Keep your focus while it works.Keep your code when it’s done.

So you never have to check.

A terminal cannot tap you on the shoulder. Claude Code, Codex, every CLI agent — they all finish into a window you are not looking at, so you poll it, and every check is a context switch you paid for. Muninn knocks instead: three soft hits, evenly spaced, never escalating, and a lower sound when it is asking rather than finishing. The panel waits in the corner until you have seen it. It does not take focus, and it does not time out.

It doesn’t guess. It asks.

Your agent ends each turn with a structured block — done, changed, verified, next, blocked, risk, explain — and Muninn renders exactly that. No second model between you and the work. MUNINN.md teaches the format in one file; a missing summary says finished, no summary, never an invented one. And tests pass, said by your agent, stays a quotation — Muninn quotes, it never vouches.

Prefer skills to rules? muninn init --skill also installs a /muninn skill — the same contract, invocable on demand: “summarize that diff as a muninn block”, or re-emit a summary the panel missed. The rule stays the backbone, because a closing summary is an every-turn obligation and skills load when the model judges them relevant — the wrong trigger for “always”. Either way your CLAUDE.md stays at one line.

So the code stays yours.

explain is held to a standard: how it works, which function does what, the constraint that ruled the other approach out — then the name of the technique, so something transfers to a project this one never touched. After reading it you should be able to predict what the code does with an input you have not seen. That is the difference between accepting a change and understanding it.

Every turn, kept — with the prompt that caused it.

The last fifty turns in a plain JSON file on your own disk, each with the prompt that started it — a conversation, not a list of outcomes. Search ranks project names first and highlights the match inside the result; export any of it as markdown. The tray keeps the recent few to hand.

Somewhere for your attention to rest.

A feed is context-switching on a loop; a maze, a minefield or a chess position is one thing at a time. Minesweeper in a 5×5×5 volume, a first-person maze, chess against a real Stockfish running as WebAssembly on your own machine, a runner, a drifting raven. Progress banks every second and the summary always wins the screen. No streaks, no scores kept overnight — a thing that wanted your attention would be the problem again, in a nicer coat.