Module graph 

The competition graph.

prep ──► implement ×N ──► judge ×M (blind) ──► split? ──► deliberate ──► vote (private)
                                                  │                          │
                                                  └──── unanimous ───────────┤
                                                                             ▼
  merge ◄── gate ◄── review ×R + E2E, fix, repeat ◄── fold losers ◄──────── tally

Every node persists before the next one starts, so a run can be resumed after a crash, a rate limit, or a reboot without re-spending the work that already landed.

The design decision that matters most is where the facilitator lives. There is no moderator agent: magi assigns the labels, decides the presentation order, relays the transcript, and collects the final votes one-to-one. A moderator that never learns an author cannot leak one.

Structs

Pause

A request to park the run at its next node boundary.

Runner

Drives one run.

Functions

fold_run

Tear a run’s worktrees and branches down.

landing_subject_source

What the squash subject falls back to when the pull request title is empty or candidate-shaped: for a review-only run the derived title, never the review prompt held in instruction.

landing_title

The pull request title to hand to land::merge_subject. A review-only run opened by an earlier build titled its pull request with the review prompt; that title is dropped (empty, so the fallback applies) rather than landed. Any other title, including an operator’s rename, passes through untouched, and so does every title of a run that implements a task.

worst_open

Severity of the worst open finding in the last review round, for reporting.