A
规则，真的接到了动作上吗？Are rules connected to actions?
它沿着 Skill、Hook、任务合同和安全门找执行证据。只有提示词或 README，没有可执行入口，不算受控。It follows Skills, hooks, task contracts, and safety gates for execution evidence. A prompt or README without an executable entry is not control.
给已有 Agent Harness 的开源验收 SkillAn open-source verification skill for existing Agent Harnesses
别只听 Agent 说“做完了”。Do not just trustan Agent saying “done.”
Harness Check 是一层独立验收：它检查你已有的规则是否真的接入执行、检查能否在本地和 CI 重跑，并把还需要真人确认的地方明确留出来。Harness Check is an independent verification layer. It checks that your rules are truly wired to execution, can run again locally and in CI, and clearly leaves the work that still needs a person.
它不搭建 Runtime，不给模型打分，也不把人工判断伪装成自动化。It does not build a runtime, grade models, or disguise human judgement as automation.
01“写了规则”，不等于规则真的能拦住错误。A written rule does not mean it can stop an error.
02“本机绿了”，不等于 CI 和下一次也能复现。“Green locally” does not mean CI or the next run can reproduce it.
03“脚本跑过”，不等于真实宿主和人工验收已经完成。“A script ran” does not mean real-host and human review are complete.
它到底检查什么What it actually checks
它不检查“结果好不好”。
它检查“工作流是否可信”。It does not judge the output.
It checks whether the workflow is trustworthy.
B
失败时，能定位也能重跑吗？Can failure be located and rerun?
它按工作区、暂存区和 CI 选择唯一入口，保留正反例、真实退出码和失败标签，不把一句“已完成”当报告。It chooses one entry for the working tree, staged changes, or CI, preserving fixtures, real exit codes, and failure labels instead of a vague “done”.
C
哪些事仍然必须由人确认？What still needs a person?
真实宿主触发、外部副作用、视觉与发布判断会单独列出。没有证据，就写“待补”，绝不伪装成通过。Real-host triggers, external side effects, visual review, and release calls stay separate. Without evidence, they remain “pending”, never passed.
你最后拿到什么What you get at the end
不是“通过”。
是一份能交接的判断。Not just “passed.”
A decision someone else can take over.
报告会明确写出：在哪个模式运行、自动化通过还是失败、失败怎样复现、宿主烟测是否做过、人工验收还缺什么。它不会把这些压成一个看似权威的总分。The report states the run mode, automated pass or failure, how to reproduce it, whether host smoke tests happened, and what human review remains. It does not compress these into a pretend-authoritative score.
查看检查范围与结论口径Read the check matrix
怎么开始How to start
先验证一份已有的 Harness。Verify one Harness you already have.
在项目根目录运行这条命令。前提是项目已经有自己的 Harness 验证入口、测试和规则真源；Harness Check 不会替你凭空生成这些东西。Run this at your project root. The project must already have its own Harness verification entry, tests, and rule sources; Harness Check will not invent them for you.
bash .harness/scripts/verify.sh --working-tree
还没有 Harness？先读 中文 Guide 建一个最小闭环；已有 Harness，再按仓库 README 安装这个 Skill。No Harness yet? Read the guide to build a minimal loop. If you have one, install this skill from the repository README.
它适合谁，不适合谁Who it is — and is not — for
你已经在维护 Agent 工作流，
才会真正需要它。It earns its place when
you maintain an Agent workflow.
适合：已有 Harness，开始需要可信交付For: an existing Harness that needs trustworthy delivery
- 你已有规则、Hook、测试或 CI，却无法确认它们有没有一起工作You have rules, hooks, tests, or CI but cannot tell if they work together
- 你需要区分本机、暂存区、CI 和真实宿主的结论You need distinct conclusions for local work, staged changes, CI, and real hosts
- 你希望失败能定位，完成能交接，而不是依赖 Agent 自报You want failures located and completion handed off, not self-reported by an Agent
不替代：搭建 Runtime、写规则或人的判断Not a replacement for a runtime, your rules, or people
- 它不是 DeepSeek Harness 一类的 Agent Runtime，不负责把模型和工具跑起来It is not an Agent Runtime such as DeepSeek Harness; it does not run models and tools
- 它不会替项目写业务规则、测试或唯一验证入口It does not write your domain rules, tests, or canonical verification entry
- 它不会宣称真实宿主、视觉、内容质量或外部副作用已经由机器定论It never claims that real hosts, visual quality, content quality, or external effects are machine-settled