Conversation and tasks

Steering ongoing work

Normal submission is steering: while the agent is working, another message is queued for the next model step. It does not abort the response or tool batch already in progress. The agent sees it before continuing with another request.

For example:

Keep the existing public API. Also add a regression test for Windows-style paths.

You can also ask a side question this way. It becomes part of the main conversation, so the agent can remember the answer and act on any resulting instructions.

The agent library additionally has an append/follow-up queue: each submission waits until current work and pending steering finish, then runs separately. Messages within one submission stay together. The CLI's normal Enter submission uses steering; there is currently no separate documented CLI follow-up shortcut.

Stopping

Press Esc or run /stop to stop the current turn or tool batch. Stopping is an explicit interruption; sending an ordinary message is not. Inspect any changes already made before asking the agent to continue—stopping doesn't undo filesystem or external side effects.

Long-running tools

Slow tool calls can become managed background tasks instead of blocking the conversation. You will see task progress/completion notifications. The agent can inspect running/recent tasks, poll new output, and stop tasks using the built-in task_list, task_poll, and task_stop tools.

A background task is not the same as a separate conversation. Its eventual result can still be delivered to the agent. Stopping the agent's current turn and stopping a background task are distinct actions; ask explicitly to stop a particular task if that is what you need.

Running a command yourself

Prefix composer input with ! to execute a shell command and include its tool exchange in the conversation:

!git status --short

This runs locally; it is not merely a request for the model to suggest a command. Only execute commands you trust. Custom commands can also execute shell snippets while their prompt is being prepared.

Being explicit

Separate planning from execution when it matters: “inspect this, don't edit yet” is clearer than relying on the agent to infer your intent. Use a restrictive permission preset when you need actual approval gates; instructions alone are not a security boundary.