---

name: little-steve-agent-guard version: 0.1.4 description: Little Steve Agent Guard: a self-evolving security system for agent skills. Wraps all skill commands with risk assessment, audit logging, approval levels, and continuous rule evolution. / 小史安全卫士：面向 Agent Skill 的自进化安全系统。为所有技能命令提供风险评估、审计日志、分级审批和持续规则进化。 homepage: https://github.com/EchoOfZion/little-steve-agent-guard requires: bins: - jq

Little Steve Agent Guard

A self-evolving security system for agent skills. Wraps all skill command execution with risk assessment, audit logging, tiered approval, and continuous rule learning.

Dependencies

• jq (required) — install via brew install jq or apt install jq

Filesystem Scope

This is a cross-skill security guard. By design, it needs read access to other skills' directories to:

• guard-exec.sh: read target scripts for static risk analysis before execution
• capability-diff.sh: compare a skill's SKILL.md declarations against its actual scripts

The guard does not write to other skills' directories. All writes (audit logs, rules) stay within its own reports/ and rules/ directories.

Bypass & Emergency Procedures

The runbook (docs/runbook.md) documents emergency bypass procedures (circuit-break, manual script execution, log reset). These are human-operator-only actions for when the guard itself malfunctions. The agent must never execute bypass procedures autonomously.

CRITICAL: Execution Rule

ALL skill script executions MUST go through guard-exec.sh. Never call skill scripts directly. Always use:

bash {baseDir}/scripts/guard-exec.sh exec <script-path> [args...]

Example:

bash {baseDir}/scripts/guard-exec.sh exec {workspaceDir}/skills/<other-skill>/scripts/<script>.sh <command> [args...]

Approval Levels

• L1 (low/medium risk): Auto-execute, audit logged
• L2 (dry-run): Preview without executing
• L3 (high risk): Block and prompt user — output warning, wait for user to reply "确认" or "confirm"
• BLOCK (critical): Reject entirely, no execution possible

When guard-exec.sh returns exit code 10 (prompt), present the warning to the user and wait for confirmation. On "确认"/"confirm", re-run with confirm instead of exec.

Agent Command Conventions

1. Execute a skill command (with guard)

bash {baseDir}/scripts/guard-exec.sh exec <script> [args...]

2. Confirm a prompted action (after user approval)

bash {baseDir}/scripts/guard-exec.sh confirm <script> [args...]

3. Preview without executing

bash {baseDir}/scripts/guard-exec.sh dry-run <script> [args...]

4. Quick risk check

bash {baseDir}/scripts/guard-exec.sh check <script> [args...]

5. Run capability consistency check on a skill

bash {baseDir}/scripts/capability-diff.sh check --skill-dir <skill-path>

6. View audit stats

bash {baseDir}/scripts/audit.sh stats

7. Generate weekly security report

bash {baseDir}/scripts/weekly-report.sh generate [days]

8. Manage rules

bash {baseDir}/scripts/promote-rule.sh list
bash {baseDir}/scripts/promote-rule.sh add --rule <name> --pattern <regex> --level <low|medium|high|critical>
bash {baseDir}/scripts/promote-rule.sh promote --rule <name>
bash {baseDir}/scripts/promote-rule.sh demote --rule <name>

9. Test candidate rules against history

bash {baseDir}/scripts/replay-verify.sh test --rule <name>
bash {baseDir}/scripts/replay-verify.sh test-all

Five Core Security Policies (Immutable)

1. Least Privilege — scripts only access their own data directory
2. Credential Protection — no secrets in args, output, or logs
3. Capability Consistency — runtime must match SKILL.md declarations
4. Outbound Control — no undeclared network access
5. High-Risk Confirmation — destructive/critical actions need human approval

Risk Classification

Level	Examples
low	read-only: list, view, status check
medium	single-item mutation: add, update status
high	delete, bulk mutation, file write outside data/
critical	network access, secret exposure, system commands

