Use when the user wants Claude Code to make progress on a Steward backlog. Two modes. With NO arguments it runs a conversational session — it shows the current backlog, asks what the user wants to accomplish (scope, how much, autonomy), does that work, then hands back a parameterized command to schedule as a nightly routine. With ARGUMENTS it runs directly with no conversation — a specific item (by id or reference), a steward's top item (by name/id), or a batch session (a max item count + scope + autonomy) suitable for an unattended nightly Claude Code Routine. A run tends in-flight backlog PRs (resolving conflicts, fixing CI, catching superseded work), claims and implements queued items through linked PRs, reconciles items whose work is already done, and reports what was done plus the decisions the user must make.
Use when defining, previewing, creating, activating, or updating a steward for a repository through the Steward MCP server. Inspect the codebase, draft a narrow rubric-centric steward spec, validate and preview it with configure_steward, create it only after user approval, then reconcile inventory and initialization artifacts from the coding agent.
Use before implementing a plan, when the user asks for a second opinion on an approach, after ExitPlanMode, or whenever a draft design is on the table. Consult the repository's stewards via the Steward MCP server, return their attributed feedback, and, if the plan crosses a domain that no steward currently owns, propose a narrow new steward for that domain before code is written.