القائمة
Use this skill for web browsing, logged-in websites, social platforms, dynamic pages, screenshots, downloads, and any task that needs a real browser session.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Guided Team Console Task design advisor and creation/update helper. Use when the user says "/team-task" or asks in natural-language to create, update, design, or complete a Team Console Task, WorkUnit, Discovery Task, reusable/template task, worklist-producing task, split-task, 任务卡片, 发现任务, 分片任务, or a multi-source research workflow. Helps non-expert users turn fuzzy goals into the right Task shape, asks plain-language questions, previews full JSON, and writes only through /v1/team/tasks after confirmation. Do not use for running, observing, or debugging Tasks.
Creates repo-local requirement/step plans and ready-to-send messages for a separate GLM/coding agent. Use when the user says to call glm-plan, asks to hand work to another agent, or wants a detailed plan/message/file for an external agent to implement.
Use when and only when the user message contains the explicit keyword "/team-plan". This skill guides the Agent through creating TeamUnit presets and Plans via the /v1/team REST API. If the user merely mentions "team plan", "团队计划", Team Runtime, or planning without "/team-plan", do not use this skill automatically; ask whether they want to start with "/team-plan".
无浏览器实体的 HTTP(S) 网络访问技能。Use when an agent needs lightweight network requests such as public web pages, JSON APIs, RSS, sitemap, metadata checks, redirects, static HTML extraction, or file downloads without browser interaction.
Use when the user asks to view, list, create, configure, switch, verify, archive, delete, or reason about independent agent profiles / operation windows in ugk-pi, including Chinese requests like “我有哪些agent”, “有哪些 agent”, “创建 agent”, “切换 agent”, “删除 agent”, “给 search agent 安装技能”, “给其他 agent 安装技能”, or “搜索 Agent 有哪些技能”. Do not use for legacy .pi/agents subagent files unless the user explicitly says subagent.
用于提醒、定时任务、周期总结、后台自动执行和一次性延迟执行。通过 `conn` 工具直接管理任务，而不是发明关键字匹配器或表单式追问。
| name | web-access |
| description | Use this skill for web browsing, logged-in websites, social platforms, dynamic pages, screenshots, downloads, and any task that needs a real browser session. |
| allowed-tools | Bash |
Use this skill as the single external entry point whenever a task needs web access. Do not present WebSearch, WebFetch, curl, Jina, raw CDP, host IPC, or Docker sidecar as choices for the agent to pick manually.
Use staged routing:
S1: static public access through WebSearch, WebFetch, or curlS2: Jina readerS3: real browser automation through this project's browser bridgeMove from a lower stage to a higher stage only after a current-round failure is confirmed. Current-round failures include fetch errors, empty or irrelevant results, login walls, anti-bot blocking, missing dynamic content, incomplete screenshots, or content that requires cookies.
Historical route cache may influence the starting stage. It must not override current evidence. If the chosen stage fails in this round, upgrade.
Primary path for Docker and Linux:
agent / skill
-> requestHostBrowser()
-> WEB_ACCESS_BROWSER_PROVIDER=direct_cdp
-> LocalCdpBrowser
-> http://172.31.250.10:9223
-> Docker Chrome sidecar
Legacy fallback for Windows host development only:
agent / skill
-> requestHostBrowser()
-> .data/browser-ipc
-> host-browser-bridge-daemon.mjs
-> Windows Chrome CDP
The function name requestHostBrowser() is legacy. In Docker sidecar mode it does not mean "use Windows host IPC"; it dispatches directly to the sidecar CDP backend before IPC is attempted.
Always start with:
node /app/runtime/skills-user/web-access/scripts/check-deps.mjs
If you are debugging from the Windows project directory instead of the /app container, use:
node runtime/skills-user/web-access/scripts/check-deps.mjs
Expected healthy output in Docker sidecar mode:
host-browser: ok (http://172.31.250.10:9223)
proxy: ready (127.0.0.1:3456)
The host-browser label is kept for compatibility with older scripts. In sidecar mode it means "browser backend is reachable", not "Windows host IPC is active".
