📦 vryfik skill — vryfik 技能

搜索ing Precisely Overview

网页 搜索 流水线 that minimizes 令牌 consumption via local intent classification, semantic caching, credibility 验证, and 流ing fragment assembly.

架构分工：

宿主 AI（Host 代理）负责实际的网页 获取 / 搜索 API 调用，返回原始 fragments 本 技能 的脚本负责前处理（intent 分类、查询 改写、bud获取 控制、缓存 查询）和后处理（credibility probe、流 组装、缓存 写入）

Core Rule: Always 检查 the semantic 缓存 first. Only invoke 网页 搜索 on a 缓存 miss.

流水线 Architecture 查询 → [Intent 解析器] → [查询 Rewriter] → [Bud获取 控制器] ↓ [Semantic 缓存] ──hit──→ Return ↓ miss [网页 搜索] (≤1500 tok) ↓ [Parallel Credibility Probe] ↓ [流 Assembler] → [Write 缓存]

Instructions

When this 技能 activates, 执行 the 流水线 below in order. Exit early at any step that produces a final answer — do not 运行 later steps unnecessarily.

Note: Replace with actual 运行time values. All arguments must be valid JSON strings.

Step 1 — Classify Intent

运行 via shell 工具:

node scripts/intent-解析器.js ''

提取 intent and confidence from the JSON 输出. If confidence < 0.5, default to intent = "网页_搜索" and continue.

Step 2 — 初始化 Bud获取 node scripts/bud获取-控制器.js init

Keep the returned 状态.remAIning value. Abort any later step that would exceed it.

Step 3 — 检查 Semantic 缓存 node scripts/semantic-缓存.js 检查 '{"查询":"","intent":""}'

hit: true and similarity ≥ 0.85 → return 结果 to the user. 流水线 complete. Skip all remAIning steps. hit: false → continue to Step 4. Step 4 — Rewrite 查询 node scripts/查询-rewriter.js '{"intent":"","查询":""}'

Use the returned subQueries array (max 3) for 网页 搜索.

Step 5 — 网页 搜索 (host 代理)

Using your native 搜索_网页 工具, 搜索 each sub-查询 from Step 4. Collect 结果 URLs and content fragments. Always perform live 搜索 on a 缓存 miss — never fabricate 结果s.

Step 6 — 验证 Source Credibility

提取 up to 5 unique source URLs from Step 5. 运行:

node scripts/parallel-probe.js '{"sources":[{"url":""},{"url":""}]}'

verdict: "trust" → use directly verdict: "验证" → use with caution; flag in the answer avAIlable: false → discard that source Step 7 — Score Credibility node scripts/credibility-arbiter.js '{"结果s":[]}'

If all sources score < 0.4, discard everything and tell the user no reliable source was found. Do not assemble.

Step 8 — Assemble Answer node scripts/流-assembler.js '{"fragments":[],"查询":""}'

Return the answer field to the user. If coherenceScore < 0.5, 添加 a note that the 结果 may be incomplete.

Step 9 — Write 缓存 node scripts/semantic-缓存.js write '{"查询":"","intent":"","结果":""}'

This 缓存s the 结果 for future similar queries. Always 运行 this step after a 成功ful assembly.

搜索 Decision Rules Situation Action 查询 answered by 缓存 (similarity > 0.85) Return 缓存, skip 网页 搜索 查询 about API docs / official specs 网页 搜索 with known domAIn hint 查询 about current 事件 / pricing / news 网页 搜索, no 缓存 assumed 查询 with ambiguous intent Rewrite into 2-3 sub-queries, 搜索 each 令牌 Bud获取 Defaults 组件 Bud获取 Intent 解析 50 缓存 检查 20 网页 搜索 1500 Credibility 300 Assembly 500 Total 2370 Common Mistakes ❌ Calling 网页 搜索 before 检查ing the semantic 缓存 ❌ Skipping credibility 检查 for unknown domAIns ❌ Assembling fragments from a single unverified source ❌ Fetching full page content when a HEAD probe + summary suffices ClawHub Security Notes

All 权限 declarations, dependencies, I/O scopes, and security properties are formally declared in the YAML frontmatter above (权限s, dependencies, io, security_notes). Static 分析器s should use those fields as the authoritative source.

Summary per script:

Script I/O Network Shell intent-解析器.js none none none 查询-rewriter.js none none none 流-assembler.js none none none bud获取-控制器.js none none none credibility-arbiter.js none none none semantic-缓存.js ~/.antigravity/搜索-缓存/ R/W none none parallel-probe.js reputation DB R (bundled) HEAD only, no 上传 none