cat-harness

deep-interview

Socratic requirements interview with mathematical ambiguity gating. Use for any implementation-shaped request with ambiguous intent, scope, or acceptance criteria (router ladder rule 2); when requirements are unclear or the user says "requirements unclear", "not sure exactly what I want", "don't assume", "interview me", "deep interview"; or when a vague idea needs crystallizing before planning. Produces a pending-approval spec, then hands off to ralplan/ultragoal/team — never implements.

ralplan

Consensus planning workflow — a planner agent drafts with RALPLAN-DR deliberation, fresh architect and critic agents review the same persisted artifact in parallel until join-gate consensus (iteration cap by risk tier — 5 for full/high-risk, 2 for lite/low-risk), then intent reconciliation, an ADR-style pending-approval plan, and a structured approval handoff to ultragoal or team. Use for clear-but-risky work — requirements known but non-trivial architecture, sequencing, or verification risk (migration, security, breaking change, data loss, multi-system; router ladder rule 3) — when the user says "consensus plan" or "$ralplan", or on an incoming deep-interview handoff. Planning only; never implements.

team

Coordinated multi-lane execution through native subagent fan-out (no tmux). Use when a task splits into 3+ independent parallel lanes, each owning its own verification. Maintains an evidence-backed task board via the sanctioned state writer; the leader may not declare success until every task has completion evidence.

ultragoal

Durable multi-goal execution engine over .cat goal artifacts with evidence-gated, receipt-verified completion. Use for clear multi-goal / multi-step execution requests (router ladder rule 4), for approved plan handoffs from ralplan, or when the user says "ultragoal" / "$ultragoal". Not for vague requests (deep-interview) or unplanned high-risk architecture work (ralplan).