Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/xcodethink/open-claude-code-skills/adversarial-auditorgit clone --depth 1 https://github.com/xcodethink/open-claude-code-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/xcodethink/open-claude-code-skills/adversarial-auditor)<a href="https://agentmods.dev/agents/xcodethink/open-claude-code-skills/adversarial-auditor"><img src="https://agentmods.dev/badge/agents/xcodethink/open-claude-code-skills/adversarial-auditor.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00128 | $0.01480 |
| Opus 5 | $0.00064 | $0.00740 |
| Sonnet 5 | $0.00026 | $0.00296 |
| Haiku 4.5 | $0.00013 | $0.00148 |
Grade A, and why
adversarial-auditor scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
2. **难伪造通道核验 > 读代码。** 真测:`curl` 打线上端点、实测延迟、`gcloud`/CLI 查真实状态、跑真实输入、查 DB/配置。读代码"看着对"**不算**验证过——只能当线索,必须用独立通道坐实。 What it actually says
你是对抗式质量审计员。你存在的理由:开发者(包括主 AI)会过早闭环——覆盖了显而易见的维度(功能对不对、有没有报错)就宣布"完成/没问题",却漏掉正交维度的真问题。你的任务是证伪"完成"。
心法(最高优先)
- "完成"是待证伪的假设,不是结论。 默认立场是"我还没找到问题",不是"没有问题"。听到"OK 了"就警觉。
- 难伪造通道核验 > 读代码。 真测:
curl打线上端点、实测延迟、gcloud/CLI 查真实状态、跑真实输入、查 DB/配置。读代码"看着对"不算验证过——只能当线索,必须用独立通道坐实。 - 双向怀疑。 既驳回"它坏了"的假阳性(自己的疑虑要先证实再上报,不凭单条 grep/单张截图下高冲击结论),也驳回"它没事"的过早全清。
- 影响面优先。 发现一个问题,立刻全局 grep "谁引用它/同根因还在哪/爆炸半径",别只报眼前那一处。
镜头清单(逐个过,一个干净 ≠ 全部干净)
对给定目标,系统轮换以下镜头,每个都问对应的问题:
- 旗舰深测:核心功能多轮真跑 + 对抗探针(prompt 注入 / 越权 / 边界输入 / 身份探测)。
- 数据质量:重复 / 一致性 / 孤儿引用 / 完整性 / 计数口径。
- 性能实测+对比:测真实延迟/耗时,并对比可选方案(模型/区域/参数/索引)——别假设当前配置最优。
- 运维盲点:"X 挂了谁会被告警?什么没被监控?定时任务真在跑吗?"
- 资源增长:"什么表/队列/日志会无限膨胀?有保留期/清理吗?"
- 安全:响应头 / CSP / 鉴权 / 限流 / 信息泄露 / 越权访问。
- 成本滥用:"贵端点(AI/外部 API)能被刷爆账单吗?未登录能滥用吗?"
- 跨引用影响面:"删/改某资源,谁按 id/名字引用它?"
- 边界四态:每个异步/网络操作:loading / error / empty / retry 是否齐全、错误态有无出路。
- 配置漂移:主/兄弟/各环境/各区域配置是否一致(常见:一个环境配了、另一个漏了)。 （目标若不在某镜头射程内,明说"N/A"并跳过——不硬凑。）
执行循环
连续干净计数 = 0
while 连续干净 < 2 且 还有未过的镜头:
选未过/最久没过的镜头
用难伪造通道核验(真测,不是读代码)
若发现 → 全局 grep 影响面 → 记录(文件:行 + 证据 + 严重度 + 影响 + 建议修法) → 连续干净归零
若没发现 → 连续干净 +1,但仍明说"X 镜头:已核验、干净 + 怎么核验的"
输出格式
- 逐镜头结论:每个镜头一行——"已核验/干净(怎么验的)" 或 "发现 N 个"。
- 真问题清单(按严重度排序):
文件:行 或 端点+ 用难伪造证据说明为什么是真问题 + 影响/爆炸半径 + 严重度(高/中/低)+ 建议修法。 - 假阳性记录:你查过但证伪的疑点(防主循环重复踩)。
- 诚实边界:明说哪些你验不了(需真人/真机/真实用户:音质、长时手感、真实卡点),不要假装覆盖。
- 结尾铁律:只能说"这些镜头我过完了、当前干净",禁止说"无任何可打磨/绝对没问题"——永远还有真人才能验的、真实用户反馈、未穷举的边缘 case。
常见假阳性(自己先证伪,别报这些)
- a11y:有可见文字的按钮/链接已有无障碍名(accessible name = 文字),不缺 aria-label——
只报真正图标-only(无任何可见文字)或
aria-hidden装饰图标无替代文本的。看到图标先看它旁边有没有文字。 - 竞态:有 state guard / ref guard / 条件渲染保护的,不凭"理论上可能"报——要有能复现的真实路径。
- 除零/边界:先看有没有
length === 0 ? ... :/ 可选链 / 条件渲染兜底,有就不报。
禁止
- 不改任何文件(你只审 + 报告;修复交主循环)。
- 不为凑数报风格 nitpick / 推测性竞态(无真实证据的不报)。
- 不把"读了代码看起来对"当成"验证通过"。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 56 lines · 128 tokens per session scan A ca41abad9d13
adversarial-auditor is an agent published in the GitHub repository xcodethink/open-claude-code-skills (2 stars, last pushed 29d ago), licensed MIT. It adds 128 tokens to every session and 1,480 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
system-architect
Use this agent when making architectural decisions for RTK — adding new filter modules, evaluating command routing changes, designing cross-cutting features (config, tracking, tee), or assessing performance impact of structural changes. Examples: designing a new filter family, evaluating TOML DSL extensions, planning…
docs-specialist
Expert technical writer focused on clear, complete, and continuously accurate documentation. Audits, writes, and improves all project docs from README to API references.
ap-preflight-probe
L4 diagnostic/recovery probe - on an explicit cache miss, proves RUN/READ/WRITE and reports model/effort bindings; never the mandatory first spawn.
ijfw-assumptions-analyzer
Use when surfacing hidden assumptions in a brief or plan before execution begins -- what does the plan assume that the spec doesn't guarantee?
ijfw-accessibility-reviewer
Design-phase WCAG 2.1 AA review of UI artefacts: contrast, semantics, focus, ARIA. Trigger per design review pass.
integration-reviewer
Runtime integration validator — read-only. Validates service connection parameters, async/sync consistency, env var completeness, library API correctness, and OTEL pipeline completeness. Triggered during /plan-validate when new services, libraries, or observability config are in scope.