delegatelisted

# Delegate Run before every `Agent()` call — cheap enough to reload each time. Spawning is the dominant token cost: every subagent rebuilds its context cache from scratch, and cache creation, not model output, is the bulk of the bill. The cheapest work is the spawn that never happens. ## Pre-Delegation Checklist 1. **Total scope size** — sum the estimated diff (excluding comments) across the whole batch, not per item. Under ~300 lines total (soft): do it in the coordinator, or fold into an agent already live in scope via `SendMessage`; do not spawn. 2. **Genuine parallelism** — a real wall-clock or file-independence need, or would sequential merely take "a bit longer"? No real need → one agent, sequential. 3. **Reuse** — an agent already live in the same file/domain scope? → `SendMessage` it (`grand-admiral` § Agent Reuse). 4. **Model tier** — set explicitly on this spawn per the table below; never leave it to the frontmatter fallback. 5. **Worktree** — code-mutating agents: pre-create it and inject the resolved SHA; never rely on `isolation: "worktree"` (`grand-admiral` § Worktree Isolation). 6. **Monitoring** — is the built-in Monitor running? An un-monitored dispatch is a doctrine violation (`grand-admiral` § Recovery → Stall Watchdog). 7. **Development work?** — brief the goal only, no file list/approach; the agent plans, the coordinator approves (`grand-admiral` § Development-Work Delegation). **Anti-pattern — file-independence is not spawn-justification.** Real case