Split this work into independent subagents that run in parallel: {{work}}.

For each subagent specify:
- A self-contained prompt (it shares no context with siblings — include file paths).
- The exact output shape you expect back.
- Read-only vs. write; if writing, isolate via a worktree to avoid conflicts.

Only parallelize truly independent tasks. Sequence anything with shared state. Merge/dedup results yourself after they return.