agent-reach

name	agent-reach
description	Use this skill when Pixiu needs platform-specific or multi-backend internet access through Agent Reach: Twitter/X, Reddit, XiaoHongShu, Bilibili, YouTube subtitles/video metadata, GitHub CLI research, RSS, V2EX, LinkedIn, Xueqiu, Exa search, or a multi-platform public-opinion/research workflow.
when_to_use	The user asks for a platform-specific internet task, authenticated/browser backed channel, video subtitle/metadata workflow, GitHub CLI research, RSS, or multi-platform public-opinion research that benefits from Agent Reach's channel routing and doctor checks.
when_not_to_use	Do not use for ordinary URL reads, simple web lookup, normal project work, or tasks where Pixiu's built-in web_fetch and web_search tools are enough.
triggers	["agent-reach","Agent Reach","Twitter","X","x.com","tweet","Reddit","XiaoHongShu","xiaohongshu","xhs","小红书","Bilibili","B站","YouTube","GitHub research","GitHub code search","RSS","V2EX","LinkedIn","Xueqiu","雪球","Exa","multi-platform research"]
required_tools	["shell","skill","request_user_action"]
risk	medium

Agent Reach For Pixiu

Agent Reach is an optional capability layer for platform-specific internet access. In Pixiu, it is a Skill-backed route to external CLIs and MCP-style tools, not a replacement for Pixiu core tools.

Routing Rules

Use Pixiu built-ins first for lightweight generic work:

• Ordinary URL reading: web_fetch
• Simple current web search: web_search
• Normal file/project work: Pixiu file, grep, shell, edit, patch, and todo tools

Use this Skill when the task needs Agent Reach's platform channels or multi-backend routing:

• Twitter/X tweets, users, timelines, articles, or search
• Reddit posts, comments, subreddits, or search
• XiaoHongShu notes, comments, search, or feed
• YouTube subtitles, video metadata, video search, or comments
• Bilibili search, video details, subtitles, hot/rank
• GitHub repository, issue, PR, release, Actions, or code research through gh
• RSS feed parsing
• V2EX public API topics, replies, users, or nodes
• LinkedIn profiles, companies, jobs, or people search
• Xueqiu stock/community lookup
• Exa semantic search or code context through mcporter
• Multi-platform public-opinion or product research

For more Pixiu-specific routing detail, load references/pixiu-routing.md.

Browser-use handoff: if a XiaoHongShu Agent Reach backend is blocked by login, QR scan, captcha, 2FA, cookie/session, browser authorization, or the user explicitly chooses browser-use/the browser route, stop Agent Reach backend probing and load Skill(browser-use). Choose a fresh task-specific browser-use session name such as pixiu-xiaohongshu-tech; then run browser-use doctor, browser-use --headed --session <session-name> open https://www.xiaohongshu.com, and browser-use --session <session-name> state. A successful browser-use doctor is only an availability check; continue to the headed open command before reporting browser-use as attempted. Do not keep trying Jina, public/private APIs, third-party aggregators, temporary MCP installs, or scraping scripts after the route is blocked or browser-use is selected.

First Check

Before platform-specific work, check whether Agent Reach is available:

agent-reach doctor --json

Use the JSON report to select the active backend for multi-backend platforms such as XiaoHongShu, Reddit, Bilibili, and Twitter/X. If agent-reach is not installed, use Pixiu's managed tool environment when the user has asked or approved installation.

Safe preview commands:

pixiu tools env status
pixiu tools install agent-reach

Full install command, only after the user wants it:

pixiu tools install agent-reach --yes

After installation, run agent-reach doctor --json again. Do not use system pip, --break-system-packages, or global package mutation.

Do not install all optional Agent Reach channels by default. Install only the channel needed for the user request, and ask before login-heavy or browser-backed channels.

Hard Stop Conditions

Stop the current execution route and call request_user_action before trying workaround commands when any of these happen:

• agent-reach is missing and the user has not already asked Pixiu to install it. If the user has asked or approved installation, install it through pixiu tools install agent-reach --yes.
• The user explicitly chose browser-use, the browser route, or a visible browser for the blocked platform task. Load Skill(browser-use) instead of continuing Agent Reach backends.
• The required platform channel is missing and installation would add external packages, browser tooling, MCP services, or persistent config.
• A backend reports missing login, cookie/session, QR scan, captcha, 2FA, browser authorization, API key, account permission, or proxy setup.
• A login command starts downloading browser automation tooling, hangs while waiting for a QR scan, or asks for interactive user input.
• Anonymous access is blocked by the platform.

Do not bypass platform authentication with ad hoc private endpoints, scraping scripts, Playwright/Camoufox experiments, third-party aggregator scraping, temporary package/MCP installs, or repeated blind retries. If the user explicitly chooses a non-Agent-Reach fallback, keep it read-only, explain the reliability limits, and do not handle credentials outside the configured Agent Reach path.

Example request for XiaoHongShu login:

{
  "title": "需要小红书登录态",
  "reason": "小红书后端需要你授权登录后才能读取搜索、Feed 或热门内容。",
  "category": "auth",
  "instructions": [
    "桌面环境：在浏览器登录小红书，并启用 OpenCLI/Agent Reach 推荐的浏览器通道。",
    "服务器环境：如果 xiaohongshu-mcp 已经配置，可在该服务的二维码登录界面完成扫码；不要在任务中临时安装或配置 MCP。",
    "如果你选择 browser-use/浏览器方案，我会加载 Skill(browser-use)，用一个新的任务专用 browser-use session 以 --headed 打开官方页面；遇到登录、扫码、验证码或 Cookie/session 要求时会停下来让你在浏览器里接管。",
    "完成后保持浏览器窗口打开并回复我继续，我会重新运行 agent-reach doctor --json 或 browser-use --session <session-name> state 并使用可用路线。"
  ],
  "resumeHint": "完成登录、Cookie 导入或浏览器授权后回复“继续”。"
}

Command References

Load only the relevant reference file when needed:

• Search and Exa: references/search.md
• Social/community platforms: references/social.md
• Video and podcast workflows: references/video.md
• Generic web and RSS via Agent Reach: references/web.md
• GitHub CLI: references/dev.md
• LinkedIn/career: references/career.md
• Pixiu-specific boundaries: references/pixiu-routing.md

Safety Boundaries

• Read/search/summarize workflows are in scope.
• Posting, commenting, liking, following, applying to jobs, account changes, repo creation, issue/PR creation, or any other write action requires an explicit user request.
• Keep cookies, tokens, and proxy credentials local. Do not paste credentials into reports, artifacts, final answers, or logs.
• Prefer Agent Reach configuration commands for credentials, for example agent-reach configure ....
• Use secondary accounts for cookie-based social platforms when appropriate.
• Do not clone upstream tool repositories or create persistent tool state inside the Pixiu workspace. Use ~/.agent-reach/ for Agent Reach config and /tmp/ for transient command output unless the user asks for a Pixiu workspace artifact.

Evidence

When producing research artifacts or recommendations from Agent Reach data, include source URLs, commands or tools used, access time, and active backend names when available.