ProxyLane Web Evidence

安装指南

1. ClawHub CLI 从源直接安装技能的最快方式。
   npx clawhub@latest install proxylane-web-evidence
2. 手动安装 将技能文件夹复制到以下位置之一
   全局模式
   ~/.openclaw/skills/
   工作区
   /skills/
   优先级：工作区 > 本地 > 内置
3. 提示词安装 将此提示词复制到 OpenClaw 即可自动安装。
   请帮我使用 Clawhub 安装 proxylane-web-evidence。如果尚未安装 Clawhub，请先安装（npm i -g clawhub）。
   更倾向于下载？
   获取包含原生技能文件的ZIP压缩包。
   下载技能 (.zip)

什么是

ProxyLane Web Evidence 是 Openclaw Skills 的一项功能，可通过提供的 ProxyLane 连接收集特定来源的网页观察结果。它支持 YouTube 字幕片段或评论、Google AI 引用、通过现有 MCP 会话获取的小红书笔记，以及已识别的商品报价。每种模式都使用统一的 JSON 结果契约，其中包含标准化观察结果、时间戳、已脱敏的原始证据、哈希、诊断信息和覆盖范围详情。
可将其用作 RAG、客户之声分析、引用比较或价格监控之前的证据收集步骤。它不保证普遍可访问，不会通过出口 IP 检查证明区域可见性，不会补充缺失内容，也不会执行后续分析。Openclaw Skills 用户获得的是明确的访问结果和可追溯证据，而非关于来源内容的无依据假设。

应用场景

收集指定语言的 YouTube 字幕片段，作为可追溯的 RAG 输入。
收集有界数量的 YouTube 评论样本，并附上来源链接和覆盖范围详情。
捕获经过审核的 Google AI 界面中可见的引用，以供比较。
使用现有 MCP 会话收集小红书笔记证据。
使用 SKU 或 ASIN 及可选变体，观察已识别的商品报价。
在后续分析前，诊断访问受阻、字幕或会话缺失、不支持的界面以及覆盖不完整等问题。
运行合成测试数据，检查本地执行和标准化流程，但不将其视为实时访问的证明。

工作原理

选择五种来源模式之一，并提供其语义目标，例如视频 ID 和字幕语言、笔记 ID 或搜索查询，或商品 SKU/ASIN 和变体。
准备隔离的运行环境和获准使用的 ProxyLane 连接。通过 PROXYLANE_PROXY_URL 或私有单行代理文件提供 HTTP/HTTPS 代理配置；小红书则要求在现有服务器上设置 XHS_PROXY。
可选运行测试数据演示，以验证本地执行和标准化流程。测试数据结果是合成的，不能证明实时目标可访问。
显式运行实时采集。CLI 会标准化所选来源的观察结果、对凭据进行脱敏、记录原始输入哈希，并强制执行总时限。
检查 JSON 状态、诊断信息、证明模式和覆盖范围。应区分 partial、blocked、no_data、missing_session 和 unsupported_surface，不要将它们视为等同的失败，也不要悄悄改用其他提供方。
保存结果和选定的脱敏原始证据。将证据交给后续分析时，应注明来源 URL、观察时间、模式、覆盖范围、诊断信息和产物哈希。

配置指南

准备隔离的运行环境并安装所需 Python 依赖，查看 references/running.md 了解各模式的专用命令。通过运行环境或私有文件提供代理 URL；切勿将凭据放入命令参数、提示、示例、技能目录或已保存的产物中。不要将依赖安装到应用依赖中。
先运行合成演示：
python3 scripts/collect.py --mode demo --fixture --out output/fixture-demo
运行本地测试：
python3 -m unittest discover -s tests -v
对于面向 JavaScript 的运行环境，请使用委托给 Python 的包装器：
node scripts/proxylane-web-evidence.mjs --mode demo --fixture
如果包装器需要特定的 Python 可执行文件，请将 PROXYLANE_EVIDENCE_PYTHON 设置为隔离的解释器。进行实时采集时，请使用 references/running.md 中对应模式的命令；保留 JSON 结果和选定的原始证据。可选的 --verify-egress 探测仅检查单独的 HTTP 出口 IP 请求，不能证明目标请求所用的 IP 或国家/地区。

