Chats

Chat is a conversation with an agent outside any issue: ask something, clarify, request a summary or a draft. An agent in chat knows the workspace — issues, projects, its own skills — and can do things as well as answer: file an issue, move a status, prepare a draft.

Two surfaces, one chat

• Chats — the workspace home. On the left, one scroll of sections: Agents, Channels and Direct messages, and, once turned on in «⋯ → Configure sections…» beside «Chats», Waiting for me and Archive. A section collapses, moves and hides, an empty one is not shown; the «+» in its header starts a thread with an agent, a channel or a direct message. An agent's row opens its forum: your threads with it and the tasks you both take part in. On the right, the open thread, or a new thread with the agent you pick when nothing is open. A thread bound to a task lives in the task and is not a chat of its own.
• The floating window — chat over any page, for asking a question without interrupting your work. Turn it on in Settings → Chat; on the Chats page it hides, so as not to duplicate what is already on screen.

Both surfaces hold the same conversations: the session history is shared.

Chat or a mention

• A mention in an issue pulls the agent inside it: it sees the issue and the thread and can change them.
• Chat is attached to nothing by itself: the topic stands alone, and the result is an answer to you rather than a record on an issue.

If the question concerns a specific issue, it belongs in a mention there; if the topic stands on its own, start a chat.

When a conversation grows into work, press Make a task: a new issue appears and the chat is bound to it. The chat stays a chat with the same agent, but everyone who sees the issue now sees it and can write in it, and the issue's feed notes it. You can bind a chat to an existing issue, move it to another or unbind it later; once unbound, everyone who wrote in it stays in it. A direct conversation or a whole channel is never bound, but a single message from one is: its branch is bound, and replies to the message keep going there. Whoever sees the issue reads the message and its branch, not the rest of the conversation or channel; the message's author stays in the branch.

The conversation

Pick an agent and write. If you are not sure where to start, the interface offers openers — show my open issues by priority, summarize what I did today, plan what to do next.

Enter sends the message and Shift+Enter breaks the line. In Settings → Preferences → Send message with you can pick ⌘↵ (Ctrl+Enter on Windows and Linux): Enter then breaks the line and continues a list. The setting applies wherever you write a message, including editing one and answering an agent; forms such as an issue description always submit with ⌘↵.

While the agent answers you can see what it is doing: Queued, Starting, Thinking, Typing, and on tool steps Running a command, Reading files, Searching the web. An answer can be expanded into its details: how many steps it took, which tools it called, and what came back. A long answer can be stopped — the partial result stays in the conversation.

A message sent while the agent works waits in the queue for its next turn. To correct the agent right now, send with ⌘↵ (with ⇧⌘↵ if you send with ⌘↵). The label beside it says what happens: Add to the current turn — the agent reads your words without stopping; Stop the step and start over — the agent stops its current step and goes on from your message; As the next turn — an agent on a laptop reads it when its turn ends. A message to another agent always waits in the queue.

Sessions

Each conversation is a session with a title, a history, and an unread marker. A session can be renamed, its current run stopped, or the whole thing deleted along with its messages — deletion is irreversible.

Chats are private. Only the person who started a session can see it; to everyone else it simply does not exist. That is deliberate: a conversation holds one person's working context and the instructions their agent acts on.

Context between messages

The agent continues the conversation rather than re-reading it: Bounc remembers the provider session and returns the agent to it with your next message.

Next

• Assigning and mentioning — handing an agent work on an issue.
• Agents — configuring who you are talking to.
• Agent tasks — the run behind every answer.