Where a mission can stop, and how it moves

Read it as a path. Ten checks before a session starts (seven stable contract conditions, three guards hardened in operation), a witness and a human gate before anything lands, one exit command whose code depends on durability alone, and a register that sends every incident back into the next contract. The watcher is drawn in two steps: what it sees, and the receive–deliver–wake route that launches a mission from an armed contract and wakes the architect when the mission returns.

mission path named refusal learning loop

The board has an architect; the tree has executing seats; authority stays human. Claude Code, in the architect seat on the developer machine, holds the contracts and the board and writes nothing in the tree: the launcher strips its identity from every mission it starts. Claude Code, Grok Build and Codex CLI execute inside their declared worktrees, under the model and effort the contract names. The watcher carries observations and routes, never permission. And when a ruling is contested, a seat consulted outside the tree, with no execution surface and so with nothing to defend, returns a read that enters the register: never a decision.

The parcels moving through the wires are the work itself. A pipe with nothing in it is a pipe that has stopped. Every parcel is filled: everything drawn here runs. Hover or focus a block to isolate its path.