start-story-multi-agentlisted

# Claude multi-agent story flow Use Claude Code to implement, Codex CLI to review in a read-only sandbox, and the human to open, review, and merge the pull request. Treat this split as a workflow default, not a quality or cost guarantee. ## Enforce the contract 1. Use one story, one worktree, one branch, and one active implementer. 2. Make specs/<ID>.md the local execution contract. If a remote source is needed, query it exactly once with the narrowest sufficient query, then work locally. 3. Resolve the base from origin/HEAD, BASE_BRANCH, or an explicit script argument. Never checkout or pull the primary worktree to create a story. 4. Review only the base-to-branch diff plus files needed to validate a concrete finding. 5. Allow at most one review-to-fix pass, followed by one re-review. 6. Let executable verification and exit codes decide whether work is ready. 7. Stop before push, PR creation, merge, rebase, or worktree removal. A human owns those external and destructive gates. 8. Keep trace and review artifacts under the story worktree's ignored .claude-flow/<ID>/ directory. ## Set up a project Require Git, Claude Code, Codex CLI, jq, and the repository's own development toolchain. Run setup from the target project root: ~~~bash mkdir -p specs artifacts scripts .claude/hooks cp "${CLAUDE_SKILL_DIR}/assets/spec-template.md" specs/_TEMPLATE.md cp "${CLAUDE_SKILL_DIR}/assets/format-hook.sh" .claude/hooks/format-file.sh chmod +x .claude/hooks/format-file.sh ~~