数据架构

CLI 会在 stdout 输出 JSON。提供 --out 时，它会保存 outcome.json 和选定的脱敏原始输入文件。统一结果契约见 references/output-contract.md；代表性的测试数据输出见 examples/fixture-output.json。
证据类别 记录的信息
来源身份 模式和特定目标身份，例如视频、笔记或商品标识符
观察结果 带时间戳的字幕片段、评论 ID 和链接、可见的 AI 答案引用、匹配的小红书笔记或已识别的报价
来源信息 来源 URL、观察时间和证明模式
覆盖范围和诊断信息 采集限制、访问状态，以及对结果不完整或不可用的解释
原始证据 选定的脱敏原始输入及其哈希，以便追溯
凭据、Cookie、API 密钥、MCP 身份验证令牌和小红书 xsec 令牌不得出现在结果或持久化证据中。适用时，覆盖范围必须说明抽样、分页、区域不确定性、商品变体或结账限制等局限。测试数据是合成的，绝不能作为实时来源证明呈现。

高级特性

五种特定来源的采集模式共用一个 JSON 结果契约和确定性标准化流程。
凭据脱敏和原始输入哈希支持安全、可审计的证据交接。
总时限和 POSIX 进程组超时可限制来源子进程及浏览器后代进程的运行时间。
可选的 --verify-egress 提供单独的出口 IP 诊断，并明确其边界：无法证明目标请求的出口 IP 或区域可见性。
测试数据模式使用合成输入运行全部五种标准化流程，但不会声称具备实时访问能力。
.mjs
入口支持面向 JavaScript 的技能运行环境，同时将采集工作委托给 Python。
访问状态会区分请求受阻、数据或会话缺失、不支持的界面和覆盖不完整，以指导有界的后续步骤。
此技能不会发布评论、修改账户或会话、安装服务器、隐式调用付费托管 API，也不承诺普遍可访问。

name: proxylane-web-evidence version: 1.0.0
description: >- Collect bounded, traceable web observations through a supplied ProxyLane HTTP
connection: YouTube caption segments or comments, Google AI citations,
Xiaohongshu notes through an existing MCP session, and identified product offers.
Use for source collection and access diagnostics before RAG, voice-of-customer
analysis, citation comparisons or price monitoring. Requires explicit source
identity and records coverage; does not promise universal site access.
triggers:

• "collect verifiable web evidence with ProxyLane"
• "fetch YouTube transcripts through my proxy for RAG"
• "collect a bounded YouTube comments sample with source links"
• "capture Google AI citations for a regional comparison"
• "collect Xiaohongshu note evidence through my existing MCP"
• "observe this identified product price through ProxyLane" tools:
• exec
• read
• write mutating: true
  eval_contract:
  goal: |
  Give an agent repeatable source observations with matching target identity,
  redacted raw evidence and honest access/coverage outcomes using a supported
  supplied ProxyLane connection. Excellent output lets another operator trace
  each observation to the captured input without confusing fixtures, samples,
  provider declarations or regional assumptions with live verified facts.
  dimensions:
• "SOURCE_IDENTITY — do captions, comments, notes and offers match the requested source and required semantic fields?"
• "TRACEABILITY — can every observation be checked against dated, hashed, redacted input evidence and source links?"
• "ACCESS_DIAGNOSIS — are challenges, missing captions/session, unsupported DOM and dependency failures distinguished with bounded next steps?"
• "COVERAGE_HONESTY — are sampling, pagination, regional uncertainty, variants and checkout limitations stated in machine-readable output?"
• "EXECUTABILITY — can the documented source commands run in an isolated runtime without hidden services or deployment assumptions?"
• "CREDENTIAL_BOUNDARY — are credentials absent from outcomes, persisted evidence and yt-dlp command arguments?" hard_fails:
• "Presenting synthetic fixtures, HTTP success, proxy declarations or an unrelated page as live source proof."
• "Persisting or printing proxy credentials, cookies, API keys, MCP authentication tokens or xsec tokens."
• "Treating organic links as AI citations, or gl/hl/exit IP alone as verified regional visibility."
• "Publishing a comment, changing an account/session, installing a server, or making a paid managed-API call as an implicit fallback."
• "Inventing missing captions, prices, currencies, variant identity, social comments or OCR content."

