Web Session Automation

Web session automation (웹세션 자동화)

Attach to one dedicated automation Chrome over CDP and reuse the logged-in sessions inside it. Uploading, downloading and chatting are the same technique with a different verb.

Foundation = one automation Chrome. Launch Chrome with your own profile dir + --remote-debugging-port=9222; log in once to the sites you'll use (claude.ai, chatgpt.com, gemini.google.com, a KakaoTalk Business chat, …). Every verb connects with connectOverCDP and drives the relevant tab.

This skill follows the open Agent Skills format and is host-neutral. Use it from Claude Code, Codex/ChatGPT desktop, Codex CLI, or Antigravity CLI. Resolve <SKILL_ROOT> to the absolute directory containing this SKILL.md; do not assume the current working directory is the skill directory.

• No LLM, no API keys for the six fixed verbs. It reuses your browser login; you log in, it never types passwords.
• agent is the exception: it shells out to your already-logged-in claude CLI session (still no API key — same subscription, no extra cost beyond normal usage) to decide each step. download/upload/chat fall back to it automatically when their fixed script fails structurally; a blocked result (login/credentials needed) never triggers that fallback, since agent can't get past that wall either.
• An elevated/admin everyday Chrome refuses CDP attach — use a dedicated profile automation Chrome.

Files

web-session-automation/
  ├─ SKILL.md       ← this orchestration file
  └─ session.js     ← engine: status / download / upload / youtube / chat / facebook / agent (Node + Playwright)

Config: env CDP_URL (default http://localhost:9222), PW_PATH (playwright module path if unresolved).

0. Pre-check

node "<SKILL_ROOT>/session.js" status

If cdp:false, launch the automation Chrome (own profile + --remote-debugging-port=9222) and log in once to the target sites.

1. upload

node "<SKILL_ROOT>/session.js" upload --file "<abs path>" [--url "<chat URL>"] [--kakao] [--input-index 0]

--kakao adds KakaoTalk Business completion polling + blocker detection. Outward action — confirm target & file before running, then verify via the returned ok/screenshot.

1-Y. youtube (YouTube Studio upload wizard)

node "<SKILL_ROOT>/session.js" youtube --file "<mp4>" --title-file "<title.txt>"|--title "<s>" [--desc-file "<desc.txt>"] --visibility public|unlisted|private [--publish] [--shot-dir "<dir>"]

YouTube upload is a multi-step wizard (details → elements → checks → visibility), not a plain file input, so it has its own verb. Flow: create → upload video → set file → title/description (clear filename prefill, then type) → not-for-kids → next×3 → visibility → (publish). Without --publish it stops right before publishing and saves screenshots (yt_details.png, yt_visibility.png) — do a dry run, review, then publish. --visibility defaults to private. If not logged in, returns blocked → log in to studio.youtube.com in the browser. Publishing is an irreversible external action — confirm before --publish. (Selectors target YouTube Studio's Polymer DOM; some text selectors assume the Korean UI.)

2. download (claude.ai / KakaoTalk Business)

# claude.ai artifacts + conversation text
node "<SKILL_ROOT>/session.js" download --out "<dir>" [--url "<chat URL>"] --artifacts all|names|none [--names "a|b|c"] --text none|handoff|full|both
# files/videos from a KakaoTalk Business chat (each message's save button)
node "<SKILL_ROOT>/session.js" download --out "<dir>" --url "<kakao chat URL>" --kakao [--limit N]

claude.ai artifact buttons matched by aria-label; with --kakao, save buttons matched by a.btn_save (--limit N caps count). Saved via Playwright download event + saveAs(). Verify via _result.json (verify.ok, sizes > 0). Ask the user what to save before running.

3. chat (ChatGPT / Gemini)

node "<SKILL_ROOT>/session.js" chat --site chatgpt|gemini --prompt-file "<file>" [--url "<conversation URL>"] [--out "<file>"] [--timeout <seconds>]

Pass the prompt as a file (avoids escaping / early submit). Waits for a stable answer. If not logged in, returns blocked: login required → log in once in the browser.

