Project Supervisornew · 1.6

A project can have a supervisor: one of the agents, chosen in the header of the project. Its Supervisor pane has two parts. Chat is the conversation of the project with it: each message the user writes wakes it, in a session of its own that keeps the thread, and it answers in the chat. Instructions is a markdown file of the project (supervisor.md): how this supervisor works here, on top of its own definition.

When a specialist moves a task to Review, the supervisor reviews it: it accepts it (Done, with a comment of why), asks for changes (a comment to the specialist and the task back to Working, where the specialist resumes), or asks the user when unsure. After three rounds with the same specialist it leaves the task to the user.

It can also be called by hand: Prompt ▸ Supervisor in the menu of a task, a comment that mentions it (@ its key) on any task, or the MCP tool anxeb_supervisor_run from the session of the project. It never acts on its own on the user's tasks. What it may do is the supervisor access of the project; the tasks it creates are always Draft, for the user to decide.

While it works the chat says "{supervisor} is working…" with Stop, and what it asks shows there as cards. Its answers ring a notification unless the chat is in view.

What it does

• Automatic review of the tasks specialists move to Review.
• Accept, ask for changes, or ask the user; at most three rounds per task.
• A project chat with the supervisor, kept across messages.
• Per-project instructions (supervisor.md), editable in the app and through MCP.
• Called by hand from the task menu, by @mention, or by MCP.

How it behaves

• Its own changes never start another of its runs.
• A supervisor never reviews its own work.

Why it matters

• Work reaches you already reviewed.
• Ask the project, not a chat window: the supervisor knows its backlog.