💀 Pulls & wipes
A pull is one attempt at the fight. Pull 1 opens when the run starts; every wipe closes the current pull, and the next pull opens with the repair. A wipe is a real failed gate — one of exactly these recorded events:
- A verification stage fails for real — a
FAILorWOBBLEtransition that drops the machine back to fixing. - A test run stays red where green was required(the implement and refactor steps). The test-writing step’s red run is expected — that one is never a wipe.
- The coding agent itself fails its task.
- A turn contract checkfails — an agent’s input or output was empty, malformed, or a refusal. Transport success is not task success; a check failure is a wipe like any other.
One mistake is one wipe: consecutive failure triggers inside the same step (say, the agent fails and its test run then fails) merge into a single wipe.
The repair is real
After a wipe, the Fix & Resume turn (flavor: the Field Medic) runs a real repair: the coding agent receives the actual failure output of the stage that wiped — the real test log, the real judge verdict — fixes the real cause in the workspace, and commits the fix. The run then resumes exactly the stage that wiped. Both the failure and the fix stay on the record; neither is ever erased.
How a pull ends
- Cleared — the run reached completion.
- In progress — the run is still writing its transcript.
- Abandoned — the transcript went silent without completing; the game says so instead of pretending the fight is still on.