conductlisted

# Conduct — the cmux Conductor Routes a work item to an **execution topology**, the way `/auto-research` routes a goal to a **skill**. The two compose: auto-research answers *which process*; conduct answers *what shape that process runs in* (and whether cmux is involved at all). Human writes intent, agent picks the topology. **Core discipline — the lowest rung that does the job.** Escalation is a brake, not an accelerator (`agent-loops/references/dispatch-policy.md` — cite, don't restate). Most work items route to the leftmost rungs; a fleet must be *earned* by the classification below, never assumed. Routing everything to a fleet is the fan-out-theater failure mode. **cmux is optional infrastructure.** Routes that need it require the cmux app + socket (`etk:cmux` skill ��� macOS only). When cmux is absent, conduct still classifies and emits; patterns 1–2 degrade to their non-cmux equivalents (noted per pattern in the routing map). ## Phase 1 — Classify Read the work item (`$ARGUMENTS`) and score **four axes**: | Axis | Values | Inferred from | |---|---|---| | **Scope** | single task · divisible job · full ticket | Classify on the **deliverable**, in this order: (1) a **research/estimation** deliverable — spike, "research the work needed", "turn X into tickets", "size it", "investigate options" — → **divisible**, *even when it carries a ticket ID*; (2) "audit/map/sweep across…" → **divisible**; (3) a ticket ID **or** "implement/build/fix X end-to-end" → **ticket** (this