model-checklisted

# Model check Use at intake or when the task shape materially changes. The output is a routing decision, not permission to begin the underlying work. Score 0–2 on: 1. reasoning depth; 2. ambiguity; 3. novelty; 4. blast radius if wrong; 5. benefit from extended thinking. Choose model capability from ambiguity + novelty + blast radius. Choose effort from reasoning depth + extended-thinking benefit + whether the work requires sustained multi-file iteration. Favor the cheaper option on a genuine tie, but never trade away a required correctness, privacy, security, migration, or production-safety capability. Prefer one capable runner for a coherent vertical slice. Use worker subagents only for genuinely independent subtasks with self-contained briefs; cold-start context duplicated across workers is token waste. A separate agent is valuable for a capability the primary runner lacks or for the one independent review required by repository policy, not as a simulated department. Before recommending any model, check whether a deterministic script already covers the work — `script-maker`'s mechanism hierarchy controls that choice, and neither a local nor a hosted LLM is a substitute for it. Only for work that still requires model judgment: when a task is a good candidate per `.claude/skills/local-worker/SKILL.md`'s bounded-delegation criteria, recommend local execution over hosted Claude only when the combined cost of dispatch + verification + correction is materially lower than do