Use when acting as main_brain inside a rendered MathorCup instance and you need to inspect the ready pool, dispatch a bounded task, observe its status, or decide close, reopen, or cancel through repo scripts instead of hand-editing runtime truth.
Use when main_brain needs to inspect one task in depth, reconstruct its event and queue history, or compare multiple worker artifacts through adjudicate_task.sh before any human-gated close, reopen, or cancel decision.
Use when acting as a bounded worker inside a rendered MathorCup instance. Read the task packet, stay inside allowed paths, and return structured feedback or retrospective artifacts through the repo gate files without closing or mutating runtime state directly.
Use when a template-source change needs a safe, low-risk smoke test. Render scaffold into a temporary instance, then run validate_agent_docs.sh and doctor.sh on the rendered target without touching backend runtime execution.
Use when maintaining the mathorcup-template repository itself rather than a rendered instance. Read scaffold as the source of truth, edit scaffold-first artifacts, preserve stable script entrypoints, and validate changes with template-source checks plus render-only smoke.
Use when you need current runtime facts inside a rendered MathorCup instance. Read .env and project/paper/runtime/paper.env as machine truth, then use runtime contracts as rendered mirrors and never confuse the native bridge with runtime state truth.