Use these project-level sidecar commands from the repository root:
npm run docker:chrome:check
npm run docker:chrome:status
npm run docker:chrome:open
npm run docker:chrome:restart
Use docker:chrome:restart only when Chrome itself needs a kick. It keeps the persistent profile but clears stale Chrome locks and crash-restore prompt state.
Only use the Windows host IPC bridge when sidecar mode is not configured or when explicitly debugging the legacy Windows path.
The legacy launcher is:
powershell -ExecutionPolicy Bypass -File .\scripts\start-web-access-browser.ps1
Do not tell a Docker sidecar user to start this launcher just because Chrome is unavailable. If WEB_ACCESS_BROWSER_PROVIDER=direct_cdp is set, first check the sidecar services and CDP endpoint.
Sidecar facts:
http://172.31.250.10:9223.data/chrome-sidecar${WEB_ACCESS_BROWSER_PROFILE_DIR:-/config/chrome-profile-sidecar}WEB_ACCESS_BROWSER_PUBLIC_BASE_URL=http://ugk-pi:3000Manual login and automation must use the same profile path. Do not split manual login and auto-start across different profile directories.
Keep passing the platform-provided scope metadata to proxy calls. Do not rewrite WEB_ACCESS_CDP_HOST during a task; it is process-global and can make concurrent work jump browsers.
web-access must not inspect, list, configure, switch, bind, clear, or change Agent / Conn browser settings. Browser assignment is a user-only Playground UI setting. If the user asks to change it, stop and tell them to use the UI instead of browser automation. Do not edit browser route caches, environment variables, Docker Compose, Chrome profiles, cookies, or login state to satisfy such a request.
Chrome must stay on DISPLAY=:0 with --ozone-platform=x11. Do not switch it back to Wayland unless Chrome top-layer UI clicks are revalidated.
Container workspace paths such as /app/runtime/..., /app/public/..., and file:///app/... are valid internal artifact inputs for this skill.
The browser bridge resolves supported local artifacts to an HTTP URL before Chrome opens them:
/app/runtime/report.html
-> http://ugk-pi:3000/v1/local-file?path=%2Fapp%2Fruntime%2Freport.html
Keep these two URL bases separate:
PUBLIC_BASE_URL: user-visible links, such as http://127.0.0.1:3000 locally or a public domain in productionWEB_ACCESS_BROWSER_PUBLIC_BASE_URL: browser-automation links inside the compose network, normally http://ugk-pi:3000Do not make sidecar Chrome open http://127.0.0.1:3000/.... Inside the browser container that points at the browser container itself, not the ugk-pi app.
Do not make sidecar Chrome open file:///app/.... The sidecar is a different container and does not own the app container's filesystem view.
Only in the final user-facing answer should you avoid raw container paths. For user delivery, return the PUBLIC_BASE_URL-reachable link or use send_file.
When a logged-in website needs a local image or other file selected through the sidecar browser, use the shared upload bridge instead of guessing profile paths:
app / agent container writes:
/app/.data/browser-upload/<file-name>
sidecar Chrome selects:
/config/upload/<file-name>
The two paths are the same host directory mounted into both containers. Do not pass /app/.data/... to DOM.setFileInputFiles; CDP resolves paths from the browser container's filesystem, so the browser-side path must be /config/upload/<file-name>.
Do not use page JavaScript fetch() as the upload transport for third-party creator platforms. It is blocked by page origin, CORS, Private Network Access, or platform logic often enough that treating it as a file picker replacement is just asking the browser to be a confused courier. Use the native file input/CDP path first, then fall back to drag/drop or clipboard simulation only after confirming the file exists at /config/upload.
When check-deps.mjs passes, a compatibility proxy is available at http://127.0.0.1:3456 inside the app container.
Browser-changing proxy calls require metaAgentScope. This is intentionally enforced at the proxy boundary so scripts cannot silently fall back to a stale long-lived browser process. If a script receives missing_agent_scope, read the current scope from CLAUDE_AGENT_ID, CLAUDE_HOOK_AGENT_ID, or agent_id, then add metaAgentScope=<scope> to the proxy URL.
