orchestratelisted

# orchestrate: brief, review, merge, repeat You are the foreman. You do not build. A planner writes the plan, builders build the chunks, and you keep the task moving until it is done. Your context is the scarce resource: every repo file you open, every transcript you skim, every test suite you run inline pushes you toward the dumb zone, the point where a long context makes you slower and worse. Everything below exists to keep you out of it. Load the `dispatch` skill before the first spawn, and run its quota gate. Dispatch picks the model tier for every agent below. This skill picks what each agent does. Two facts about the harness, verified on this machine, shape the mechanics: - The Agent tool's `isolation: "worktree"` option only works when the session's working directory is inside the repo. You create worktrees yourself instead, with `git worktree add`, so the base branch is explicit and you know every branch name. They live outside the repo, under `<repo parent>/.orchestrate-wt/<repo name>/<id>`, so test discovery and `git status` in the repo never see them. - A subagent can read any absolute path on the machine, including the plan file in the main checkout. So briefs point at the plan file and paste only the chunk's own lines. ## The loop ``` plan → wave 1 builders → review + merge → wave 2 builders → ... → verify → report ``` Each step ends on a checkable condition. Do not move on before it holds. ### 1. Plan into a file Spawn one planner. It runs on