Fleet docs v0.2.0-rc.448

Admin

The admin page, at /admin, is where an admin changes the fleet by hand, without a chat or an agent. Admin in the top bar opens it. Only an admin sees the link and the page: anyone else gets 403 Forbidden, and so does every call the page makes.

The page has six tabs: Accounts, Secrets, People, Scheduler, Config and Change log.

Accounts

One card per inference account of yours: name, kind, user, a default mark, status (ok, limited with its reason, or paused), the 5-hour and weekly windows, this month's spend and the vendor's balance (what the usage popover in the top bar shows). Pause stops new turns on the account until Resume; turns running now finish. Set default, Rotate (a new value) and Remove (type the name to confirm) use /v1/profiles. Add an account takes an API key or a long-lived token: the value is typed once into a password field and is never shown again. The Pacing box edits pace.five_hour, pace.turn_cost, pace.max_concurrent and usage_threshold, shows the file value beside the database one, and resets. Each change is in the Change log. The page lists the accounts the API lets the signed-in admin manage: their own, and the box's own login when no user owns it.

Sign in (#616) adds a subscription account with no terminal. Pick the vendor (Claude, or ChatGPT for Codex), a name and what it is for, and press Sign in. Fleet runs the vendor's login on a node that has its CLI and shows a card with a Sign in to ... link. Open it and approve in your own browser, on your phone or laptop:

  • Claude shows a code after you approve. Paste it into the card's box and press Send code.
  • ChatGPT (Codex) shows a one-time code on the card. Enter it on the vendor's page; fleet sees the approval by itself.

Fleet then proves the login with one real call (Claude: a one-word Haiku turn) and stores it as the account. If the proof fails, nothing is stored, the card says why, and Try again starts it once more. A name you already have for the same vendor is renewed. An agent can ask for a sign-in with the login_request tool: the card names the agent, the inbox bell (and Telegram, for approvals) links you here, and the agent goes on when the account is stored. The link and code are shown only on this card, only to you, and are kept in the server's memory: never in a log, the ledger, the inbox or Telegram. From a terminal: fleet profile signin claude <name> --reason TEXT, fleet profile signins (the link), fleet profile signin-code <id> (the code on stdin).

Live settings

The Scheduler tab shows the live settings. serve.yaml gives each one its starting value. A value an admin saves here is kept in the database and wins over the file. It applies at the next decision of the scheduler, the node claims and the dispatcher: there is no restart. Reset deletes the database value, and the serve.yaml value applies again. The Source column says which one is in force.

Setting What it does
pace.five_hour The pace budget, a percent of each profile's 5-hour window. 0: no budget.
pace.turn_cost Percent of the window one turn in flight is expected to use. 0: 5.
pace.max_concurrent Turns of one profile at once, for a profile with no cap of its own. 0: no cap.
usage_threshold A profile at or above this percent of a window starts no new turn. 0: off.
dispatch.max_workers The hard ceiling on dispatched workers at once; node load decides below it.
dispatch.ceiling_usd_per_hour The dispatcher's machine spend ceiling, when max_workers is 0.
dispatch.model, dispatch.contract_model The models of dispatched workers: a list grouped by provider (Claude, ChatGPT, GLM). It is the model list of GET /v1/models, and fleet serve adds new models to it every 6 hours from models.dev. A model of another provider runs on its own CLI with a profile of it; with no such profile, the dispatcher says so and starts no worker.
dispatch.projects The projects the dispatcher serves. An empty list turns it off.

A value out of range is refused with the reason. A change to a pace setting writes a line to the scheduler log (fleet autoscale log), and each turn the pace then holds writes a held: line that names the budget in force. Listen addresses, TLS, sign-in and paths stay in serve.yaml only.

Scheduler tab and emergency controls

The Scheduler tab pauses and resumes the dispatcher, shows what serve.yaml sets for the scheduler read-only (local kinds, memory admission, spawn limits, live rules), and holds the emergency controls. Each one asks you to type its word before the button works, and each leaves an event in the Change log:

Control Word What it does
Pause every account PAUSE ACCOUNTS Every profile starts no new turn until resumed (profile.limited).
Pause the dispatcher PAUSE DISPATCHER No new ticket gets a worker (dispatch.state). Running workers keep running.
Stop every running agent STOP ALL Stops every agent that has not ended, live chats included (agent.stop each, one admin.emergency with the ids). Conversations and worktrees stay. Pause the dispatcher first, or it starts the next ticket's worker.
Prune ended agents PRUNE Removes every user's done, failed and stopped agents (admin.emergency, action prune). Their events and costs stay.

Change log

The Change log tab lists the admin changes in the ledger, newest first: settings (setting.change, with the value before and after), the dispatcher's pause switch (dispatch.state), the emergency actions (admin.emergency), and the profile.*, member.*, token.* and secret.* events. Pick a kind to filter it; Older reads the next 50.

For scripts

GET /v1/settings, PUT /v1/settings/{key} with {"value": ...}, DELETE /v1/settings/{key} and GET /v1/admin/log?kind=&before=&limit=, POST /v1/admin/stop-all (answers the ids it stopped) and GET /v1/admin/serve. All need an admin's token; a change needs a write token.