Chat Mode & Work Mode

Every chat is in one of two modes. You switch between them, and the conversation carries over: the agent that helped you plan an issue is the one that builds it.

Codex gets the same split through its sandbox: read-only in Chat Mode, workspace-write in Work Mode, with its approval requests shown as the same cards.

The composer shows the mode: Chat, or a hammer with the issue ID in Work Mode. A divider in the thread marks every switch: "Entered Work Mode on myapp-a1b2c3", "Entered Chat Mode".

Chat Mode

The default. The agent reads the repository and changes issues: it can list and search the backlog, file and update issues, write specs, and use the read-only git tools to see what an issue changed. Ask it for a code change and it offers to file an issue or start work on one. To start, it calls start, and you approve it in a card that reads "Enter Work Mode on …?". Once you allow it, the chat continues in Work Mode on its own.

Work Mode

Work Mode

The agent runs in the issue's worktree: it reads, edits and writes files and runs commands there without a card. It loads the xpo skill first and follows your project's workflow (spec, implement, test, hand off) including where it says to stop for your review. It also gets a task list and can hand work to subagents.

Each step shows live on one line: the current action and a timer while it works, then "6 actions · 2 failed". Click it to see every action on a timeline.

merge always asks first. After a successful merge, the chat goes back to Chat Mode by itself.

Getting into Work Mode

• Start Work on a PLANNED issue opens a new chat in Work Mode, titled after the issue, with a first message ready to send. If a chat is already working on the issue, it opens that chat.
• Continue Work on an in-progress issue opens its Work Mode chat, or creates one on the existing worktree.
• Enter work mode on a past chat — from its row actions (right-click, ⋯) — puts it back to work on the issue with its whole conversation.

Only one chat can be in Work Mode on an issue at a time. Start Work is disabled while the issue is in BACKLOG, while a blocker is unresolved, or when the project isn't a git repository. Its tooltip says why.

Exit work mode, in the composer's mode menu, returns the chat to Chat Mode. The worktree stays, and the issue stays in progress.

Background tasks

The agent can start things that keep running, such as a dev server or a test watcher. The turn still ends when the agent replies, and the composer shows "Monitoring N tasks" (or "N agents working" for subagents). Click it to see the tasks and stop them. If a task finishes and the agent has something to say, its reply shows as its own turn.

Stop interrupts the current turn only; background tasks keep running. Switching modes restarts the agent and stops its tasks.