GOAL Use Claude Code and GitHub to complete the currently actionable AFK implementation work under jeanbrianc/BullyBearAI epic #1090. Use subagents for independent slices and leave reviewable draft PRs. CONTEXT Read the current epic, linked issues, open and merged PRs, CLAUDE.md, and available delivery instructions. Inspect the actual checkout and refresh the AFK/HITL classification. Follow the runtime, usage, and worker limits in long-run-brief.md. Confirm the execution environment can stay available for this run. EXECUTION The main agent owns scope, dependencies, and the run ledger. Give each implementation worker one bounded issue, a separate worktree/branch, explicit file ownership, acceptance checks, and a required return format. Serialize overlapping changes. Implement, run focused tests and required repository gates, update docs, commit, push, and create or update the draft PR. Record the issue, base/head commits, checks, artifacts, dependencies, and next action after each slice. After each batch, inspect what remains. Create further children only for real gaps within the epic. Keep a written checkpoint before a handoff or interruption so a later session can verify what is complete and resume safely. BOUNDARIES This pass authorizes code, tests, docs, issue progress, branches, and draft PRs. Do not merge, deploy to AWS, access brokers or credentials, change trading flags, send notifications, create approvals/preflight records, or place orders. Leave parent issues open when real customer or runtime evidence is still missing. DONE MEANS Stop when the actionable AFK backlog is exhausted, the configured budget is reached, or the remaining work needs a human, unavailable access, or unresolved dependencies. Do not invent work to extend the run. Return the PR list, verified commands/results, dependency-aware review order, failed or skipped checks, and exact remaining HITL work.