Coding chats

A coding chat is one task in one of your projects. It gets its own git worktree and branch, so several agents can work on the same repository without tripping over each other.

Agents and models

Choose Claude Code, Codex or OpenCode in the composer, along with the model and reasoning effort. You can switch agent part-way through a task. Claude models that support it offer a 200k or 1M context window.

Permissions

Open the model menu in the composer and choose Permissions. The choice is saved per chat and applies from the next reply.

Codex: Full access (the default), Workspace write, or Read only. The restricted modes use Codex's own sandbox, and Workspace write also turns off network access.
Claude: Full access (the default), Accept edits, or Plan mode. These are Claude's own permission modes, not a filesystem sandbox.

Moneypenny doesn't show approval dialogs: a request outside the chosen permissions fails instead.

Commands

Type / at the start of a message for the chat's commands: /compact (Claude), /clear for a fresh agent session, and the task panels /tools, /diff, /evidence, /pr, /terminal and /activity. Your shared skills are listed there too. Type @ to mention a bot, another chat, or a note to attach as context.

Queue messages while it works

While the agent works, Queue message keeps what you typed, and each queued message runs in order when the turn before it ends. A stopped or failed turn holds the queue, so nothing runs after a failure that you didn't mean to. Choose Send now to go ahead anyway.

Invite a chat into another chat

Pick a coding chat with @, or Invite a chat, to ask it a question from another chat. It answers from its own worktree and its own agent session, so it knows its own work, and it sees the inviting chat's recent conversation.

Watching background work

When Claude answers while something it started is still running, such as a test run in the background, the chat is watching: Claude keeps listening and answers each time that work reports. Stop watching ends it and every command it started.

Diffs, pull requests and the terminal

Review the diff and open a pull request from the chat when the work is ready. The terminal is your own shell in the chat's worktree. Close worktree removes the worktree but keeps the branch, commits and conversation.

Pages and diagrams in replies

Agents can answer with charts, tables, diagrams and mockups drawn inside the reply, and Mermaid diagrams render in the app's colours. Pages run in a sandbox that cannot reach the app or your account.

Next: Plan review and proof