본문으로 건너뛰기
TakoAPI
스킬로 돌아가기

policy-engine

joetomasone 제작

Coding Agents & IDEs
—아직 평가 없음
내 평가

요약

Deterministic governance layer for OpenClaw tool execution.

설명

이 스킬이 하는 일

A governance plugin for OpenClaw that intercepts tool calls before execution to enforce allowlists, deny patterns, path restrictions, and risk tiers. Every decision is logged for audit. Ships with dry-run mode, escalation tracking, and hot-reload support.

Unlike ad-hoc prompt-based restrictions, it enforces policies deterministically at the hook level so prompt injection or model drift can't bypass them.

사용 시점

  • Limiting a research sub-agent to read-only tools
  • Blocking dangerous exec commands like rm -rf or docker push
  • Enforcing write paths so agents can't touch files outside a workspace
  • Testing policy rules in dry-run before enforcing them in production
  • Auditing which tools each agent attempted to use across a session

예시 워크플로

입력

User wants to restrict a research agent to read-only tools and block it from writing files

에이전트
  1. 1

    Define an allowlist profile named 'readonly' with [read, web_fetch, web_search, message]

  2. 2

    Map the research agent ID to the 'readonly' profile via the routing config

  3. 3

    Enable dry-run mode to verify what would be blocked without enforcing

  4. 4

    Check logs for DRYRUN entries, confirm the policy behaves as expected

  5. 5

    Disable dry-run to enforce blocking in production

출력

Research agent is restricted to read-only tools; any write or exec attempt is blocked and logged with a remediation message

설치

clawhub install joetomasone/policy-engine

CLI 명령을 실행하거나, 프롬프트를 Claude Code, Cursor, Codex 등 사용하는 코딩 에이전트에 붙여넣으세요.

링크