Stopping and steering a run

Applies to: Hosted session and Your runner.

An agent run is not a black box you wait out. While it is working you can send it guidance, or stop it. When it ends, the people who care are told once.

This page covers the live controls and the run-end notification. For what a run is and what its timeline shows, see Tasks and pull requests. For the GitHub side, see Agents and GitHub.

Where the controls are

On the run view at /{workspace}/working/{runID}, and in the run sheet that opens when you click a run on Home.

You see Stop and the message box when you are a member of the task's project, or when you are the person who delegated the task. Everyone else can watch the run but not steer it.

Stopping a run

Stop asks the run to end. It is a request, not a kill.

Hosted session. The agent gets a short time to commit what works and push it, then the run ends as stopped.

Your runner. The runner checks for the request on its own schedule, then ends the provider process and finishes the run as stopped. The default check is every 15 seconds, so expect a short delay. You can change the interval with limits.control_poll_seconds in the config of ithura-agent.

What a stop leaves behind:

  • The run ends with the status stopped.
  • What worked so far is committed as "WIP: stopped by a person" and pushed to a branch named after the task branch with -wip on the end. It is never pushed to the task branch. If there was nothing to keep, nothing is pushed. A hosted session gets a short time to do this; if it runs out, the run ends as stopped without a -wip branch.
  • The task goes idle. It is not picked up again until someone delegates it again, so a stop is a real stop rather than a pause.
  • A line is posted in the agent's chat thread, once.

Pressing Stop twice is safe. The first request wins and the thread line is still posted only once. Stopping a run that is already finished returns a conflict rather than pretending to work.

Steering a run with a message

Under the live step log there is a message box. What you send becomes guidance for the agent, and it is posted in the agent's chat thread so the rest of the team can see what was asked.

What happens next depends on the provider.

A provider that can be steered mid-run, Claude Code among them, is interrupted and carries on in the same session with your guidance folded in. The runner acknowledges the guidance once the adapter has taken it, and it is never offered again.

A provider that cannot has your guidance recorded in the timeline once and left unacknowledged, which starts a fresh run after the current one ends. So the guidance is never lost, it just arrives at the start of the next run instead of the middle of this one.

Either way the guidance is stored as a comment on the task, so the record of what was asked lives with the work rather than in a chat scrollback.

Cost after an interruption

Interrupting a provider mid-run costs something, and the provider does not always say how much.

When a Claude Code run is interrupted, the run keeps the tokens for the interrupted part and says plainly that the dollar cost of that part was not reported by the provider. The run does not quietly drop the tokens, and it does not invent a figure to fill the gap.

Being told when a run ends

When a run ends, one notification goes out, so nobody has to keep Home open to find out what happened.

A run end means any of these: it finished, it failed, it asked a question, it found nothing to change, or it was stopped.

Who hears about it

The person who delegated the task, and the task's assignees. In both cases they must be active members of the task's project, so a run cannot notify someone who has lost access to the work.

The agent's own token user is left out, unless that user is also the delegator. An agent does not need telling about its own run.

One notification, not three

A single run notifies a person once, and retries do not produce a second notification.

Before this existed, one moment could raise several rows: a delegation change and a chat question could both fire for the same run ending. The run-end notification is now the single owner of a run end, and the older paths stand down when it is in play. The chat thread post and its unread badge still appear, since those are the conversation rather than a notification.

One honest caveat. Events can arrive from more than one place, and their order is not guaranteed: the pull request worker can move a task to in review a moment before the run reports finishing. So if you were already told about that moment within the last two minutes, the run-end notification is skipped for you, and the check works in both directions. It is a two-minute window rather than a strict rule, which is the right trade for not being told the same thing twice.

Channels and defaults

ChannelDefault
In-app inboxAlways
Web pushOn for questions and failures. Off for finished, nothing-to-change and stopped
EmailOff

Push is on by default for exactly the two cases where a person is blocking something: the agent asked you a question, or the run failed. The rest can wait until you next look.

Your own push and email master switches still apply on top of these, so turning push off everywhere turns it off here too.

Change any of it in notification settings, under the Agent runs group.

Do not disturb and quiet hours

Both are respected, and they drop every channel for the notification rather than just the noisy ones. A run that ends inside your quiet hours is still on the run view and in Home when you come back to it.

Finding something in a long run

A run's steps have a search box, and the API takes the same search as ?q=. On a run with hundreds of steps it is faster than scrolling for the moment something changed.

Search never matches content that has been deleted, below.

If a step shows a secret

Sometimes a command echoes a token, or a test prints a key. The step is a record of what happened, so the secret is now sitting in the run.

Workspace admins can delete it, and so can the person who delegated the task, as long as they still have access to its project. Having delegated something once does not keep the power to edit its record after you leave. From a step's menu you can delete that one step, or delete every step in the run that contains an exact piece of text, which is what you want when the same token was echoed twenty times.

What deletion does:

  • The content goes, everywhere. Not hidden: gone from the steps, the run detail, the receipt and the API.
  • The step stays, marked "Deleted by" with who did it and when, so the run's history does not acquire a silent hole.
  • The run's totals do not change. Deleting a step does not rewrite what the run cost or how long it took.
  • Every deletion is recorded in the audit log.

It cannot be undone. There is no copy kept anywhere to restore from, which is the entire point: a secret you can recover is a secret you have not removed.

Delete by text matches the text you give it. The same secret can sit in a run in more than one shape, raw in one step and escaped inside a JSON blob in another, so search for the other shapes too rather than trusting one pass to have found everything.

Deleting the step is not the end of the job. Treat the credential as exposed and rotate it, because it existed in a log, and anyone who could read the run before you deleted it has already seen it.

Troubleshooting

Stop did nothing for a few seconds. Expected. The runner checks for a stop request every 15 seconds by default, so there is a gap between the request and the run ending.

I do not see Stop or the message box. They are shown to members of the task's project and to the person who delegated the task. Watching a run needs less access than steering one.

My guidance did not change anything in this run. Check the run's provider. Only providers that can be steered mid-run fold guidance into the run in progress; the others start a new run with it once the current run ends.

I was not told my run finished. Check that you are still an active member of the task's project, and whether you were told about the same moment by another notification in the two minutes before it ended.

I got nothing on my phone when a run finished normally. That is the default. Push is on for questions and failures only. Turn on the rest in notification settings under Agent runs.