'm logged in" — simple, but useless for a 2am job, since nobody's logged in at 2am.
  • "Run whether I'm logged in or not" — this is what fleet-routine uses. You type your password in once, during setup. Windows itself remembers it securely (this project's code never sees or stores it), and from then on, it can start a task under your identity even while you're asleep, and it still looks, to every program running inside it, like you logged in normally.

That second option is the one setting that decides whether a scheduled AI job can find your saved logins. Get it wrong (for example, run the job as a generic system account instead of your account) and every AI tool silently can't find its saved login — the exact failure Gate 0 exists to catch. Want the deeper technical version? See "Windows Task Scheduler — S4U logon" in docs/technical/references.md.

The one-time secret word (why it's random every time)

When Gate 0 checks that an AI tool is working, it doesn't just ask "are you there?" — a tool could answer that question even while badly broken (e.g., by printing an error message that happens to say something). Instead, it asks each tool to say back a random secret word that was invented seconds ago, fresh, for this one check only.

Why does that matter? Because it rules out two sneaky ways a check could lie to you:

  • An old, leftover result file can't contain today's brand-new random word — it was written before the word existed.
  • A generic error or help message ("please log in first") also can't contain today's word — it was written by the tool, not copied from a script.

So if the exact word comes back, that's real proof a real AI model actually ran, right now, not just that some program exited without crashing. This same idea — a fresh, unpredictable check value that can't be faked by replaying an old answer — shows up all over security engineering under names like "nonce" or (in stronger, cryptographic forms) "HMAC." See "HMAC / nonce canaries" in docs/technical/references.md for the technical reading.

"Fail closed, not silently" (why a missing tool still counts as a failure by default)

There's a design choice buried in Gate 0 worth calling out because it's easy to get backwards: if a tool is missing or broken, the default answer is fail the whole check, not quietly skip that tool and call everything else good enough.

There is an escape hatch — you can explicitly say "I know this one tool isn't installed on this computer, that's expected, don't fail the whole check over it." But that escape hatch only works for genuinely absent things (a program that isn't installed, a local model server that isn't running). It is specifically blocked from working for a broken login. You cannot use the escape hatch to wave away "this AI tool's saved password doesn't work" — that always fails the check, on purpose, because letting that slide would defeat the entire reason Gate 0 exists.

The scheduler that will (eventually) be a Claude-Code-style "routine," house-wide

Claude Code already has its own version of "run this automatically on a schedule" for a single tool. The plan here generalizes that same everyday idea — schedule something, walk away, trust it ran — across every AI tool the household uses, on every house computer, instead of one tool on one machine. That generalized version is what "cron" traditionally means in computing: a way of saying "run this job on this recurring schedule," going back decades before any of this AI tooling existed. See "cron / scheduling background" in docs/technical/references.md.

See also

  • docs/plain/OVERVIEW.md — what routines are and what Gate 0 proves, without the background reading.
  • docs/plain/making-of.md — how the project got here.
  • docs/technical/references.md — the same topics, with real links and API docs.