Common endpoints:
AGENT_SCOPE="${CLAUDE_AGENT_ID:-${CLAUDE_HOOK_AGENT_ID:-${agent_id:-}}}"
curl -s http://127.0.0.1:3456/health
curl -s http://127.0.0.1:3456/targets
curl -s "http://127.0.0.1:3456/new?url=https%3A%2F%2Fexample.com&metaAgentScope=${AGENT_SCOPE}"
curl -s "http://127.0.0.1:3456/session/target?metaAgentScope=${AGENT_SCOPE}"
curl -s -X POST "http://127.0.0.1:3456/session/target?target=ID&metaAgentScope=${AGENT_SCOPE}"
curl -s -X DELETE "http://127.0.0.1:3456/session/target?metaAgentScope=${AGENT_SCOPE}"
curl -s -X POST "http://127.0.0.1:3456/session/navigate?url=https%3A%2F%2Fexample.com&metaAgentScope=${AGENT_SCOPE}"
curl -s -X POST "http://127.0.0.1:3456/session/close-all?metaAgentScope=${AGENT_SCOPE}"
curl -s "http://127.0.0.1:3456/info?target=ID&metaAgentScope=${AGENT_SCOPE}"
curl -s -X POST "http://127.0.0.1:3456/eval?target=ID&metaAgentScope=${AGENT_SCOPE}" -d 'document.title'
curl -s -X POST "http://127.0.0.1:3456/type?target=ID&metaAgentScope=${AGENT_SCOPE}" -H "Content-Type: text/plain; charset=utf-8" -d 'text to insert'
curl -s -X POST "http://127.0.0.1:3456/key?target=ID&key=Enter&metaAgentScope=${AGENT_SCOPE}"
curl -s -X POST "http://127.0.0.1:3456/enter?target=ID&metaAgentScope=${AGENT_SCOPE}"
curl -s "http://127.0.0.1:3456/navigate?target=ID&url=https%3A%2F%2Fexample.com&metaAgentScope=${AGENT_SCOPE}"
curl -s -X POST "http://127.0.0.1:3456/click?target=ID&metaAgentScope=${AGENT_SCOPE}" -d 'button.submit'
curl -s "http://127.0.0.1:3456/scroll?target=ID&direction=bottom&metaAgentScope=${AGENT_SCOPE}"
curl -s "http://127.0.0.1:3456/screenshot?target=ID&file=/tmp/page.png&metaAgentScope=${AGENT_SCOPE}"
curl -s "http://127.0.0.1:3456/close?target=ID&metaAgentScope=${AGENT_SCOPE}"
When passing a nested URL into /new, /navigate, or /session/navigate, URL-encode it first. Otherwise query params inside the target URL can be mistaken for proxy params.
For React, Draft.js, ProseMirror, and other rich text editors, prefer /type after focusing the editor with /eval, for example editor.focus(). The /type endpoint uses CDP Input.insertText, which follows the browser text input path more closely than document.execCommand('insertText'). It inserts at the current cursor position and does not clear existing content.
For rich editors that must create real paragraph nodes, use /key?key=Enter or /enter between /type calls instead of inserting HTML paragraph tags. This follows the browser keyboard event path and lets the editor update its internal state.
Define the current agent scope once when browser work may span multiple commands:
AGENT_SCOPE="${CLAUDE_AGENT_ID:-${CLAUDE_HOOK_AGENT_ID:-${agent_id:-}}}"
Use that scope on proxy requests with metaAgentScope=${AGENT_SCOPE} so the proxy can reuse the current agent-owned target instead of competing for a shared page. For manual page changes, prefer POST /session/navigate?url=...&metaAgentScope=${AGENT_SCOPE}. It reuses the scoped default target and creates one only when missing. /new is a lower-level escape hatch; within the same scope it replaces the old default target before registering the new one, so do not use it to accumulate scratch tabs.
For a concrete URL, prefer the automatic runner:
node /app/runtime/skills-user/web-access/scripts/staged-route-cli.mjs run-url --url "https://example.com" --task-kind "open_page"
Only use recommend or report for manual route debugging.
Prefer eval for extraction. Use /session/navigate for the current agent-owned page. Only use /new when you intentionally need to replace that page or create an explicitly managed target, and close explicit targets you created unless you intentionally keep an agent-owned page alive for follow-up work.