🗺 What is this game?
This is a software-delivery pipeline you can watch. Every project is a world. When you post a quest, real AI agents — real model calls plus a real coding agent in a real git workspace, driven by a real state machine — do the actual software work: they write a failing test, make it pass, refactor, commit, review the diff, and re-verify the result. The game layer is a skin over the recorded history of that work, nothing more.
The honesty contract
The game renders exactly what was recorded. Concretely:
- Every state transition in a run carries a basis: the real recorded evidence that earned it (a test run, a commit, an agent diff, a model verdict) — or an explicit runner policy marker (a step with no evidence to have) or an honest UNPROVEN marker. The UI shows the basis verbatim.
- Failures are wipes and are never hidden: a failed gate goes on the record, shows in the UI, and counts against the run.
- Nothing is simulated or injected. There are no scripted failures, no fake recoveries, no fabricated health numbers. A recovery story appears only when something really failed and an agent really repaired it.
What a run actually builds
A run for a world with no bound repository builds a throwaway demo project in a temporary folder on the machine running the game — real work, sandbox destination. A world bound to a local git repo (the path chip on its card) builds in that repository: the run validates it first (a real repo root, a clean checkout — refusals are recorded wipes) and its commits land on the current branch and stay there as the deliverable. Either way nothing is pushed and nothing is deployed anywhere, and the run page states which destination it used.