agent-foremanlisted

# Agent Foreman ## Preserve root authority Treat the model currently conversing with the user as Main and Root. Keep user-intent interpretation, architecture selection, shared-seam integration, completion authority, and final communication with Main. Never replace Main because another model appears stronger. Never infer a model identity from a requested selector. Record an actual model identifier only when the host provides authoritative readback. Record unavailable identity verification explicitly. Treat a worker, advisor, or evaluator as a bounded contributor, never as a replacement Root. Treat deterministic evidence as authoritative over every model verdict. Do not create a child merely because this skill is active. Check the host rules, repository instructions, user authorization, and available delegation controls before dispatch. Continue with `main_direct` when child creation is unavailable, unauthorized, unverifiable, or wasteful. ## Load references progressively Read [routing-and-authority.md](references/routing-and-authority.md) on every use. Read only the selected profile sections in [plan-contract.md](references/plan-contract.md). Consult [plan.schema.json](references/plan.schema.json) when creating or validating canonical artifacts. Read [failure-taxonomy.md](references/failure-taxonomy.md) when lifecycle, persistence, security, concurrency, replay, transport, migration, cancellation, or background work applies. Read the relevant patterns in [gate-catalog.md]