novera-skills

Novera Skills — Trust layer for AI agents

Deterministic, offline Agent Skills for auditing the trustworthiness of AI-produced content and claims. Built on the open Agent Skills standard (SKILL.md + agents/openai.yaml), compatible across clients, zero third-party dependencies, pure Python standard library.

Status: early and small on purpose. Five skills are released; more are on a public roadmap. Code over claims.

What is this

Each skill in this repository audits one class of claims that AI-generated assets tend to assert — from geographic claims and directory materials to skill packages and e-commerce briefs. Everything runs offline with deterministic scripts: same input, byte-identical report. Findings always carry file:line locations, and every judgment that depends on external facts is labeled with an evidence strength instead of being asserted.

Skills

Roadmap

Quick start (geo-evidence-audit)

No installation. Python 3.9+ required, nothing else.

# Text report
python skills/geo-evidence-audit/scripts/audit.py path/to/page-or-dir

# JSON report
python skills/geo-evidence-audit/scripts/audit.py path/to/page-or-dir --format json

# Only contradictions, drop unverified claims
python skills/geo-evidence-audit/scripts/audit.py path/to/page.md --min-strength explicit

Exit codes: 0 no findings · 1 findings present (contradictions or unverified claims) · 2 usage or input error.

As an Agent Skill, point your client at skills/geo-evidence-audit/ — SKILL.md carries the trigger conditions, agents/openai.yaml the cross-client description.

Repository layout

novera-skills/
├── skills/<slug>/        # one self-contained skill per directory:
│   ├── SKILL.md          #   triggers, workflow, boundaries (frontmatter: name + description only)
│   ├── agents/openai.yaml
│   ├── scripts/          #   deterministic stdlib-only tooling, entry audit.py
│   ├── references/       #   data provenance (URL + retrieval date + evidence strength), rules, evidence model, scope
│   └── fixtures/         #   audit test fixtures with expected exit codes
├── tests/<slug>/         # pytest suite driving the real CLI
├── tools/py39_gate.py    # CI syntax gate (ast feature_version=(3,9))
└── docs/                 # internal design docs (Chinese, not part of any release)

Skills directories intentionally contain no README, install guide, or changelog — everything consumer-facing lives here.

Per-skill publishing

Each skill is versioned and released independently with a namespaced tag:

<slug>-v<semver>     e.g. geo-evidence-audit-v0.1.0

MAJOR = breaking change to triggers / CLI / output schema · MINOR = new signals, rules, or report fields · PATCH = false-positive/negative fixes, docs, fixtures. Bare v* tags and floating latest are never used. To consume a single skill, fetch the tag and take its skills/<slug>/ directory.

Constraints

• Pure Python standard library, zero third-party dependencies, Python 3.9+ (CI-enforced syntax gate)
• Fully offline: audit scripts never make network requests
• Deterministic: same input produces byte-identical output
• Findings always carry file:line; evidence strength is labeled, not asserted
• No performance claims anywhere: numbers only after they are measured
• CI gates: py39 syntax, stdlib-only import scan, SKILL.md < 500 lines, full pytest matrix (ubuntu/windows × 3.9/3.12)

License

MIT

Novera Skills — AI 智能体的信任层

确定性的、完全离线的 Agent Skills，用于审计 AI 产出内容与声称的可信度。基于开放 Agent Skills 标准（SKILL.md + agents/openai.yaml），跨客户端兼容，零第三方依赖，纯 Python 标准库实现。

状态：刻意保持早期与小体量。已发布五款，更多在公开路线图上。代码先于声称。

这是什么

本仓库的每款 skill 审计一类 AI 生成资产容易出错作假的声称——从地理声称、目录站材料到 skill 包与电商 brief。全部检测由离线的确定性脚本完成：相同输入，逐字节相同的报告。每条 finding 必带 file:line 定位；任何依赖外部事实的判断都标注证据强度，而不是直接断言。

技能列表

路线图

快速开始（geo-evidence-audit）

无需安装。仅需 Python 3.9+。

# 文本报告
python skills/geo-evidence-audit/scripts/audit.py path/to/page-or-dir

# JSON 报告
python skills/geo-evidence-audit/scripts/audit.py path/to/page-or-dir --format json

# 只要矛盾类 finding，过滤 unverified 声称
python skills/geo-evidence-audit/scripts/audit.py path/to/page.md --min-strength explicit

退出码：0 无 finding · 1 有 finding（矛盾或 unverified 声称）· 2 用法或输入错误。

作为 Agent Skill 使用时，将客户端指向 skills/geo-evidence-audit/——SKILL.md 载明触发条件，agents/openai.yaml 提供跨客户端描述。

仓库布局

novera-skills/
├── skills/<slug>/        # 每款 skill 自成一体的目录：
│   ├── SKILL.md          #   触发条件、工作流、边界（frontmatter 仅 name + description）
│   ├── agents/openai.yaml
│   ├── scripts/          #   确定性纯标准库工具，入口 audit.py
│   ├── references/       #   数据来源（URL + 抓取日期 + 证据强度）、规则、证据模型、分工
│   └── fixtures/         #   审计测试夹具与预期退出码
├── tests/<slug>/         # 驱动真实 CLI 的 pytest 套件
├── tools/py39_gate.py    # CI 语法门禁（ast feature_version=(3,9)）
└── docs/                 # 内部设计文档（中文，不进任何发布物）

skill 目录内刻意不放 README、安装指南、变更日志——面向使用者的信息全部在本文件。

按目录发布

每款 skill 独立版本化，用带 slug 前缀的 tag 发布：

<slug>-v<semver>     如 geo-evidence-audit-v0.1.0

MAJOR = 触发语义 / CLI / 输出结构破坏性变更 · MINOR = 新增信号、规则或报告字段 · PATCH = 误报漏报修复、文档、夹具。永不使用裸 v* tag 和浮动 latest。只需单个 skill 时，取对应 tag 下的 skills/<slug>/ 目录即可。

工程约束

• 纯 Python 标准库，零第三方依赖，Python 3.9+（CI 语法门禁强制）
• 完全离线：审计脚本绝不发起网络请求
• 确定性：相同输入产生逐字节相同的输出
• finding 必带 file:line；证据强度只标注、不断言
• 任何地方不做性能声明：实测之后才写数字
• CI 门禁：py39 语法、纯标准库 import 扫描、SKILL.md < 500 行、全量 pytest 矩阵（ubuntu/windows × 3.9/3.12）

许可证

MIT