Skip to main content
After a session starts, the agent moves the work forward on its own, and you can add instructions or change direction at any time. When it needs a decision from you, the agent pauses and shows a request above the input box. It continues automatically after you respond.

Add messages while the agent works

The input box stays available while the agent works. There are two ways to send a message:
  • Queue: The message waits above the input box and is sent in order after the agent finishes its current task.
  • Steer: The message goes to the working agent immediately, and it continues with the new instructions.
The default is queue. When the input box has content, a Queue / Steer toggle appears next to the send button. Pressing ⌘Enter (Ctrl+Enter on Windows) sends with the opposite of the default. To change the default, go to Settings › General and change Messages sent mid-turn.
Steer is currently supported only by Claude Code and Codex. With other agents, messages are always queued.
You can drag queued messages to reorder them, click Steer to send one immediately, delete one to withdraw it, or edit it from the … menu.
Queue and steer

Queued messages appear above the input box and can be switched to steer at any time

Approve or deny the agent’s actions

When permissions are set to Ask for approval, or when Approve for me detects a potentially unsafe action, the agent asks for your approval before acting. The command it wants to run appears above the input box.
  • Allow: Approves this one time only.
  • Allow for this task: Approves this time and automatically allows similar actions for the rest of this session.
  • Deny: Doesn’t run it. The agent tries another approach.
  • Something else…: Write your instructions and press Enter. This denies the action and tells the agent what to do instead.
When focus isn’t in the input box, press 1 to allow, 2 to allow for this task, or 3 to deny. Changing permissions to full access midway doesn’t automatically approve requests that are already waiting. For how to choose permissions, see Start a session.
Approve a command

The agent asks for approval before installing dependencies from the internet

When there are several requests, use Previous request and Next request to switch between them. The agent continues only after you handle all of them. Press Esc to hide them for now, then click Needs your input above the input box to reopen them. When you’re not in this session, Needs input appears next to the session in the sidebar, and you also get a system notification. Press ⌥⌘A (Ctrl+Alt+A on Windows) to jump to the next session waiting for you.

Answer the agent’s questions

When the agent is missing information, it asks you directly. When it needs you to pick an approach, it offers options.
1

Answer

Type an answer or choose an option (you can press 1 to 9). When the options end with Other…, you can write your own.
2

Go through each question

When there are several questions, one appears at a time. Click Continue after answering, or click Previous to look back.
3

Submit

On the last question, click Submit, and the agent continues working. If you don’t want to answer, click Skip.
Answer a question

The agent asks you directly when it's missing information

Provide secrets

When the agent is missing a secret (for example SENTRY_AUTH_TOKEN), it explains why and asks you to enter it. The value you enter is used only in the current task and doesn’t appear in the conversation. For secrets you need long term, save them in the project’s secrets and environment variables. See Projects and environment setup.

Help manually on the cloud server

Some steps the agent can’t do for you, such as signing in to a website and entering a two-factor verification code. In that case, the request above the input box states what you need to do.
1

Open the screen

Click Take control to open the cloud server desktop and take over the keyboard and mouse, or click View cloud server to watch without interacting. When the session runs on your device, do it directly on that computer.
2

Complete the step

Follow the prompt to sign in, verify, or complete other steps. The agent doesn’t pause while you work.
3

Tell the agent to continue

Go back to the session and click Resolved. If the step is no longer needed, click Not now, and the agent tries another approach.
Mobile web can only view the cloud server desktop and can’t take control. For more about the remote desktop, see Cloud server.
Also, when running on your device, the agent may ask you to confirm before accessing sensitive locations such as SSH or the keychain. See Run on your device. After you connect Gmail, the agent always asks you to confirm before sending an email. See Connectors.

Stop the agent

While the agent works and the input box is empty, the send button becomes a stop button. Click it or press Esc to stop. Completed steps and file changes are kept, and you can send another message to continue. Requests still waiting for you are canceled, and queued messages are sent in their original order.

Edit a message and rerun

If you notice your instructions were wrong, you can edit a message you sent and have the agent redo the work from that message.
1

Start editing

Hover over your message and click Edit (pencil icon). When the input box is empty, press ↑ to edit your last message directly.
2

Edit and rerun

After editing, click Rerun or press Enter. Press Esc to discard. If the agent is working, it stops first and then reruns.
The original records after this message collapse and are marked as discarded. Editing doesn’t undo file changes that were already written. To undo them, tell the agent directly. The edit button appears only when the current agent supports rewinding.

Fork from a reply

Forking copies the session record up to a reply into a new session, where you can continue in a different direction. The original session stays unchanged. Hover over a completed reply and click the fork button below it. You can also choose Fork › Fork chat in same worktree from the … menu in the session’s top bar to fork from the latest completed reply. The new session opens automatically.
A forked session shares the same project directory and branch as the original. If both run at the same time, they overwrite each other’s changes.
  • The new session copies only the record. It doesn’t copy tasks in progress, pending requests, or queued messages.
  • Sessions running on your computer can’t be forked from the middle. Use Fork › New chat in this worktree from the … menu instead. It keeps the project, model, and permissions but doesn’t bring the conversation record.
For problems such as sending failures or the agent stopping midway, see FAQ and troubleshooting.