Agent-agnostic task dispatcher. Reads backlog on main to find active changes, fetches the feature branch, reads tasks.md, picks up the next executable task group, implements it, and pushes to remote. Runnable by any agent (Codex / Claude Code / Cursor / cron / GitHub Actions / one-shot) — only requires git + shell + network.
Propose an OpenSpec change — scan backlog or accept a specific B-NNN, check PRD readiness and dependencies, generate four-piece set (proposal/specs/design/tasks) on feature branch, create Draft PR, and update main indexes. Works both as one-shot (`/change-propose B-045`) and scheduled by a runner.
Interactive AI skill (Claude Code or Codex) for reviewing dispatch-completed changes. Uses branch-centric model where review happens via PR. Handles PR review, fractal documentation sync on branch, PR merge to main, three-dimension verify, delta specs sync, archiving, and backlog status update. Use when dispatch runner has completed task groups (status review) or when manually triggered.
Interactive AI skill (Claude Code or Codex) for designing product modules AND decomposing them into backlog idea entries in a single conversation. Reads design/inputs (brainstorming / figma / interviews), builds the module design document (modules/M-NNN-<slug>.md), splits the module into N backlog entries with auto-suggested types, and writes both the module and the decomposed backlog. Also syncs design/roadmap.md AUTO sections in real time. Use when the user wants to design a new module, revise an existing one, or decompose a big idea into actionable backlog entries.
Interactive AI skill (Claude Code or Codex) for generating Product Requirements Documents (PRDs). Takes a backlog entry (already tagged with a module) and produces a structured PRD through user dialogue, grounded in the module design document and referenced design inputs. Use when the user wants to explore, analyze, or define requirements for a backlog item.