beastmode-claude-codelisted

# Beastmode Claude Code — Claude Code Harness Adapter This skill is **only the harness mechanics**. The framework, tier-routing rule, verifier-first design principle, and self-improvement loop live in the canonical `beastmode` skill (v2.3.0) — load it first and follow it. ## Only the harness mechanics | Universal seat | Claude Code implementation | |---|---| | Director (frontier) | The Claude Code session itself — frontier, e.g. opus or fable | | Watcher (frontier) | A second frontier session, prefer cross-family via an external lane | | Executor (economy) | Task subagents, or parallel `claude -p --model <economy>` lanes | | Loop engine | Director self-prompts between subagent results | | Anti-spin | Director judgment + usage budget; abort on 3 identical blockers | | Worker-contract enforcer | `--permission-mode` plus a starter worker contract in the prompt | | Live progress | Tail Task subagent transcripts or tmux panes running `claude -p` | | Goal / self-improvement log | `<repo>/GOAL.md` and `<repo>/.learnings/BEASTMODE.md` (universal) | ## ACN — async parallel sub-agents ACN is the universal Beastmode fan-out layer. In Claude Code it is the Task tool plus parallel `claude -p` invocations. Task spawns an in-session subagent (multiple Task calls in one turn run sequentially). `/batch` fans worktree-isolated tasks out across separate sessions for large parallel changes; each task gets its own branch and worktree. Multiple background tmux / `claude -p` processes are the