ProxyLane Web Evidence

Contract

One capability, five source modes: collect verifiable observations using a supplied ProxyLane connection before an agent analyzes or recommends anything. Source modes share a JSON outcome contract, deterministic normalization, credential redaction, raw-input hashes and a total deadline. Analysis, RAG indexing, representativeness and commercial conclusions are downstream work.

This release accepts a supplied HTTP/HTTPS proxy URL with an explicit port from PROXYLANE_PROXY_URL or a private one-line --proxy-file. No credentials are arguments. Xiaohongshu instead requires XHS_PROXY already configured on the existing server; the client can record that declaration but cannot prove or change server egress. Provider provenance is user supplied. A proxy can be blocked and does not provide login, missing content, authorization, stable IP retention or universal access.

The closest existing capabilities are proxy setup, general web research, video summarization and data research. Use those for connection provisioning, discovery/synthesis, or Brain ingestion. This skill remains separate because it owns source-specific identity checks, bounded raw observations and portable evidence outcomes, and does not replace their workflows. Manifest/resolver overlap was reviewed before creation; publishing must add routing entries at their owner, not a vendor cache.

Phases

Choose a mode and the semantic target. Read source modes. A video needs its id and caption language; a note needs id/xsecToken or a search query; a price needs SKU/ASIN and optional variant. Google needs a reviewed, current AI surface DOM contract. Do not convert a search result or login page into a successful target observation.
Prepare the isolated runtime and approved connection. Follow running. Reuse a supplied approved secret through the runtime environment; do not copy it into examples, the skill directory, prompts, command arguments or artifacts. Check executable/dependency versions. Never install into application dependencies.
Run a fixture demonstration first when evaluating the skill. python3 scripts/collect.py --mode demo --fixture --out output/fixture-demo. This calls all five normalization pipelines with explicitly synthetic inputs. It validates local execution, not target access, browser selectors, live captions, sessions or proxy reliability.
Run one explicit live collection. Use the mode command in running, retain the JSON outcome and raw evidence. Optional --verify-egress performs a separate HTTP exit-IP probe; it cannot prove the target request's IP or country. A browser and an API may see different IPs on a rotating connection.
Interpret status and coverage before analysis. Read output contract. partial can contain useful real observations with incomplete or unverified context. blocked, no_data, missing_session and unsupported_surface require different remedies. Do not retry a login/CAPTCHA indefinitely or silently substitute a managed provider.
Deliver evidence with its limits. Cite source URLs, observation time, mode, proof_mode, coverage, diagnostics and artifact hashes. Separate observed content from analysis. Compare regions or repeat prices only after controlling the relevant language, surface, session, product variant and delivery context. Regional causality and checkout totals remain outside this collector's guarantees.

Output Format

The CLI emits JSON on stdout and, when --out is supplied, saves outcome.json and selected redacted raw input files. Observations include timestamped caption segments, comment ids/links, a visible cited AI answer, a matching XHS note, or an identified offer. See the field contract and synthetic representative output.

The .mjs entrypoint delegates to Python for runtimes that expect a JavaScript skill script: node scripts/proxylane-web-evidence.mjs --mode demo --fixture. Set PROXYLANE_EVIDENCE_PYTHON to an isolated interpreter when needed. POSIX process-group deadlines bound source subprocesses and browser descendants.

Release Verification

Verification boundaries and KNOWN_GAPS distinguish executed checks from unproven coverage. The first-release cross-modal gate was explicitly waived after an inconclusive attempt; no authoritative PASS is claimed. Local unit/integration and proposed routing tests are runnable with python3 -m unittest discover -s tests -v. Global resolver registration and publication receipts belong to the release owner. Do not describe this candidate as published or production validated before its actual delivery receipts exist.

Meta Ads 报告
neo_ppt_skills