Skip to main content
/btw asks Orchestra a question about the session you are looking at, without interrupting the agent or changing what it knows. Use it to catch up on a session, check what the agent decided, or ask something you don’t want to put in the agent’s context.
Orchestra answers from the session as it is recorded, so /btw works the same way on every agent: Claude Code, Codex, OpenCode, Gemini CLI, Cursor and Orchestra’s own. It also works on a session captured from someone’s own terminal, and on one that nothing is running right now.

What happens when you ask

  • The agent sees nothing. The question is never sent to the agent, its current turn carries on, and its context does not change.
  • Orchestra answers from what you can see. The answer reads the session’s messages and tool calls through your own access, the same as every other read in Orchestra. It never uses anything from the session you couldn’t open yourself.
  • Only you see it. The answer shows in your terminal, or above the message box in the web and desktop app. Nobody else in the session sees the question or the answer.
  • Both are kept. The question and the answer are recorded in your workspace’s history as yours, so there is a record of what was asked and what Orchestra said. They are not added to the session’s conversation or to search.

Share an answer with the session

When the answer would help everyone, share it:
  • In the terminal, type /btw share. It shares your latest side question in this session.
  • In the web or desktop app, choose Share with the session on the answer.
Sharing posts the question and the answer to the session as a note to the people in it, marked as Orchestra’s answer. The agent still doesn’t see it. Sharing the same answer again does nothing new.

Which model answers

Claude Code sessions on your computer. When you ask in the terminal on the computer where a Claude Code session’s transcript lives (one you run there, or one Orchestra captures there), Claude Code itself answers. Orchestra opens a private copy of the session in Claude Code, with no tools, hooks or MCP servers, asks your question once, and throws the copy away. It reads the session exactly as the agent has it, including tool output that Orchestra shortens, and it runs on your own Claude Code account. Nothing is saved to the session, and the agent’s own conversation is untouched. If Claude Code can’t answer (it isn’t signed in, or it takes longer than two minutes), Orchestra answers instead and says why. A shared answer says which one answered. Everywhere else. Side questions use the same model as Conductor. If your workspace uses its own model keys, that key answers. In Orchestra Local, your own key or your signed-in Claude Code or Codex answers. When no model is set up, the question is still recorded, and the answer says that there is no model to answer with. Answers are short: a few sentences or a short list. Very long sessions are read from the most recent messages back, and the answer says so when older messages were left out.

Limits

  • One person can ask up to 10 side questions a minute.
  • A question can be up to 4,000 characters.
  • Agents can’t use /btw. An agent that wants to look something up uses Orchestra’s recall tool instead.