Vigilance

Evaluate-before-Execute (EBE) guardrail for OpenClaw agents. Issues a mandatory GO / NO-GO decision before any high-stakes tool call. Enforces child-safety p...

This package contains two related but distinct evaluator scripts and inconsistent metadata/paths — verify which eval_engine.py your agent will run and where you must place SENTINEL_CONFIG.md. Before installing: (1) Inspect both eval_engine.py files yourself or run them in an isolated environment; (2) be aware that when LLM judge mode is enabled the exact --data you pass (commands, URLs, booking details) will be sent to external provider APIs if you set ANTHROPIC_API_KEY or OPENAI_API_KEY — avoid including secrets or sensitive personal data in --data; (3) confirm which file paths the agent expects and fix the SKILL.md/paths if needed; (4) consider disabling LLM judgment (use rule-based fallback) if you cannot accept third‑party data transmission; and (5) if you rely on Chain-of-Thought remaining private, ensure stderr is not forwarded to external logs — CoT is printed to stderr by design. These issues look like sloppy packaging and disclosure rather than clearly malicious code, but they merit manual review before trusting the skill in production.

Type: OpenClaw Skill Name: vigilance Version: 1.0.3 The bundle provides two defensive tools for OpenClaw agents: 'agent-sentinel' (a safety guardrail) and 'agent-eval-engine' (a quality scorer). Both tools use rule-based regex checks and LLM-as-judge calls (via Anthropic, OpenAI, or Ollama) to evaluate agent actions and outputs for safety, compliance, and accuracy. The SKILL.md files contain strict instructions to ensure the agent invokes these evaluators before executing high-stakes tools like shell commands or payments. No evidence of data exfiltration, malicious execution, or persistence was found; the code is transparent, well-documented, and aligned with its stated purpose of providing safety and quality gates.