take-on-storylisted

You are replacing the cascade, not the roles. Every task still runs in CI; you decide *when*, watch *whether it worked*, and keep the story's state legible. The conduct rules are in `rules/claude-session.md` §Driving a story — this skill is the procedure. ## 0. Refuse the wrong repo Read the stub (`.github/workflows/claude.yml`): a bot named in `allowed_bots` means a live cascade, and two drivers race for the same labels. Say so and stop. Drive dark repos only. ## 1. Orient Fetch. Read the story issue fresh: its branch line, its sequencing section, its open tasks, any handoff comments already posted. The team's own files define what these mean — read them from the consuming repo, don't assume. Post one comment on the story: driving begins, which wave is next. ## 2. The wave loop 1. Reconcile: re-read the story and its tasks from GitHub. Never act on remembered state. ⚠️ **One batched query per transition, and prefer REST.** `gh issue view` / `gh pr view` / `gh run view --json` bill the **GraphQL** pool; `gh api repos/{o}/{r}/issues/{n}` and `.../actions/runs/{id}` bill **core** — two independent budgets, and concurrency drains GraphQL first. Driving three stories exhausted it twice in one hour while core sat near a quarter used (#93). Ask for every task and story in one aliased GraphQL query rather than one call each; it is also the only *consistent* read, since separate calls describe separate moments. Read immutable state — titles, the task list,