← 返回 Skills 市场
74
总下载
0
收藏
0
当前安装
2
版本数
在 OpenClaw 中安装
/install persona-evaluator
功能描述
Audit any OpenPersona (or peer LLM-agent) persona in three complementary modes: structural (CLI, deterministic, CI-friendly: 4 Layers × 5 Systemic Concepts ×...
安全使用建议
This skill is coherent with its stated purpose: structural audits use the OpenPersona CLI and semantic modes either read a provided pack-content JSON or perform consented A2A probes. Before installing or using it, confirm your agent host has the OpenPersona CLI available (npx openpersona >= 0.2.0) or be prepared to install it. When using black-box mode, only run probes where you have permission — the skill's protocol relies on voluntary responses and explicitly caps confidence/score based on consent and data source. The skill does not ask for secrets or perform disk writes per its rules, but verify the host environment enforces those constraints (e.g., the agent runtime actually prevents silent disk writes or unexpected network relays). If you need higher assurance, check the upstream repository (metadata references acnlabs/OpenPersona) and review the published OpenPersona CLI code before granting the agent network or A2A messaging privileges.
功能分析
Type: OpenClaw Skill
Name: persona-evaluator
Version: 0.3.4
The persona-evaluator skill is a comprehensive auditing tool for LLM personas within the OpenPersona framework. It implements a structured '4+5' evaluation methodology across structural, white-box, and black-box modes. The bundle is well-documented, with clear safety guidelines in BLACK-BOX.md (e.g., explicit consent requirements and prohibitions against harmful probes) and a transparent development history in CHANGELOG.md. No evidence of malicious intent, data exfiltration, or unauthorized command execution was found; the allowed tools are strictly limited to the 'npx openpersona' CLI.
能力标签
能力评估
Purpose & Capability
Name/description match the instructions: structural mode uses `npx openpersona evaluate`, white-box requires `--pack-content` (reads persona JSON), black-box uses A2A probe/consent. The CLI and A2A messaging are appropriate capabilities for an audit tool; nothing in the files asks for unrelated cloud credentials or system-wide access.
Instruction Scope
SKILL.md explicitly limits actions (structural: run CLI; white-box: read pack-content JSON only; black-box: send consented probes or observe public material). It prescribes consent rules, caps on confidence, and 'never write back to disk' for semantic mode. Note: black-box mode involves active messaging to remote agents (A2A / consent probes) which can collect voluntarily disclosed content — this is expected but a privacy consideration for users and subjects.
Install Mechanism
This is instruction-only with no install spec and no code files to execute. It relies on the existing OpenPersona CLI (npx openpersona), which is proportional to the stated structural features. No downloads or extracted archives are requested in the skill bundle.
Credentials
The skill declares no required environment variables, credentials, or config paths. The instructions reference agent-card / ACN auth as a trust model for A2A but do not request host secrets or tokens. The lack of credentials is appropriate for a reviewer/auditor skill.
Persistence & Privilege
Flags show always:false (no forced inclusion) and normal autonomous invocation allowed (platform default). The skill does not request permanent presence or to modify other skills' configs. Autonomous invocation + messaging is normal for an agent skill; no additional privilege escalation is requested.
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 persona-evaluator - 安装完成后，直接呼叫该 Skill 的名称或使用
/persona-evaluator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
v0.3.4
v0.3.4 docs-only patch: trim off-target darwin-skill comparisons (What/When intro + Relationship table); refresh stale 'will be published' standalone-distributable note in Install section; tighten body (drop single-point Black-box backref + fold Re-evaluate sub-section into Apply-fixes-via-refine). SKILL.md 282->274 lines. No functional change to CLI/scoring/rubric/tests. Rationale for off-target comparisons in docs/SKILL-RUBRIC.md v0.1.5.
v0.3.3
Initial publish of persona-evaluator v0.3.3.
Audits any OpenPersona persona pack across 9 dimensions in three complementary modes:
- Structural CLI (deterministic, CI-friendly): 4 Layers x 5 Systemic Concepts + Constitution gate, role-aware severity.
- Semantic white-box (LLM rubric scoring of --pack-content JSON).
- Semantic black-box (A2A handshake / 10-probe protocol for remote agents you cannot read on disk; Tier 1/2/3 confidence caps).
v0.3.3 includes a critical schema-bifurcation fix (W6): evaluator now reads the flat on-disk persona schema correctly via a nested-first / flat-fallback helper, mirroring lib/lifecycle/refine.js. Both authored personas (entrepreneur-skill) and generator-pipeline output (persona-secondme-skill) re-validate at 7/10 Good.
Test surface: 879 tests pass, 8/8 skills pass spec, 7 W6 regression tests added with parity check between flat and nested fixtures.
Source: https://github.com/acnlabs/persona-evaluator (also bundled in github.com/acnlabs/OpenPersona main repo).
元数据
常见问题
persona-evaluator 是什么？
Audit any OpenPersona (or peer LLM-agent) persona in three complementary modes: structural (CLI, deterministic, CI-friendly: 4 Layers × 5 Systemic Concepts ×... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 74 次。
如何安装 persona-evaluator？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install persona-evaluator」即可一键安装，无需额外配置。
persona-evaluator 是免费的吗？
是的，persona-evaluator 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
persona-evaluator 支持哪些平台？
persona-evaluator 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 persona-evaluator？
由 acnlabs（@neiljo-gy）开发并维护，当前版本 v0.3.4。
推荐 Skills