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 skills/summersec/sumsec-skills/semantic-analyzernpx skills add SummerSec/SumSec-Skills --skill semantic-analyzergit clone --depth 1 https://github.com/SummerSec/SumSec-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/skills/summersec/sumsec-skills/semantic-analyzer)<a href="https://agentmods.dev/skills/summersec/sumsec-skills/semantic-analyzer"><img src="https://agentmods.dev/badge/skills/summersec/sumsec-skills/semantic-analyzer.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 5.1 | $0.00052 | $0.01474 |
| Opus 5 | $0.00026 | $0.00737 |
| Sonnet 5 | $0.00010 | $0.00295 |
| Haiku 4.5 | $0.00005 | $0.00147 |
Grade A, and why
semantic-analyzer scanned grade A with 0 findings 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 2d 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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 101 lines — stays where its author put it; the contents beside it link to each section on GitHub.
指令语义深检(第 3 层)
利用 Claude 对语义的模型侧推理,对 Skill / Prompt / Agent 指令做第三层审阅:找出词典未收录、仍可能拉宽输出边界的用词与句式。
定位:与 Hook 检测的关系
| 层级 | 机制 | 速度 | 覆盖面 |
|---|---|---|---|
| 第 1 层:词典匹配 | Hook,已知词对 | 毫秒级 | 固定 |
| 第 2 层:结构规则 | Hook,固定模式 | 毫秒级 | 固定 |
| 第 3 层:语义深检 | 本 Skill,按需 | 秒级 | 动态 |
Hook 覆盖已入库的 wide 词;本 Skill 补未入库项与语境叠加。
Gotchas
- 模板与 fenced 示例(如 STEP 2b 代码块内的演示词)仅用于说明维度,勿原样复制进生产 Skill;本仓库会对
plugin/skills/下 Markdown 跑 linter。 - 四维打分带主观性;是否收录词典须与用户对齐。
- 收录新词:仅允许改
references/semantic-trap-lexicon.md→npm run build-lexicon→npm test;禁止把lexicon-data.js当手写主路径。 - 已知 wide 词以
require('./lib/lexicon-data.js')的wideWordsZh/wideWordsEn(或zhPairs/enPairs)为准做去重。
语义陷阱的四个判定特征
- 程度性 vs 二元性:陷阱词偏程度连续;安全词偏是否、通过/不通过。
- 展望性:「潜在」「可能」等拉远时间轴。
- 主观判定性:依赖评判者标准。
- 语义发散度:联想网大、难收敛。
原理:满足特征越多,边界越宽,越易突破 Prompt 约束。
工作流
STEP 1:获取目标文本与已知词典
- 文本来源:用户粘贴、给定路径、或当前编辑中的指令文件。
- 已知词典:在仓库根下执行
const lex = require('./lib/lexicon-data.js')
使用lex.wideWordsZh、lex.wideWordsEn判断某 surface 是否已登记为 wide;已登记则本层不再重复展开(Hook 已覆盖)。
STEP 2:逐句深度语义扫描
对每句有实质内容的正文(跳过标题、代码块、空行):
2a. 提取关键词:动作指令词、目标对象词、约束修饰词、输出形态词。
2b. 四维评分(0–3 / 维):
词汇: "优化"
┌──────────────┬───────┬──────────────────────────────────┐
│ 特征维度 │ 评分 │ 判定理由 │
├──────────────┼───────┼──────────────────────────────────┤
│ 程度性 │ 3/3 │ 优化是连续性的,没有明确终点 │
│ 展望性 │ 2/3 │ 隐含"变得更好"的未来预期 │
│ 主观评价性 │ 3/3 │ "更好"的标准因人/场景而异 │
│ 关联松散度 │ 3/3 │ 性能/可读性/架构/安全均可被优化 │
├──────────────┼───────┼──────────────────────────────────┤
│ 总分 │ 11/12 │ → 高危陷阱词 │
└──────────────┴───────┴──────────────────────────────────┘
- 0–3 分:无 → 强
- 得分落档:10–12 → critical;7–9 → high;5–6 → medium-high;3–4 → medium;0–2 → low(可忽略)
2c. 排除已知 wide:见 STEP 1。
2d. 收窄方案:对总分 ≥ 5 的词给出更窄、可执行、同句可替换的表述。
STEP 3:上下文与句式隐患
- 语义叠加:多枚中高严重度词同句共现 → 整体严重度上调。
- 隐式宽边界:如「全面 + 检查」、递进修饰 + 开放式动词等,会拉宽动作范围。
- 否定句反转:如
不要忽略任何潜在风险→ 模型为「不忽略」反而扩联想(示例仅在行内代码中展示,避免 linter 误报)。
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 2d ago First seen · 101 lines · 52 tokens per session scan A 4ec8fab5f438
semantic-analyzer is a skill published in the GitHub repository SummerSec/SumSec-Skills (8 stars, last pushed 21d ago), licensed Apache-2.0. It adds 52 tokens to every session and 1,474 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.
Other skills, from other repositories
dsh-web-pre-push-checks
Use before pushing, opening or updating a pull request, or claiming dsh-web checks pass. Selects the required repository gates and diff-specific generation, build, and GUI evidence.
dsh-web-documentation
Use when adding or editing dsh-web README files, docs, AGENTS.md instructions, user-facing configuration text, or bilingual documentation pairs.
dsh-sdk-upgrade
Safely select and install a compatible official @deepseek-ai SDK release for dsh plugin projects (dsh-web, dsh-trading, and similar monorepos) from npm using an isolated worktree, explicit cohort review, CI-equivalent validation, and controlled rollout — including syncing the project's declared DSH host-version floor…
dsh-web-sdk-compatibility
Adapt and repair dsh-web after an approved official @deepseek-ai SDK/runtime cohort is selected or installed. Compare public API, type, service-injection, module-table, protocol, and behavior changes; map every change to repository consumers; implement the smallest fixes and durable compatibility contracts; handle…
ov-experience-memory
Retrieve and apply OpenViking Experience memories through the Agent runtime's generic OpenViking search and read tools. Use before or during executable, multi-step, or tool-based work such as coding, file or data changes, configuration, deployment, workflow execution, and failure recovery when prior operational…
task-planning-arch
计算任务 gap 并产出下一步可执行子任务 List[TaskSpec];gap 已闭返回空数组。对齐 arch 场景(架构师名册/技术栈概览/双视角分析)确定式分解——按根目标交付物集合 + donechildren 查表(参照 task-planning storage 特例,非自由 LLM 分解)。.