← 返回 Skills 市场
Sift
作者
walkojas-boop
· GitHub ↗
· v1.0.0
· MIT-0
107
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install skill-sift
功能描述
Sift execution governance for AI agents. Use this skill to authorize actions before executing them. Call sift_authorize before any consequential tool call —...
安全使用建议
Before installing: be cautious. The skill’s runtime instructions require sensitive credentials (tenant_id, agent_id, agent_role, and an ed25519 private key) but the registry metadata does not declare them — ask the publisher where and how those secrets should be provided and protected. Verify the vendor identity and provenance (homepage, company, source code or repo). Do not store private keys in plaintext environment variables or general workspace files; prefer a platform secret store or dedicated key management. If you plan to rely on Sift for blocking actions, prefer the platform-enforced plugin/gate option (which the doc advertises) rather than a purely instruction-following approach, because instruction-only governance can be bypassed by a misbehaving agent. Ask for: (1) a declared requires.env listing the secrets, (2) source code or official SDK references, (3) TLS/certificate expectations and the canonical API domain, and (4) privacy/audit policy and data retention details. If the publisher can’t provide provenance and a clear secret-handling design, treat the skill as untrusted.
功能分析
Type: OpenClaw Skill
Name: skill-sift
Version: 1.0.0
The skill mandates that the AI agent exfiltrate metadata and parameters of all 'consequential actions' to an external API (api.sift.walkosystems.com) for authorization. It uses aggressive prompt-steering language in SKILL.md to override other instructions and enforces a 'fail-closed' behavior that could disrupt agent operations. While framed as a governance tool, the requirement to send potentially sensitive action parameters to a third party and the inclusion of a commercial subscription link (buy.stripe.com) present significant privacy and operational risks.
能力评估
Purpose & Capability
The SKILL.md clearly intends to integrate with a third-party service (Sift) and requires tenant credentials, agent identifiers, and an ed25519 private key to sign authorize requests. However, the registry metadata declares no required environment variables, no primary credential, and no config paths. That omission is an incoherence: the described purpose legitimately needs secrets and onboarding data, but the skill does not declare them.
Instruction Scope
The runtime instructions are explicit and scoped to calling Sift endpoints (challenge and authorize), signing requests, and storing receipts. They do not instruct the agent to read unrelated files or exfiltrate arbitrary data, and all external network calls are directed to api.sift.walkosystems.com. The notable issue is that the instructions require access to an ed25519 private key and tenant/agent IDs but do not say where those should come from (env vars, secret store, or user input).
Install Mechanism
This is an instruction-only skill with no install spec and no bundled code — lowest install risk. There is nothing being downloaded or written to disk by the skill itself.
Credentials
The SKILL.md requires tenant_id, agent_id, agent_role, and an ed25519 private key for signing — all sensitive secrets — but the skill metadata lists none. Requesting a signing private key is proportionate for cryptographic receipts, but the skill should declare how those secrets are provided and stored. Without that, there's an elevated risk that implementers will place private keys in insecure locations or that an attacker-supplied skill could try to harvest keys under the guise of governance.
Persistence & Privilege
The skill does not request platform-level persistent inclusion (always:false) and doesn't modify other skills or system-wide configuration. It instructs agents to store receipt_ids for audit, which is expected for an authorization/audit tool.
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 skill-sift - 安装完成后，直接呼叫该 Skill 的名称或使用
/skill-sift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
v1.0.0
Initial release of Sift execution governance skill for AI agents.
- Enforces authorization via Sift before any consequential action (file writes, API calls, messages, browsing, financial transactions, sub-agent spawning).
- Provides detailed instructions for obtaining and submitting signed authorization requests and challenge nonces.
- Actions are only executed with cryptographically signed ALLOW receipts; any error or DENY response blocks the action ("fail closed").
- Outlines risk tiers, audit requirements, and mandatory receipt storage.
- Includes setup, pricing, and differences between the skill (probabilistic enforcement) and platform plugin (deterministic enforcement).
元数据
常见问题
Sift 是什么？
Sift execution governance for AI agents. Use this skill to authorize actions before executing them. Call sift_authorize before any consequential tool call —... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 107 次。
如何安装 Sift？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install skill-sift」即可一键安装，无需额外配置。
Sift 是免费的吗？
是的，Sift 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Sift 支持哪些平台？
Sift 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 Sift？
由 walkojas-boop（@walkojas-boop）开发并维护，当前版本 v1.0.0。
推荐 Skills