One task, from the request to the verdict. The agent is never asked to cooperate: every step below is driven by a hook event it was going to fire anyway, and the evidence is read out of command output it was going to produce anyway. Read top to bottom — session and claim, then the edits that grow the task's own scope, then the command results that become evidence bound to exact bytes, then the stop where a state is computed rather than believed.
The system in tiers. core/ never imports the plugin, and the plugin is deliberately thin — everything interesting is importable, and therefore testable. hooks.json is generated from wiring.py, because the scope guard once shipped completely inert while every unit test passed. eval/ is a separate authority: it grades in an evaluator-owned workspace from an exported patch, never inside the candidate's, and it shares none of the runtime's judgement.
Read each lane left-to-right. A task is what the shipped plugin does today. Evaluation is one paired run of the harness. The third lane is planned and not built — it is drawn so this page shows the intended shape as well as the current one. Everything in the Overview tab is real code today; lane three is not.