governance-method-suite

One playbook for the full AI-agent governance loop: red-team drill -> prompt-injection shield -> prompt-leak scan -> document desensitization -> kill-switch -> badge issue/verify -> skill & publish quality gates. Orchestrates 9 independently published sub-skills by slug; does not vendor their internals. Use when the user wants an end-to-end governance plan, a security audit, or a release-compliance review.

凡自治之物 · LGD 治理学 · 创造术家族 ｜ 评测协议：skill-quality-gate

English

What this is

A collection / orchestration skill: one playbook that wires the full AI-agent governance loop end to end. It does not re-implement security logic — it routes you, stage by stage, to 9 sub-skills that are already published and maintained on their own. Treat this as the table of contents; the sub-skills are the chapters.

The flow is:

red-team drill -> injection shield -> leak scan -> desensitization -> kill-switch -> badge issue/verify -> skill gate -> publish gate

Entry index (slug | stage | when to use | call note)

#	Stage	Sub-skill slug	When to use	Call note	Expected output
1	Red-team drill	agent-redteam-kit	Before launch: stress-test the agent adversarially	Invoke the sub-skill; run its drill	Scored attack-surface report + fix list
2	Injection shield	prompt-injection-shield	Hostile user/tool input must be blocked	Invoke the sub-skill on the prompt	Guard allow/block decision
3	Leak scan	prompt-leak-scanner	Before shipping a system prompt: catch internal/secret leakage	Invoke the sub-skill on the prompt	Leak findings (line-level)
4	Desensitization	doc-desens-scanner	Before a doc/code/log goes external	Invoke the sub-skill on the file	Masked copy + audit ledger
5	Kill switch	agent-kill-switch	Before deploying a high-risk/funds agent	Generate then --check the kill card	Complete kill card; blocks if fields missing
6	Badge issue	lgd-badge-issuer	After three-law assessment passes	Invoke the sub-skill to issue	Badge certificate + SHA-256 fingerprint + ledger row
7	Badge verify	lgd-badge-verify	Before trusting an external badge	Invoke the sub-skill on the badge	Verify pass/fail with reasons
8	Skill quality gate	skill-quality-gate	Before admitting a skill to the library	Invoke the sub-skill on the skill dir	Gate verdict + blocking findings
9	Publish gate	publish-quality-gate	Right before external publish	Invoke the sub-skill	Publish verdict; blocks launch on blockers

Per-stage notes

• 1. Red-team first. You can't defend what you haven't attacked. Run agent-redteam-kit early and feed its findings into stages 2–4.
• 2–4. Defense in depth. Shield runtime input (2), scan the prompt itself (3), and desensitize anything that leaves the boundary (4).
• 5. Control plane. Even with good scans, a high-risk agent needs a pre-authored kill card — --check makes completeness a hard gate.
• 6–7. Provenance. Issue a badge only after gates pass; verify someone else's badge before relying on it.
• 8–9. Release gates. Two gates, two purposes: library admission (8) vs. external publish (9). Both must be green.

Demo

bash

# list the whole flow
python3 scripts/governance_flow.py --list

# get the call note for one stage
python3 scripts/governance_flow.py --route redteam

# ...or by sub-skill slug
python3 scripts/governance_flow.py --route lgd-badge-verify

Scope & red lines

• This suite vendors no sub-skill internals. It references sub-skills by slug only; their code stays in their own packages.
• Pure documentation + a read-only router. No network, no business-data writes, no acting on your behalf.
• Sub-skill release status is tracked separately in the governance ledger; this suite does not republish or fork them.

中文

这是什么

一个合集/编排技能：用一套 playbook 把智能体治理的完整闭环串起来。它不重写安全逻辑——而是按环节，把你路由到 9 个已独立发布、各自维护的子技能。本合集是目录，子技能是章节。

治理闭环：

红队演练 -> 注入防护 -> 泄漏扫描 -> 文档脱敏 -> 停机开关 -> 徽章签发/核验 -> 技能质量门 -> 发布门

入口索引（slug | 环节 | 何时用 | 调用说明）

#	环节	子技能 slug	何时用	调用说明	预期输出
1	红队演练	agent-redteam-kit	上线前对抗性压测智能体	调子技能跑演练	攻击面评分报告 + 修复清单
2	注入防护	prompt-injection-shield	需拦截恶意用户/工具输入	对 prompt 调子技能	放行/拦截判定
3	泄漏扫描	prompt-leak-scanner	发系统提示词前查内部/密钥泄漏	对 prompt 调子技能	泄漏发现（行级）
4	文档脱敏	doc-desens-scanner	文档/代码/日志对外前	对文件调子技能	脱敏版 + 留痕清单
5	停机开关	agent-kill-switch	高危/资金类智能体部署前	先生成再 --check 制动卡	完整制动卡；缺项即阻断
6	徽章签发	lgd-badge-issuer	三律考核通过后申请徽章	调子技能签发	徽章证书 + SHA-256 指纹 + 台账行
7	徽章核验	lgd-badge-verify	信任外部徽章前	对徽章调子技能	通过/不通过 + 原因
8	技能质量门	skill-quality-gate	技能入库存准入前	对技能目录调子技能	门禁结论 + 阻断项
9	发布门	publish-quality-gate	对外发布前最后一道	调子技能	发布结论；有阻断项即不许上线

环节说明

• 1. 先红队。 没攻过就谈不上防。尽早跑 agent-redteam-kit，把发现喂给 2–4。
• 2–4. 纵深防御。 运行时拦注入（2）、扫提示词本身（3）、对出边界的东西脱敏（4）。
• 5. 控制面。 即使扫描到位，高危智能体也要预先写好制动卡——--check 把字段齐全做成硬门禁。
• 6–7. 凭证。 门禁全过才发徽章；信任别人的徽章前先核验。
• 8–9. 发布门。 两道门两种用途：入库准入（8）与对外发布（9），两道都要绿。

示例

bash

# 列出全流程
python3 scripts/governance_flow.py --list

# 查某个环节怎么调
python3 scripts/governance_flow.py --route redteam

# 也可按子技能 slug 查
python3 scripts/governance_flow.py --route lgd-badge-verify

边界与红线

• 本合集不内嵌任何子技能内部实现，仅按 slug 引用；子包代码留在各自包内，不改不复制。
• 纯文档 + 只读路由脚本：无网络、不写业务数据、不代你执行任何动作。
• 各子技能的发布状态在治理台账单独跟踪；本合集不重发、不 fork 它们。