Code / Messages and steering

Messages and steering.

Type in the composer at the bottom of a session and press Enter to send. If the agent is idle, your message starts a turn. If it is working, your message steers the running turn. To end the turn instead, press Stop.

The session view

A running chat session titled "Show a delivery window on the orders table". The breadcrumb ends in the branch chip, chat-4kd9wp. On the left, the conversation, with each reply folding its steps into Explored and Wrote rows, and the composer beneath with the model, Opus 5.5 (xHigh), and Attach files, Dictate and Send buttons. On the right, the Files tab lists the session's QA Report and Review files, with the QA Report open beside the list.

The conversation is on the left. The panel on the right has these tabs.

TabWhat it shows
PreviewThe app running in the session's sandbox. See Preview and sandboxes.
DesktopThe sandbox's desktop.
FilesDocuments, mockups, images and videos the agent wrote. See Files of a session.

Any member of the workspace can send messages in any session. Each message also tells the agent which page the Preview is showing, so "this page" means what you are looking at. When the sandbox is paused, sending a message starts it again.

Send a message

  1. Type in the composer. Press Shift+Enter for a new line.
  2. To add files, press Attach files or drag them onto the composer. You can also paste an image.
  3. Press Enter or the send button.

Type / to open the menu of skills and commands. In a Codex session, type $ instead. See Skills.

Attachments

LimitValue
Files per message20
Size per file10 MB
ImagesJPEG, PNG, GIF, WebP
DocumentsPDF, Word (.docx), Excel (.xlsx, .xls)
Text.txt, .md, .html, .htm, .csv, .tsv, .json, .xml, .yaml, .yml, .log
VectorSVG (.svg), read by the agent as markup, so any model can take it

Some models read fewer kinds of file:

ModelImagesPDFs
Anthropic modelsYesYes
OpenAI modelsYesNo
GLM 5.3NoNo
GLM 5.3-FlashYesYes
Kimi modelsYesYes
DeepSeek V4.1 FlashYesNo

The composer refuses a file the model cannot read and says why. If you switch to such a model, those files stay in your draft and the send is held until you remove them or pick a model that can read them.

Steer the running turn

Steering adds your message to the turn that is already running. The agent changes course without starting over.

Send your message while the agent works. It goes into the running turn at once, whoever started the turn. A teammate's message steers it too. Under your message, "Steered, 12:04" shows when it went in.

If the turn ends before the agent reads your message, also after Stop, the message is sent at once as the next turn. It then shows "Sent as the next Turn, 12:04".

Some models cannot read messages mid-turn. While their turn runs, Send is disabled and the composer says to wait for the turn to end or stop it. Your draft stays.

Sync and the other buttons that send a message are unavailable while a turn runs.

The Deploy Chat and the Merge agent chat steer the same way. Up next is Sensei's only: in a Sensei thread, a message sent while Sensei works waits in Up next. See Sensei.

Stop a turn

Press Stop, the square button in the composer, or press Esc. The agent stops where it is. Send a message to carry on.

Answer a Clarification

Sometimes the agent asks a Clarification before it goes on. It appears in the transcript, one tab per question, and the turn ends. The task's card on the Tasks page shows "Your turn" to the session's Assignee.

  1. On each question tab, pick an option. To write your own answer, pick Other and type it.
  2. Open the Summary tab to check your answers. Add anything else under Extra information (optional).
  3. Press Submit answers. It appears once every question has an answer.

Your answers go to the agent as your next message, and it carries on. You can also type a reply in the composer instead.

Clear or compact the conversation

Two commands in the skills menu act on the agent's memory of the session. They are marked Runtime. Neither starts a turn. They are unavailable while a turn runs, and work again once it ends. In a Codex session, $clear and $compact work as well as /clear and /compact.

CommandWhat it doesAvailable in
/clearStarts a fresh conversation for the agent. The transcript marks it "New conversation from here".Claude, Codex, GLM, Kimi and DeepSeek
/compactSummarizes the conversation so far to free up context. Add text after it to say what the summary should keep.Claude, Codex, GLM, Kimi and DeepSeek

Codex compacts without a prompt, so in a Codex session the text after $compact stays in the conversation's history as a note, which Codex may not follow.

While a Runtime Command runs, the composer names what is happening, such as "Compacting the conversation". Send is disabled until the command finishes, and your draft stays. The composer offers no Stop during the command. One command runs at a time.

The messages stay on screen either way. After /clear, the agent remembers nothing said before it. The branch and the files in the sandbox are unchanged.

Watch the context

The ring beside Attach files shows how full the model's context window is. Hover over it to see the tokens used. It turns red at 90%. When it gets close to full, run /compact or /clear.

See what subagents are doing

When the agent starts helper agents, a row of their icons appears in the chat. Press it to open Agent activity. It lists the helpers as Active, Failed, Stopped or Done. Pick one to read its own conversation, and press Back to the main chat to return.

In a chat session, every helper is a Delegated Run. On Subscriptions or workspace Credentials, a helper that names neither a provider nor a provider to avoid runs on the turn's own model and effort. If that provider reached its limit or cannot take the helper, the helper does not start and the agent reads why. A helper that names a provider runs on it, on a model about as strong as the turn's own, or that provider's strongest. A helper that names a provider to avoid, such as a review, runs on another provider that has a model about as strong as the turn's own, and stays on the turn's own model when none has. On Subscriptions or workspace Credentials, the row names the helper and its model, such as "reviewer (Opus 5.5, xHigh)", and the chat header names the model while you read it. On Free, helpers always use their parent Turn’s model and Source, even when the agent names another provider or model. Their model, Source and provider stay hidden in the chat, run lists and results. If the Source reaches its limit, is paused or is removed, Sensei moves the work to another Source running the same model, or stops it when none is available. The agent cannot start subagents of its own here. Crons and the deploy agent still can.

The agent can steer a helper that is still running: its message reaches the helper at its next step, or joins its prompt when the helper is still queued, and it shows in the helper's conversation as a new message. When a helper has finished, the agent can continue it. That starts a new run that resumes the helper's session on the same model, so it keeps what it read and did, and it shows in the same row with its prompt as a new message. A helper can be continued while its session is still on the sandbox, also in a later turn, and one run at a time. When it cannot, for example after the sandbox was replaced, the agent starts a new helper and gives it the earlier output.

A helper has no time limit while it works. When a helper shows no activity for 25 minutes, its row says how long it has been quiet and what it did last, and the agent decides to steer it, stop it or wait. A long build or test can be quiet that long. Sensei stops a helper by itself only after 3 hours. A helper stopped that way keeps its session, so the agent can continue it.

When a turn fails

A banner above the composer says the turn failed and why. Press Try Again to send your last message again. When the banner offers Reconnect or Fix it instead, the provider needs attention first. For workspace work, ask an admin to fix workspace Credentials. For personal work, fix your Subscription, or use Sensei free if you hold none. Reassigning never supplies personal Subscriptions to workspace work. See AI providers.

Under Review, Completed and Archived sessions

An Under Review session takes your messages and stays Under Review: the message goes out as it would in an Active session, and the pull request stays ready. When the agent changes the pull request, it turns it back into a draft first, and the session follows GitHub to Active. A sync alone keeps it ready. Above the composer, a banner reads "Ready for review. Send a message or Resume to keep working." It links to the pull request by its number. Press Resume to turn the pull request back into a draft without sending anything.

Completed and Archived sessions take no messages. To work on an archived session, restore it. See Chat sessions and tasks.

Next steps