Agent harness 架构设计模式知识库。当设计多阶段任务的上下文传递、选择多 agent 协调方式、或规划 hook 系统架构时使用。纯设计指南，无可执行代码。不用于安装 hook 脚本（用 agent-hooks）或运维监控（用 agent-ops）。

This skill is a documentation-only knowledgebase about harness design — it doesn't ask for credentials or install code itself. Before adopting its patterns: 1) recognize the guide instructs creating hooks and shell scripts that will read/write session files (sessions/*, TMPDIR, ~/.openclaw/shared-context) and use common shell tools (bash, jq, grep); ensure those paths/tools are acceptable in your environment. 2) The SKILL.md references env vars (HARNESS_PROFILE, HARNESS_DISABLED_HOOKS) — these are not secrets but are expected to be set by the implementer. 3) If you implement the examples, review any hook scripts for correct file paths, locking behavior, and to avoid accidentally exposing sensitive files; test in a sandbox before using in production. 4) The doc cites internal Claude Code behaviors and external links that may change; validate assumptions against your agent runtime. Overall the package is coherent and appropriate as a design reference, but treat code snippets as templates to review rather than ready-to-run artifacts.

Type: OpenClaw Skill Name: harness-design-patterns Version: 1.1.0 The 'harness-design-patterns' bundle is a comprehensive architectural knowledge base for designing AI agent execution environments. It provides well-documented patterns for context survival (handoff documents), multi-agent coordination (Coordinator/Fork/Swarm), and safety mechanisms (Denial Circuit Breakers). While it contains shell script snippets for tasks like atomic file writing (references/06-atomic-write.md) and task complexity triage (references/16-adaptive-complexity.md), these are local, non-malicious utility functions aligned with the stated purpose. No evidence of data exfiltration, unauthorized remote execution, or harmful prompt injection was found; the content focuses on improving agent reliability and observability.