parallel-agents-lightlisted

# Parallel Agents Light Use this skill to get useful parallelism without paying the coordination cost of a full campaign. The controller stays responsible for repo state, integration, verification, and the final answer. ## Decision Ladder Choose the lightest mode that can finish the task safely: 1. **Local cycle**: use the main Codex thread when the work is one objective in one tight file set. Pull in `$repo-conventions` if repeated inspect/edit/verify cycles are enough. 2. **Read-only sidecars**: launch 1-3 bounded subagents for independent evidence gathering, review, docs drift checks, test mapping, or risk scans. Keep all writes with the controller. 3. **Split implementation**: launch 2-3 workers only when each worker has a disjoint write set and a clear verification command. Keep one integration owner in the controller thread. 4. **Full campaign**: escalate to `$manager` when the work needs durable plan artifacts, worktrees, dependency groups, managed merges, or more than three write lanes. 5. **External terminal orchestration**: when the user explicitly wants dmux, tmux, panes, or separate terminal sessions, apply the same ownership and integration rules here and use only tooling that is actually available. ## Preflight Before launching workers: - Inspect current repo state with `git status --short` when a git repo exists. - Read the files needed to define ownership. Do not launch blind workers from a vague prompt. - State or intern