Monitor, steer and resume runs.

Keep the run ID returned when Crew accepts work. A durable run is independent of the interactive conversation, so reconnect to that run before controlling it. This guide covers CLI 0.9.0.

#Find the intended run

Inside Code:

/run runs
/run use <run-id>
/run status
/run progress
/run watch

use selects a run for this workspace. watch follows bounded progress until a terminal or held state, its time limit, or your exit key. Leaving watch does not cancel the run. Use /run result and /run evidence to inspect saved work and verification, including an incomplete attempt.

/run progress --cursor N requests progress after a returned cursor. Status, progress, result, and evidence also support --json. Read current help for the placement of an explicit run ID rather than assuming every verb has the same arguments.

#Choose the control that matches your intent

Running work can be paused at an activity boundary and resumed under its existing authority. Interrupt holds a run; a bounded continuation may be needed to move held work forward. Cancel permanently ends that run. Steering supplies bounded instructions for a later activity boundary. Watching only observes. The slash surface provides pause, resume, and cancel; steer, interrupt, and continue are available through approved model tools or the direct CLI.

The example run R1 has a finished inspection, current change activity and later verification. Pause holds the next activity at a boundary; resume continues R1; cancel ends its continuation. The watcher follows a separate observation line.

Intent	Control and effect
Observe without changing work	Status, progress, watch, result, or evidence.
Hold at a safe activity boundary	Pause; it may not be immediate.
Continue paused work with the same authority	Resume.
Add bounded direction	Steer for the next applicable boundary.
Hold an attempt for intervention	Interrupt, then inspect its recovery guidance.
End this run permanently	Cancel; preserve any useful work before deciding.
Add bounded time or authority to eligible held work	Continue with explicit grants, using the run's returned recovery guidance.

#Use the right control surface

The in-session mutating commands are /run pause, /run resume, and /run cancel. They read fresh status first; cancel asks you to type the run ID. There is no /run steer, /run interrupt, or /run continue shortcut. Ask Code for those actions through its Crew tools, or use the direct CLI.

For direct control, first read the current state:

kaitoi crew status <run-id> --repo /path/to/project

Mutating direct commands require the current --expected-state-version N. Use the value returned by status; if another actor changes the run, refresh status and reassess instead of reusing a stale or guessed version. Steer also takes an instruction. The exact continuation command is returned in recovery guidance because eligible grants depend on why this run stopped.

#Resume is not a new budget

Resume releases paused work within its existing authority. Continue can add explicit bounded grants for an eligible held or budget-limited attempt. Time, dispatches, tokens, costs, preparation, and network authority are distinct choices; do not blindly add every available grant.

A run may have a configured automatic-continuation allowance. Reaching a time boundary therefore does not always mean it is waiting for you. Inspect the current state and remaining bounds. Operator /run actions occur outside a model turn; the model learns the change when it next reads appropriate status.