A decision is never edited into a new meaning — it is superseded, and both stay readable. Recording the alternatives is what makes it reviewable later; a decision without them is an assertion.

Agents are shown what is approved now and what is explicitly excluded, with the reason. Work linked to anything other than approved now is refused at dispatch. DUEX cannot judge semantically whether an objective strays — so an agent that thinks its work conflicts raises a concern here for a human, rather than deciding alone.

Every agent receives active memory before it starts, grouped by kind. One live statement per subject — a correction supersedes the old one rather than sitting beside it, and both stay readable. Memory is never enforced: an agent whose work appears to contradict a remembered statement raises a conflict here instead of refusing, because a stale note should not be able to stop work.

A node never stores its own status. One that is about a task, a decision or a scope item shows that thing’s current state, read fresh every time the canvas loads — so the graph cannot quietly fall out of step with the work the way a diagram does. A node is not a task and does not map one-to-one onto one. Dependency cycles are refused when you draw them: a loop means nothing in it could ever start. Drag a node to place it; the position is saved.

The runner clones this, so it must be reachable from the runner's machine — an SSH remote with a deploy key, or a local path like /opt/myapp. Never include a username and password in the URL; DUEX refuses to store git credentials.

Only allowlisted GitHub accounts can sign in. Removing someone revokes their live sessions immediately.