A stuck worker: retry, hold, ask a person
A worker can stop without progress in four ways. It goes silent (watched, not stopped). It ends turns that do nothing.
It ends a turn with no done. It asks a question. fleet answers each one differently.
Silent: watched, never stopped
Nothing stops a turn for being silent, and no turn has a default time or budget limit (owner
ruling 2026-11-05, #230). The node watches the output of the turn stream and the CPU of every
process of the turn, and the agent has one of three states, which the page and fleet ls show:
- active: the turn has output.
- busy: no output, but the tools of the turn (the processes below the model CLI, whose own idle CPU is ignored) use CPU, such as a long test run.
- silent: no output and no CPU for
--silent-after(20 min). The node writes oneSILENT: ...status report. The turn keeps running. When it moves again one more report says so (fleet/internal/node/hang.go).
A process that dies is a crash, not silence: the supervisor tries again after 30 s, 2 minutes and
8 minutes. A time limit exists only when a person asks for one with --time-cap-s.
Each done, failed, blocked, needs_input and silent is pushed to the one who started the
worker: a parent agent gets a note from fleet in its inbox, which wakes it. A worker started by
a Claude session (CLAUDE_CODE_SESSION_ID) goes to the webhook in push_webhook_file of the
serve config, as JSON. A worker started by a person goes to their Telegram chat
(fleet/internal/push/). A chief subscribes with fleet watch <url>, which sets that webhook
on the server (an admin only; it is kept in the server's dir as push-webhook and wins over the config file; fleet watch --off stops it).
Turns that do nothing: the hold
A turn ends with no done
A question to a person
| End of a turn | State | What wakes it |
|---|---|---|
fleet status --blocked in the turn |
blocked |
A message, or resume |
| An ask in the turn | needs_input |
The answer |
| Tool denials | needs_input "denied: ..." |
A message |
| Messages wait for it | working |
It starts again at once |
None of these, and no done |
needs_input |
A message |
- The model ladder is
DefaultLadderinsupervisor.go. An agent with a model that is not on the ladder goes toblockedafter the second stuck retry. - Code:
fleet/internal/supervisor/supervisor.go,watchdog.go,guard.go,fleet/internal/api/lifecycle.go,fleet/internal/notify/. - If the owner has no linked Telegram chat, the question waits in
fleet asks. - An ask to
parentor to an agent name goes through the mailbox, not Telegram. - A plain Telegram message with no reply goes to the only agent that waits. If more than one agent waits, the bot tells you to reply to the question.
- Not built: a limit on the number of turns of one agent.
- The page shows each case in the workers panel and the supervisor log.