How the loop runs a card

Two commands that never overlap. Plan writes every card and builds nothing. Run builds the cards and writes none.

The plan phase turns a brief into reviewed cards in a vault. The build phase picks one startable card, proves the gate red, has a reviewer read the card, builds it in a private worktree, runs the gate, has a fresh reviewer read the diff, and commits. A card that keeps failing parks, and the next plan phase slices it again. — Everything the build phase cannot finish goes back to the vault as a status on the card. The next plan phase is what picks it up. On a narrow screen, drag the picture sideways.

When a card fails

The gate fails or the diff is rejected, so the same builder tries again in the same worktree, three rounds at most. Still failing, the card parks, and nothing in the build phase rewrites it. Triage then says who was wrong: the work, so the next plan phase slices the card again, or the machine, so the card goes back once.

When the loop stops

A usage limit or a provider outage is the only thing it waits on, and it waits for as long as that takes. It also waits while another agent holds a claim on a card, because finishing that claim can free work here. When nothing is startable and nobody is working, waiting cannot help, so it hands back to the supervisor, which plans again and starts a fresh run.

Who does what

A fast model builds, with a stronger one behind it. Reviews go to models from another provider, on their own list. The model that wrote a thing never judges it, and no builder writes a gate or a card.

A real card Back to the start