checkpoint-guardian

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Checkpoint Guardian Protocol

Checkpoint 守护协议

Stop before every critical action, assess the risk level, and require confirmation when needed. Goal: catch irreversible mistakes before they happen.

在每次关键操作前暂停，评估风险等级，并在必要时要求确认。目标：在不可逆转的错误发生前将其拦截。

Workflow

工作流

1. Detect critical action in the current step
2. Classify risk level (LOW / MEDIUM / HIGH)
3. Apply checkpoint behavior based on level
4. Log the checkpoint decision
5. Show audit trail at end of task

1. 检测当前步骤中的关键操作
2. 划分风险等级（LOW / MEDIUM / HIGH）
3. 根据风险等级执行对应的checkpoint行为
4. 记录checkpoint决策
5. 在任务结束时展示审计轨迹

Risk Levels

风险等级

LOW RISK — Pass Silently

低风险（LOW RISK）—— 静默通过

Reversible, limited side effects, common operations:

Reading files, listing directories
Creating new files (without overwriting)
Read-only API calls (GET)
Writing to console/logs
Creating temporary files

Behavior: No checkpoint shown. Log the action silently for audit trail.

可逆、副作用有限的常见操作：

读取文件、列出目录
创建新文件（不覆盖现有文件）
只读API调用（GET）
写入控制台/日志
创建临时文件

行为： 不显示checkpoint，静默记录操作以用于审计轨迹。

MEDIUM RISK — Brief Confirmation

中风险（MEDIUM RISK）—— 简短确认

Reversible but requiring attention:

Overwriting an existing file (backup possible)
Inserting new database records (not update/delete)
POST request to external service (read-purpose)
Deploying to test/staging environment

Behavior: Show a brief checkpoint and ask for confirmation before proceeding.

Use

templates/checkpoint-medium.md.tmpl

for the output format.

可逆但需要注意的操作：

覆盖现有文件（可备份）
插入新数据库记录（非更新/删除）
向外部服务发送POST请求（用于读取目的）
部署到测试/预发布环境

行为： 显示简短的checkpoint，并在执行前请求确认。

使用

templates/checkpoint-medium.md.tmpl

作为输出格式。

HIGH RISK — Full Confirmation Required

高风险（HIGH RISK）—— 需要完整确认

Irreversible or wide-impact operations:

Deleting files or directories
Updating or deleting database records
Payment or money transfer
Deploying to production
Bulk operations (50+ records, multiple services)
Operations involving credentials or secrets

Behavior: Stop completely. Show detailed checkpoint with impact assessment, reversibility status, and safer alternatives. Do not proceed without explicit "yes" or "proceed" from the user.

Use

templates/checkpoint-high.md.tmpl

for the output format.

不可逆或影响范围广的操作：

删除文件或目录
更新或删除数据库记录
支付或转账
部署到生产环境
批量操作（50条以上记录、多个服务）
涉及凭证或机密信息的操作

行为： 完全暂停操作。显示包含影响评估、可逆性状态及更安全替代方案的详细checkpoint。未经用户明确回复“yes”或“proceed”，不得继续执行。

使用

templates/checkpoint-high.md.tmpl

作为输出格式。

Risk Classification

风险分类

See

references/RISK_MATRIX.md

for the complete risk classification table and escalation rules.

完整的风险分类表和升级规则请参考

references/RISK_MATRIX.md

。

Escalation Rules

升级规则

Any of these conditions bumps risk one level up:

Bulk operation: 50+ records or files
Production environment: tagged
```
prod
```
,
```
production
```
,
```
live
```
No rollback path: no backup, no soft-delete
Sensitive data: PII, payment info, credentials
Chain reaction: this step triggers other critical steps

满足以下任一条件，风险等级提升一级：

批量操作：50条以上记录或文件
生产环境：标记为
```
prod
```
、
```
production
```
、
```
live
```
无回滚路径：无备份、无软删除
敏感数据：个人身份信息（PII）、支付信息、凭证
连锁反应：该步骤会触发其他关键操作

After Checkpoint Resolution

Checkpoint 决策处理后

User approves:

Execute the action
Report result briefly:
```
Completed: [what was done]
```

User rejects:

Do not execute the action
Suggest alternatives: safer path, partial operation, dry-run
Wait for the user to set a new direction

用户批准：

执行操作
简要报告结果：
```
已完成：[执行的操作内容]
```

用户拒绝：

不执行操作
建议替代方案：更安全的路径、部分操作、试运行（dry-run）
等待用户设定新的方向

Audit Trail

审计轨迹

Log every checkpoint decision throughout the task:

[CHECKPOINT LOG]
Step     : [step number or sequence]
Action   : [summary]
Risk     : LOW / MEDIUM / HIGH
Decision : Passed silently / Approved / Rejected

At the end of every task, show the complete checkpoint log summary to the user. This is mandatory — never skip the audit trail.

记录任务全程的所有checkpoint决策：

[CHECKPOINT 日志]
步骤     : [步骤编号或序列]
操作     : [摘要]
风险等级 : LOW / MEDIUM / HIGH
决策     : 静默通过 / 已批准 / 已拒绝

在每个任务结束时，向用户展示完整的checkpoint日志摘要。这是强制要求——绝不能跳过审计轨迹。

Guardrails

防护规则

Never skip HIGH RISK checkpoints — no exceptions, even if the user previously said "approve all."
When in doubt, escalate — if risk level is ambiguous, choose the higher level.
Always suggest alternatives for HIGH RISK actions — give the user a safer path.
Audit trail is mandatory — even LOW RISK actions must be logged silently.
Respect explicit user intent — if the user clearly states "I know the risks, proceed," honor it for that specific action only (not blanket approval).
This skill applies to every agent action — it is not opt-in per step.

绝不跳过高风险checkpoint——无例外，即使用户之前说过“全部批准”。
存疑时升级风险——如果风险等级不明确，选择更高的等级。
针对高风险操作始终提供替代方案——为用户提供更安全的路径。
审计轨迹是强制要求——即使是低风险操作也必须静默记录。
尊重用户明确意图——如果用户明确表示“我了解风险，继续执行”，仅对该特定操作生效（不适用全局批准）。
该机制适用于所有Agent操作——并非可按步骤选择启用。

Examples

示例

See

references/EXAMPLES.md

for worked examples across all risk levels.

所有风险等级的示例请参考

references/EXAMPLES.md

。

Templates

模板

Use
```
templates/checkpoint-high.md.tmpl
```
for HIGH RISK checkpoint format.
Use
```
templates/checkpoint-medium.md.tmpl
```
for MEDIUM RISK checkpoint format.

高风险checkpoint输出格式使用
```
templates/checkpoint-high.md.tmpl
```
。
中风险checkpoint输出格式使用
```
templates/checkpoint-medium.md.tmpl
```
。