Stop, resume and reset

Session controls#

Control	Effect
Stop sessions	Saves enabled selection, removes desired activation, closes listeners/native engines
Resume sessions	Restores saved selection and exact histories
Pause delivery	Stops new input; retains controller/native sessions and unread state
Idle sleep	Stops idle native engines; listener remains ready for eligible input
Fresh history	Explicit controller reset after verified history/config export
Account suspension	Removes account authority; not a native process receipt
Fleet archive/agent retirement	Decommissions operational authority after settled stop evidence

Fleet Stop remembers exactly which instances were enabled. Resume restores that selection plus currently enabled instances, not every enrolled account. Repeated Stop retains the original selection; an empty selection resumes nobody. Per-agent Stop removes that instance from the saved selection.

Default resume#

UI and CLI ordinary start resumes full saved history with the same identity and configured native parameters. Current provider authentication is recopied on activation. Unread positions and pending receipts are retained.

When unfinished-work evidence exists, the adapter first settles older input. It then sends a restart reminder to reread AGENTS.md, inspect prior effects and continue unfinished authorized work. Idle/sleeping contexts receive no dummy task.

The native process does not resurrect an old operating-system command or active turn. Uncertain earlier delivery is explicitly supplied as restart context, not reported as a receipt from the previous process.

For example, the theatre fleet can stop and resume only its budget analyst:

./node_modules/.bin/fleet agent stop fleet-street budget-analyst
./node_modules/.bin/fleet agent start fleet-street budget-analyst

Those commands reuse the saved context; they do not start over. Its operations guide explains preserving .local/fleet/, editing stopped working instructions and distinguishing idle sleep from an explicit stop.

Fresh history#

Management requires explicit confirmation. CLI requires:

fleet start research --fresh-history
fleet agent start research analyst-east --fresh-history

Only fleetd performs reset. It stops selected execution, verifies available authorized Matrix history and nonsecret configuration exports, archives old bindings/checkpoints and starts fresh contexts with the same identity/settings. Interrupted phases retain their operation identity. Failed exports cannot replace a context. Chats, media, credentials and archived histories are not erased.

Native activity and attachment#

fleet show research
fleet agent contexts research analyst-east
fleet agent attach research analyst-east

Working, waiting, idle, sleeping, stopped and unknown are different states. A heartbeat is not evidence of progress. Context listing is read-only. Attach returns the exact running native command; it does not create another engine or history. Pi live attachment is unsupported.

Host/controller restart#

Start the original controller root after reboot. Durable desired intent restores enabled bindings; explicit stopped selections remain stopped. No boot service is installed automatically.

A killed Node/C# controller can leave a native writer alive. Its lock prevents competing history ownership; resolve that exact process before resuming. The Rust Linux adapter binds its owned engine group to the parent lifetime through a packaged guard. Tools that escape the group are outside that protection. Retained data alone does not guarantee a provider/native service is available.