Use GitHub to inspect BullyBearAI epic #1090, its linked workstreams, current issues, and open or merged PRs. Read CLAUDE.md and the available project delivery instructions before changing the backlog. Refresh the epic around its customer outcome, measurable success criteria, scope, non-goals, dependencies, and closing evidence. Reuse existing issues. Create a bounded child only for a verified gap that is still within the epic. For each child, record: user outcome; exact scope; acceptance checks; parent and dependencies; allowed actions; expected output; and proof still needed from a person or live system. Classify the implementation as AFK-ready, HITL-required, or blocked by a named dependency. Use those classifications in the issue body; apply the repository's agreed labels, or propose a consistent label scheme if none exists. Keep execution-environment labels such as local/cloud separate. Split mixed work: deterministic evidence-lineage code can be AFK, while a study showing that it helps a real customer remains HITL. Never mark the parent complete just because its code child is implemented. Return the dependency-ordered backlog and a concrete first AFK batch.