Agent Workflow Enforcer

让 AI Agent 的执行流程稳定可控。 通过 Gate 门禁、强制输出格式、Style Context 持久化， 把"建议"变成"必须"，解决 Agent 选择性执行的问题。

Security Scan

The skill's purpose and the included gate script are coherent and low-risk, but the runtime instructions reference additional scripts and files that are not present and the source/homepage look like placeholders — these inconsistencies merit caution before installing.

Purpose & Capability

Name/description (enforcing agent workflows with gates, output format, and style persistence) match the included gate.py and the SKILL.md instructions. No unrelated credentials, binaries, or installs are requested.

Instruction Scope

SKILL.md instructs the agent to run several scripts (gate.py, create_style_context.py, detect_learning.py) and to add enforced blocks to AGENTS.md/system prompts. Only gate.py is present in the package; create_style_context.py and detect_learning.py are referenced but not included. The instructions also ask agents to read/write local files (style-context.yaml, learnings.jsonl) and to modify AGENTS.md/system prompts — these actions are expected for the stated purpose but you should confirm the missing scripts and file-write behavior before use.

Install Mechanism

No install spec is provided and the skill is instruction-first with one small Python script. Nothing is downloaded or installed by the package itself.

Credentials

The skill declares no required environment variables, credentials, or config paths. The gate.py script only prints checklists and does not read secrets or network resources.

Persistence & Privilege

Flags show always=false and the skill is user-invocable; it does not request permanent agent presence. The SKILL.md recommends editing AGENTS.md / system prompts, which is a user action and not a stealthy privilege request.

What to consider before installing

This skill appears to do what it says: gate checks and enforcing output-format conventions. However: - SKILL.md references create_style_context.py and detect_learning.py but only scripts/gate.py is included. Ask the publisher for the missing scripts or inspect them before running anything that would create or learn from files. - The homepage/source points to an example GitHub URL (github.com/example/...), and the package author is anonymous; prefer code from a real, trusted repo or request provenance. - The gate script itself is harmless (prints checklists) but the missing scripts could perform file writes or network activity — validate their contents and run in a sandbox first. Practical next steps before installing/using: 1. Request the full repository or the missing script files and review them for file I/O and network calls. 2. Test locally in an isolated environment (or container) to observe any file writes (style-context.yaml, learnings.jsonl) and ensure no unexpected network requests occur. 3. If you will add the suggested snippets to AGENTS.md or system prompts, do so intentionally and keep backups of existing prompts/configs. If you can't obtain the missing scripts and provenance, treat the package as incomplete and avoid running anything that would modify agent/system files.

核心原则：用代码控制流程，不靠 AI 记忆。

为什么需要这个？

AI Agent 会"忘"规则：

• Context 压缩后，早期指令被淡化
• 写在 prompt 里的规则是"建议"，Agent 可以不听
• 分段任务时风格不一致

解决方案：把规则写进流程，没过检查点 = 不能继续。

---

机制一：Gate 门禁

任务开始前，强制运行检查脚本：

python3 {baseDir}/scripts/gate.py <task_type> [options]

支持的任务类型

类型	说明
content	内容创作（文章、推文、配图）
code	代码修改
deploy	部署操作
custom	自定义（读取 gate-config.yaml）

示例

# 内容创作门禁
python3 {baseDir}/scripts/gate.py content --platform x --account your-account

# 代码修改门禁
python3 {baseDir}/scripts/gate.py code --scope "核心文件"

输出

═══════════════════════════════════════
🚧 Gate Check: content
═══════════════════════════════════════

📋 Checklist (必须全部确认):
[ ] 1. 已读相关 SKILL.md
[ ] 2. 已确认目标账号/平台
[ ] 3. 已检查 style-context（如适用）

⚠️ 没有完成以上检查 = 不能继续执行
═══════════════════════════════════════

规则：Gate 输出后，Agent 必须逐项确认才能继续。

---

机制二：强制输出格式

在 AGENTS.md 或 system prompt 中添加：

## 强制输出格式

### 内容创作任务

第一条回复必须包含：

📋 Content Checklist
□ 已跑 gate
□ 已读相关 skill
□ 已确认账号: [账号]

**没有这个块 = 不能开始任务。**

### 任务完成时

必须包含：

📋 Pre-publish Checklist
□ 格式检查: ✅
□ 风格检查: ✅
□ 输出位置: [路径/链接]

**任何一项 ❌ = 不能发布。**

原理：Agent 可能忘了读文件，但格式要求写死在流程里，它不会忘输出格式。

---

机制三：Style Context 持久化

分段任务时，第一批完成后创建 style-context.yaml：

python3 {baseDir}/scripts/create_style_context.py \
  --project "项目名" \
  --style "风格名" \
  --output ./style-context.yaml

模板

# style-context.yaml
project: project-name
created: 2026-02-25

style:
  name: style-name
  description: "风格描述"
  
  colors:
    primary: "#XXXXXX"
    background: "#XXXXXX"
    
  elements:
    - "元素1"
    - "元素2"
    
  avoid:
    - "避免1"
    - "避免2"

prompt_template: |
  [完整的 prompt 模板]

existing_outputs:
  file1.png: "描述"
  file2.md: "描述"

后续批次规则

在 AGENTS.md 中添加：

## 分段任务规则

检测到分段任务时：
1. 先检查 style-context.yaml 是否存在
2. **存在** → 读取，按记录的风格继续
3. **不存在** → 不能继续，先创建

**没有 style-context = 不能继续分段任务。**

---

机制四：自动学习（可选）

当用户修改 Agent 输出时，检测差异并学习：

python3 {baseDir}/scripts/detect_learning.py "原文" "修改后"

输出：

{
  "detected": true,
  "original": "进行分析",
  "corrected": "分析",
  "suggested_rule": "避免「进行+动词」冗余结构"
}

确认后写入 learnings.jsonl，下次自动应用。

---

快速开始

1. 安装

# OpenClaw
npx clawhub@latest install agent-workflow-enforcer

# 手动
git clone https://github.com/example/agent-workflow-enforcer ~/skills/agent-workflow-enforcer

2. 配置 AGENTS.md

把以下内容加到你的 AGENTS.md：

## 🚧 Workflow Enforcer

### Gate 门禁
任务开始前运行：`python3 ~/skills/agent-workflow-enforcer/scripts/gate.py <type>`

### 强制输出格式
- 任务开始：必须输出 Checklist 块
- 任务完成：必须输出 Pre-publish 块
- **没有格式块 = 任务不完整**

### 分段任务
- 有 style-context.yaml → 读取后继续
- 没有 → 先创建

3. 测试

python3 ~/skills/agent-workflow-enforcer/scripts/gate.py content --platform x

---

文件结构

agent-workflow-enforcer/
├── SKILL.md              # 本文件
├── scripts/
│   ├── gate.py           # 门禁脚本
│   ├── create_style_context.py
│   └── detect_learning.py
├── templates/
│   ├── style-context.yaml
│   ├── gate-config.yaml
│   └── agents-snippet.md  # AGENTS.md 代码片段
└── examples/
    └── content-workflow/

---

核心原则

规则写在 prompt 里 = 建议 Agent 可以不听。

规则写在输出格式里 = 强制 没有格式块 = 回复不完整。

规则写在 gate 脚本里 = 硬门禁 没跑脚本 = 后续不执行。

不是靠 AI "记住"，是靠流程卡死。

---

License

MIT