manager-looplisted

# Manager Loop One coordinator, many workers, a heartbeat instead of a human trigger between steps. The coordinator never implements — it dispatches, watches, routes feedback, and gates merges. Each worker gets its own git worktree, branch, and PR, and is told explicitly to keep going until the deliverable is actually mergeable, not just "implemented." This is the outer loop that runs many `evaluator-optimizer` convergence loops in parallel (each worker converging its own PR against a pass/fail bar), with `advisor-checkpoint` as the review step when a worker's approach needs a second opinion before or after the work lands. ## When to use it Fits a batch with several genuinely independent units — files, projects, or PRs that don't depend on each other's outcome. Doesn't fit a single task with one thread of work (there's nothing to coordinate), or units so interdependent that running them one at a time, in order, with full context carried forward is actually required. For a single feature-level task with several sequential subtasks (one branch, one PR, tasks that build on each other) rather than a batch of independent units, use `superpowers:subagent-driven-development` instead — same session, fresh subagent per subtask, code review between each. Reach for this skill only once the work is actually a multi-unit batch; standing up worktrees, branches, and PRs per subtask for what is really one feature is the heavier tool for the wrong job. ## The shape 1. **Dispatch.** For