What Nightdesk will never do
Nightdesk watches your fleet and launches things. That is the whole of its remit, and most of what follows is a consequence of taking it seriously.
It never decides that something should happen
Section titled “It never decides that something should happen”There is no scheduler, no playbook, and nothing that decides on its own that a session should be started, stopped or retried. Nightdesk acts because you pressed something. The single exception is the stall nudger below: off until you arm it, and bound to one fixed word into a session the harness itself reported stalled.
That is the test to apply to anything it does: who decided it should happen? Restarting an agent on the newly installed CLI, for instance, is a button press and a confirmation, on the row in front of you — you decided, once. Choosing an approval policy for a Codex session is naming one of the harness’s own modes at launch, the way you would name a model. Neither is the app acting on its own initiative.
It never answers for you
Section titled “It never answers for you”Questions and permission prompts go to you. Nightdesk will not answer a question an agent asked, clear a permission prompt, or approve a tool call — not on a timer, not on a rule, not ever.
The one exception, deliberately made and deliberately narrow, is the stall nudger: off by default, and when you arm it, it re-sends one fixed word to a session the harness itself reported stalled on an API error. It answers nothing: not a question an agent asked, not a permission prompt, and never a session the harness has not reported stalled.
The other thing Nightdesk will put into a running session is a task handoff — text you composed and saw, sent once, at your press, into a session that is idle rather than waiting on you. It asks for work on a task; it never answers anything.
Its hooks never gate your agents
Section titled “Its hooks never gate your agents”The hooks Nightdesk installs report what your sessions are doing. None of them can block a tool call, hold up a turn, or change what an agent is permitted to do. If Nightdesk is not running, or is broken, your agents carry on exactly as they would have.
It never owns your rules
Section titled “It never owns your rules”Your rules live in CLAUDE.md and AGENTS.md, where the agents read them. Nightdesk does not
keep a copy, does not add to them, and does not enforce them. If you want an agent to behave
differently, you change the files it reads — not a setting in here.
It never replaces your configuration
Section titled “It never replaces your configuration”Everything it installs is additive and removable, and where a file belongs to a harness, Nightdesk asks that harness’s own tool to change it rather than editing the file. What Nightdesk writes to your machine is the full list, each item with the way to take it back out.
It never starts a session you did not ask for
Section titled “It never starts a session you did not ask for”Attaching to a session opens what is already there; if Nightdesk cannot find a running background session in the harness’s listing, it shows you the harness’s own picker rather than starting something. Attach never restarts.
And if launching on one of your profiles fails — the folder, the sign-in — nothing starts. It never quietly falls back to your default account, because a session on the wrong account looks exactly like a session on the right one.
There is no Nightdesk account and no Nightdesk server
Section titled “There is no Nightdesk account and no Nightdesk server”It is one app on your Mac, reading what your harness writes on disk. The only things it sends anywhere are the requests it makes to GitHub and to your providers on your behalf, using your own accounts.