run-parallellisted

# Run parallel Parallel workers are only worth their cost when every rule that judges them is fixed before the first one runs. The enemy is the run that decides what counts as winning after reading the results, so the pick bends to whichever output looks familiar. The overcorrection is a parallel run for work one delegate could do, which buys N reports and a merge for no gain in coverage. ## When to use - The user invokes it with a shape: coverage splits a scope into slices, race runs one brief N times, gauntlet runs N distinct checks against one artifact, contest has N candidates build the same thing. - Not for a plan's tasks: `run-plan` fans those out in waves. - Not for one read-only lookup: the `exo:locate-code` agent answers it. ## The run Each delegate brief follows the delegation contract of the user's `CLAUDE.md` and the reader budget in route-skills; this skill adds only what N workers need. 1. **Frame.** Before any spawn, run `node "${CLAUDE_SKILL_DIR}/../../lib/scratch-exclude.mjs"`, so `.exo/` stays untracked, then `node "${CLAUDE_SKILL_DIR}/../../lib/scratch-path.mjs" run-parallel/<slug>`, which prints `<run dir>`, and write `<run dir>/frame.md`: the done predicate, the shape, N, and each worker's slice, arm or check with its own output path, inside that worker's checkout: a worker in a worktree the run made writes to its worktree's `.exo/run-parallel/<slug>/`, read there before the worktree is removed, because a path outside the worktree stops the worker.