Data Files

• reports/audit-events.jsonl — audit log (auto-created)
• reports/failure-dataset.json — failure samples for evolution
• rules/active/*.rule — active custom rules
• rules/candidates/*.rule — candidate rules pending promotion

---

小史安全卫士

面向 Agent Skill 的自进化安全系统。为所有技能命令提供风险评估、审计日志、分级审批和持续规则进化。

依赖

• jq（必须）— 通过 brew install jq 或 apt install jq 安装

文件系统范围

这是一个跨技能安全卫士。按设计，它需要读取其他技能目录的权限：

• guard-exec.sh：执行前读取目标脚本做静态风险分析
• capability-diff.sh：对比技能的 SKILL.md 声明与实际脚本行为

卫士不会写入其他技能的目录。所有写入（审计日志、规则）都在自身的 reports/ 和 rules/ 目录内。

绕过与紧急操作

运行手册（docs/runbook.md）记录了紧急绕过操作（熔断、直接执行脚本、日志重置）。这些是仅限人工操作员的紧急措施，用于卫士本身出故障的情况。Agent 绝对不可以自主执行绕过操作。

关键规则：执行约束

所有技能脚本执行必须通过 guard-exec.sh。 不要直接调用技能脚本，始终使用：

bash {baseDir}/scripts/guard-exec.sh exec <脚本路径> [参数...]

审批分级

• L1（低/中风险）：自动执行，记录审计日志
• L2（预览）：只预览不执行
• L3（高风险）：阻断并提示用户——显示警告，等待用户回复"确认"
• 阻断（严重）：直接拒绝，无法执行

当 guard-exec.sh 返回退出码 10（提示）时，向用户展示警告并等待确认。用户回复"确认"后，用 confirm 替代 exec 重新执行。

Agent 执行约定

1. 执行技能命令（带防护）

bash {baseDir}/scripts/guard-exec.sh exec <脚本> [参数...]

2. 确认被提示的操作（用户批准后）

bash {baseDir}/scripts/guard-exec.sh confirm <脚本> [参数...]

3. 预览不执行

bash {baseDir}/scripts/guard-exec.sh dry-run <脚本> [参数...]

4. 快速风险检查

bash {baseDir}/scripts/guard-exec.sh check <脚本> [参数...]

5. 对技能做声明-行为一致性检查

bash {baseDir}/scripts/capability-diff.sh check --skill-dir <技能路径>

6. 查看审计统计

bash {baseDir}/scripts/audit.sh stats

7. 生成周报

bash {baseDir}/scripts/weekly-report.sh generate [天数]

8. 管理规则

bash {baseDir}/scripts/promote-rule.sh list
bash {baseDir}/scripts/promote-rule.sh add --rule <名称> --pattern <正则> --level <low|medium|high|critical>
bash {baseDir}/scripts/promote-rule.sh promote --rule <名称>
bash {baseDir}/scripts/promote-rule.sh demote --rule <名称>

9. 测试候选规则

bash {baseDir}/scripts/replay-verify.sh test --rule <名称>
bash {baseDir}/scripts/replay-verify.sh test-all

五条核心安全策略（不可变）

1. 最小权限 — 脚本只能访问自身数据目录
2. 凭证保护 — 参数、输出、日志中不得出现密钥
3. 能力一致性 — 运行时行为必须与 SKILL.md 声明一致
4. 外发控制 — 不得有未声明的网络访问
5. 高风险确认 — 破坏性/严重操作需人工审批

风险分级

级别	示例
low	只读操作：列表、查看、状态检查
medium	单项变更：新增、更新状态
high	删除、批量变更、数据目录外写文件
critical	网络访问、密钥暴露、系统命令

数据文件

• reports/audit-events.jsonl — 审计日志（自动创建）
• reports/failure-dataset.json — 失败样本（用于进化）
• rules/active/*.rule — 活跃自定义规则
• rules/candidates/*.rule — 候选规则（待晋升）