shield_lock
金大哥
← 返回 Skill 列表
extension
分类: 开发与工程无需 API Key

harness-engineer

一个面向代理优先开发的完整工程工具包——搭建框架、诊断故障、清理过时文档、以及初始化新会话。当用户说：'设置一个框架', '我的代理卡住了', '代理循环中断了', '清理框架', '开始一个新的会话', '这个项目的框架', '代理不断失败', '断路器', '清理我的文档', 或者 '代理弄得一团糟' 时使用。

person作者: jakexiaohubgithub
open_in_new仓库download下载资源包

Harness Engineer Plugin

A complete harness engineering toolkit synthesized from OpenAI, Anthropic, and LangChain's 2026 research. Prevents stuck loops, orients fresh agents, repairs failures, and fights entropy.

Plugin Map

| Skill | Command | Trigger | |---|---|---| | harness-init | /harness-engineer:init | New project, new repo, needs harness scaffold | | harness-doctor | /harness-engineer:doctor | Agent failed, loop broke, session went wrong | | harness-gc | /harness-engineer:gc | Docs stale, constraints violated, features inconsistent | | harness-onboard | Auto (SessionStart hook) | Every new Claude Code session in a harness project | | harness-engineering | /harness-engineer:theory | Theory, CLASS metrics, Constitutional AI, evaluation design |

Orchestrator Agent

For multi-stage workflows, the orchestrator (agents/orchestrator.md) classifies requests, sequences skills, and synthesizes outputs. Three operation modes:

  • Guided (default) — single command, no orchestration
  • Supervised — orchestrator recommends, user confirms
  • Autonomous — orchestrator diagnoses + executes with circuit breaker safety rails

Circuit Breaker System

Three deterministic hooks wired via .claude/settings.json:

| Hook | File | Fires | |---|---|---| | Circuit Breaker | hooks/circuit-breaker.sh | PostToolUse Write/Edit — hard stash at 10 edits | | Dead Man's Switch | hooks/dead-mans-switch.sh | PostToolUse * — auto-stash at 30min no-commit | | Watchdog | hooks/watchdog.sh | PostToolUse * every 10 calls — env health + lock cleanup | | Pre-Completion | settings.json Stop hook | Before agent exits — forced verification |

Quick Start

# Install into any project
bash .claude/plugins/harness-engineer/install.sh

# Scaffold a new harness
/harness-engineer:init

# After an agent failure
/harness-engineer:doctor

# Weekly maintenance
/harness-engineer:gc

Examples

Input: "My Claude Code agent went into a loop and corrupted three files before I stopped it."

What happens: Run /harness-engineer:doctor — the skill analyzes the last 20 tool calls, identifies the loop trigger, proposes harness constraints to prevent recurrence, and outputs a CONSTRAINTS.md patch ready to apply.

Input: "Starting a new project and want to set up a proper agent harness from scratch."

What happens: Run /harness-engineer:init — scaffolds HARNESS.md, PROGRESS.md, FEATURES.md, CONSTRAINTS.md, and wires the three-layer circuit breaker hooks into .claude/settings.json.

Input: "My harness docs are stale — constraints reference files that no longer exist."

What happens: Run /harness-engineer:gc — audits all harness docs against the current codebase, removes stale references, updates feature states, and outputs a clean diff.

Research Foundation

  • OpenAI (Feb 13 2026): Context engineering, architectural constraints, garbage collection agents
  • Anthropic (Nov 26 2025): Dual-agent architecture, progress files, JSON feature lists, git checkpoints
  • LangChain (Feb 17 2026): LoopDetectionMiddleware, PreCompletionChecklist, reasoning sandwich, trace analysis

Read individual skill SKILL.md files for deep implementation details.

References

  • PLUGIN-README.md — Full installation guide, hook wiring, and advanced configuration