Backlog Refine

Batch orchestration of issue-refine; use it directly for one issue.

Project facts from AGENTS.md: § Project Board (which status is backlog, which is ready), § Repositories, § Code Layout & Tech Stack. Missing section: name it and stop.

Scope: backlog status by default; otherwise issue list, label query, or milestone. Explicit board: none is valid per issue-plan: skip board operations, report refinement without transition; stop without explicit scope. Missing board facts: stop.

Rules

• Apply dev-cycle § Subagent Dispatch to every worker and re-dispatch: foreground by default, background only on explicit user request. Same-wave concurrency does not imply background permission.
• Non-interactive workers return questions and already-clear items; the caller holds the user gate and re-dispatches with answers.
• Mandatory human gate: no fan-out beforehand, refinement with unanswered questions, or invented scope. "Skip", "defer", and "close" are valid answers.
• Only issue-refine moves an item to ready, updates a body, or splits (parent kept, native sub-issues). This skill writes nothing but .tmp-* body files.
• Do not close, relabel, or reassign beyond what the user approved at the gate.

Flow

1. Enumerate all pages: gh api --paginate 'repos/<owner>/<repo>/issues?state=open&per_page=100'; exclude pull_request entries and apply scope filters. For boards, use issue-plan § Board Status Transitions' complete lookup, filtering to consuming-repo Issues then backlog status. Incomplete enumeration: stop. Retain canonical URLs/repo identity in tracking and dispatch; scope number-based commands to § Repositories; reject explicit outside targets. Flag obsolete/duplicate items, never drop them.
2. Map scope and parent/child dependencies from issue text and a quick code search. Wave 1 has no unresolved dependencies; later waves depend only on earlier ones; same-wave items are independent.
3. Assess the What using issue-refine dimensions: body, comments, labels, linked context; code only for feasibility. Surface all material ambiguity, not just blockers; defer design to issue-plan.
4. Gate once: numbered questions grouped by wave/item, ordered by wave, priority, age, plus go-ahead; list bugs needing diagnosis (issue-refine § Bugs) with go-ahead to diagnose. Stop for answers; approval covers only cleared items.
5. Track item, wave, state (needs-answers / needs-diagnosis / clarified / refined->ready / skipped / deferred), notes; in this conversation, never in a project file.
6. Diagnose approved needs-diagnosis items with one architect issue-diagnose worker each, then refine them in a later wave. Refine waves with one issue-refine worker per item (sequentially here if unavailable). Pass clarifications as facts, go-ahead as approval, and unique .tmp-refined-body-<number>.md paths. Start only after prerequisite items are refined; split parents precede children in later waves.
7. Confirm each item reached ready (issue-refine verifies), or was refined with board: none; record failures honestly.
8. Report: refined and readied; diagnosed (RCA links); skipped, deferred, closed; still blocked on answers; wave structure; recommended next batch.