← 返回 Skills 市场
711
总下载
0
收藏
0
当前安装
7
版本数
在 OpenClaw 中安装
/install safety-checks
功能描述
Verify before you trust — model pinning, fallbacks, and runtime safety validation
安全使用建议
This skill is instruction-only and appears coherent with its goal of local safety checks. Before installing or running it: (1) review .openclaw/safety-checks.yaml and .claude/safety-checks.yaml contents so the skill audits the right config, (2) back up any workspace or cache you care about because flags like --clear and --clear-state will modify/delete local files, (3) if you plan to install the recommended auxiliary skills (constraint-engine, context-verifier, failure-memory), inspect those packages separately (they are not bundled here), and (4) confirm your agent runtime exposes only non-sensitive session metadata because the skill reads the runtime's model-version metadata rather than calling external APIs. If you want higher assurance, ask for the exact file-level commands the skill will run (or a code implementation) before enabling destructive flags.
功能分析
Type: OpenClaw Skill
Name: safety-checks
Version: 1.5.2
The skill bundle is classified as benign. The `SKILL.md` documentation explicitly states that all operations are local-only, involve no network calls or external services, and do not access sensitive system resources (e.g., environment variables, other skills' configurations). It clearly defines its limited file system interactions to specific configuration and output paths (`.openclaw/safety-checks.yaml`, `output/safety/*.log`). There is no evidence of prompt injection attempts, obfuscation, or any intent for data exfiltration, malicious execution, or persistence.
能力评估
Purpose & Capability
Name/description (model pinning, fallback, cache, session hygiene) align with the declared requirements: no env vars, only two config paths, and local workspace output. The skill only requires reading local config/metadata and writing to output/safety/, which is proportionate to the stated purpose.
Instruction Scope
SKILL.md confines operations to local checks (reading .openclaw/.claude configs, workspace files, lock files, and agent session metadata) and supports clearing stale entries. This is within scope, but the commands that clear state (--clear, --clear-state) imply file deletion/modification — users should be aware these options will modify or remove local files. Also, the skill reads 'agent session metadata' (model strings exposed by the runtime); ensure your runtime exposes only intended data.
Install Mechanism
Instruction-only skill (no install spec, no code files) — lowest risk for supply-chain install. The README recommends installing auxiliary skills (constraint-engine, context-verifier, failure-memory) via openclaw install, but there is no automated install spec in the manifest. Verify any recommended dependency packages (the leegitw/* skills) before installing them.
Credentials
No environment variables or credentials requested. The only declared required config paths (.openclaw/safety-checks.yaml and .claude/safety-checks.yaml) are directly relevant to its functionality.
Persistence & Privilege
always is false and the skill is user-invocable; it does not request persistent system-level privileges or attempt to modify other skills' configs. It writes results to a declared workspace path (output/safety/), which is consistent with its purpose.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install safety-checks - 安装完成后，直接呼叫该 Skill 的名称或使用
/safety-checks触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v1.5.2
- Clarified that `/sc fallback` now performs a config file audit only: it checks fallback entries exist in your config, but does not make network calls or test real connectivity.
- Updated usage documentation for `/sc fallback` to reflect its new scope and clarify validated config structure.
- Adjusted fallback output examples and logic section to match config-audit behavior.
- No functional logic changes outside fallback validation scope.
v1.5.1
- Tightened workspace and config scope to this skill's files only.
- No longer requires or scans other skills' workspaces or config files.
- Clarified that all safety checks are strictly local (no API or model calls).
- Updated documentation to reflect narrower file paths for session and cache verification.
- Adjusted session check outputs and instructions for improved clarity.
v1.5.0
- Metadata and documentation updated to reflect new operation boundaries; now clarifies that the skill works fully within your agent's trust boundary.
- Removed the `disable-model-invocation` setting from metadata.
- Revised description and data handling sections to clarify that safety verification leverages the configured model but does not rely on external APIs or third-party calls.
- Adjusted skill tags for improved discoverability.
- No changes to sub-commands, arguments, logic, or outputs.
v1.4.0
- Expanded workspace scanning in metadata: now includes `.claude/` directory for cross-session and config checks.
- Documentation clarifies that model version detection reads agent runtime metadata, not calling APIs directly.
- No logic changes to safety-checks skill behavior; update is documentation and config metadata only.
- Version bumped from 1.2.0 to 1.4.0.
v1.2.0
- Added a `metadata.openclaw.requires` section specifying required config files and workspace paths.
- Updated metadata to clarify required paths for configuration and workspace access.
- No changes to usage, arguments, or core functionality.
v1.1.0
**Changelog for safety-checks 1.1.0**
- Clarified and tightened security boundaries: the skill is now strictly "instruction-only" and does not invoke AI models or access external APIs.
- Updated documentation to explicitly limit checks to workspace and config paths declared in metadata.
- Revised session safety logic and examples to remove mention of global environment scanning—now only scans within `.openclaw/`, `.claude/`, and `output/safety/`.
- Enhanced data handling section with more precise explanation of what the skill does and does not access.
- Removed `.claude/settings.json` from config_paths for simplification.
v1.0.0
Initial release summarizing four core runtime safety checks in one skill.
- Combines model version pinning, fallback chain validation, cache staleness detection, and cross-session state checks into a unified suite.
- Provides CLI interface: `/sc <sub-command> [arguments]` with clear, actionable output.
- Supports configuration via standard YAML/JSON files, with sane fallbacks and clear precedence order.
- Outputs detailed check results and remediation advice for common safety risks.
- Works standalone or with constraint-engine for integrated enforcement.
- No external data calls outside your agent’s configured model runtime.
元数据
常见问题
Safety Checks 是什么？
Verify before you trust — model pinning, fallbacks, and runtime safety validation. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 711 次。
如何安装 Safety Checks？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install safety-checks」即可一键安装，无需额外配置。
Safety Checks 是免费的吗？
是的，Safety Checks 完全免费（开源免费），可自由下载、安装和使用。
Safety Checks 支持哪些平台？
Safety Checks 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Safety Checks？
由 Lee Brown（@leegitw）开发并维护，当前版本 v1.5.2。
推荐 Skills