Delegate coding tasks to Claude Code via ACPX from OpenClaw. Runs Claude Code as root with IS_SANDBOX=1 bypass, full file-write permissions, Opus model support, and persistent named sessions. Use when: (1) The user asks to write, review, or refactor code in a project directory (2) The user wants a persistent coding session ("start a session", "continue working on X") (3) Multi-file code generation or editing is needed (4) The user says "use Claude Code", "code this", "write code for", "build", "implement" (5) Complex coding tasks that benefit from Claude Code's tool-use (file read/write/edit, terminal) Not for: simple one-line edits (use the edit tool directly), reading files (use read tool), non-coding tasks, or tasks that don't need file system access.
SuperClaw AutoDream: 记忆整理持久 agent。定期整理 OpenClaw 记忆文件。 触发条件：距上次 >= 24 小时 AND >= 5 个 session。 由 agent-check.sh 自动触发，也可用户手动 invoke。
SuperClaw Phase 3: 执行。通过 superclaw 让 Claude Code 自治执行 plan。 使用场景：plan 已 approve，选定了执行方式。 终态：执行完成 → 自动触发 verify（通过 hook 或轮询）。
SuperClaw Phase 2: 技术对齐。OpenClaw 与 Claude Code 协商，把 spec 变成可执行的 plan。 使用场景：align 阶段完成、spec 已 approve。 终态：plan approved → invoke execute。
Meta-skill that teaches OpenClaw its role and boundaries in the SuperClaw system. Use at the start of any conversation that might involve coding tasks. Establishes: what you are, what you're not, when to activate the workflow, and how to interact with Claude Code through the superclaw bridge.
SuperClaw Phase 4-L2: OpenClaw 独立验收。不信任 Claude Code 的自我声明，独立跑测试和检查。 使用场景：execute 阶段完成，收到执行报告。 终态：验收通过 → invoke deliver；不通过 → 回 execute 修复。
SuperClaw Phase 1: 产品对齐。将用户模糊需求转化为结构化 spec。 使用场景：用户提出开发需求（一句话、一段描述、一个链接都行）。 唯一终态：用户 approve spec → invoke plan。 借鉴 gstack office-hours 的深度对齐方法论，包括 Forcing Questions、 Premise Challenge、Anti-Sycophancy、Spec Review Loop。
SuperClaw Phase 4-L3: 人类最终验收交付。把验证通过的产出交给用户做最终确认。 使用场景：verify 阶段通过。 终态：用户 approve → 完成；用户要调整 → 回 verify/execute；用户丢弃 → 结束。