The plain-language board

Models keep getting better at the work and harder to follow: their progress reports read like notes for another machine. mu does not ask the working model to report. With the board on, a second model, chosen only because it speaks plainly, tells you what is going on, and Jev decides what is worth telling.

What the board shows

• Where things stand. What the agent is doing now, how many of the acceptance items are done, and what waits for you.
• A running account. A line for every step as it happens: which file changed, whether a check passed, which command ran. A string of file reads folds into one line, read 5 files.
• News, retold. Whenever the agent says something or finds something, board.read asks the judge whether it is news or routine. Only news is retold on the spot by the plain-speaking model.
• A summing up. When a run ends, its news is picked once more for a summary, the last line of the account, with the steps kept above it. So you see not only done, but what was done.

At the top, two numbers: how full the context is and the cache hit rate, the direct result of the context and cache decisions.

The board is for you only; it never enters the model's context. The working model keeps working in its own language.

Switching it on

The board is switched per project: /board on and /board off in a session, or the switch on the app's board tab. The first time, you pick the model that writes it; Claude Opus 4.6 is recommended, then Gemini 3.8 Flash. /board model changes it.

It is off for new projects unless you set features.board.defaultOn, because every update costs a model call.

How often it looks

The judge looks whenever the agent says something, after a check or a ticked acceptance item, every five tool calls, and whenever the agent stops, but not more often than every 20 seconds while it works. The board follows the permission mode, a running goal and sub-agents, so it says waiting for your permission or two sub-agents are working when that is the case.