handoverlisted

# Handover ## Boundary Use this skill when one unfinished, mechanical, file-scoped unit should run in a fresh conversation. The original orchestrator keeps the plan, acceptance, review, and git ownership. Use `sprint` for sequential plan-and-build, `advisor` for judgment, and `ship` for closing finished work. Do not invoke raw `codex exec` or `claude -p`: the handover protocol supplies the allowlist, recursion guard, preflight, sandbox, result classification, and model anchor. ## Native read-only workers For **Explicit Agent Worker or parallel read-only research**, start at most two depth-one workers, disable nested delegation, and keep every pilot worker read-only. Each request carries `objective`, `scope`, `deliverable`, `acceptance`, `permissions`, and `budget`; each WorkerResult returns `status`, `findings`, `evidence`, and `gaps`. The main agent verifies evidence, deduplicates findings, and records elapsed time, resolved model, and runtime events. Record token usage only when the runtime reports it, never from worker estimates. ## Modes | Invocation | Mode | Ownership | |---|---|---| | `/handover [--agent ...] [slug]` | sync fresh worker | source host keeps git | | `/handover --async [slug]` | async Codex payload only | human runs it later | Use sync when this turn should wait for one bounded result. Use async only when the handoff artifact should be stored without spawning a worker. ## Gate Before either mode, require: supported Claude/Codex platform; no `VERB