• --url targets an existing conversation (e.g. https://chatgpt.com/c/<id>). If that conversation is not reachable — wrong account, deleted — the verb returns blocked instead of posting into a different or brand-new chat.
• The submit is verified. The inserted text is checked, then Enter, then the send button as a fallback. With no evidence of a send (new user message / emptied composer / generation started) you get ok:false, reason:'submit failed …'.
• Only a new answer counts. The last answer is snapshotted before sending; ok:true requires the collected text to differ from that snapshot. Previously a failed send would return the previous answer as if it were fresh.
• Do not baseline on message counts — ChatGPT virtualises the DOM and the rendered count is not monotonic. Also note ChatGPT reuses one submit button whose data-testid toggles send-button ↔ stop-button, so matching only [data-testid="send-button"] never finds it.
• status reports logins[site] === null when no tab for that site is open (state unknown); true means a sign-in button was actually absent from the page, not merely that the URL looked fine.

4. facebook (own profile: draft / schedule / read stats)

node "<SKILL_ROOT>/session.js" facebook status
node "<SKILL_ROOT>/session.js" facebook stats    --url "<post URL>" [--shot-dir "<dir>"]
node "<SKILL_ROOT>/session.js" facebook insights --url "<content/insights URL>" [--shot-dir "<dir>"]
node "<SKILL_ROOT>/session.js" facebook draft    --text-file "<body.txt>" [--image "<img>"] [--shot-dir "<dir>"]
node "<SKILL_ROOT>/session.js" facebook schedule --text-file "<body.txt>" [--image "<img>"] --date "Jul 23, 2026" --time "11:00 AM"

Read modes (status/stats/insights) only look; write modes (draft/schedule) never publish immediately — Post is never clicked, only Save or Schedule for later → Schedule. schedule re-reads the date/time fields and aborts before committing if they didn't set cleanly. Everything lands in Content Library → Drafts / Scheduled; screenshots (fb_staged.png, fb_draft_done.png, fb_schedule_pre.png, …) go to --shot-dir. Selectors assume the English UI. Automating your account is at your own risk under Meta's terms — use it on your own profile, at human pace.

5. agent (general-purpose fallback)

node "<SKILL_ROOT>/session.js" agent --goal "<plain-English goal>" [--url "<start URL>"] [--max-steps N] [--run-dir "<dir>"]

Loop: screenshot + interactive-element list → ask claude -p (your existing CLI login, structured JSON output) for one next action → click/type/select/attach_file/scroll/navigate → repeat, up to --max-steps (default 25). Stops itself (blocked) on anything a human must do — login, identity verification/e-signature, CAPTCHA, payment, or an irreversible final confirmation — enforced both by the prompt and, independently, in code (a hard refusal on input[type=password] and on button text matching a sensitive-action pattern, regardless of what the model decided). download/upload/chat retry through this automatically when their fixed path fails structurally. Every run logs to graduation_log.json; a host with 3 successful runs gets flagged graduation_candidate: true — a nudge to write it a dedicated fixed verb, not an automatic rewrite.

🗣️ Confirm before acting — same rule as the fixed verbs: confirm the target/content with the user before giving agent a goal that sends, publishes, pays, or deletes something.

Note — when Claude in Chrome may be a better fit: the agent verb is for driving session.js headlessly from a script (each step spawns a fresh claude -p process). If you're instead in a live, interactive Claude Code session and the user just wants a browser task done right now, claude --chrome (Anthropic's own Chrome extension, exposed as MCP tools — read_page/click/type/etc.) lets Claude Code drive the browser directly, with no session.js/CDP/decision-loop involved at all. That's often simpler for a one-off interactive task. Prefer --chrome for that case and reserve agent for scripted/repeatable use. (Not installed or verified as of this writing — requires the Chrome extension and a direct Anthropic subscription login.)

Adding sites

Add {url, input, answer, stop} selectors to the SITES map in session.js to support more chat targets. Upload/download depend on the target page's file input / download-button pattern (locale-specific selectors may need adjusting).

Do not

1. Type credentials for the user — logins / 2FA / extra-auth are theirs to do.
2. Send (upload/chat) without confirming the target and content.
3. Attach to the everyday admin Chrome — dedicated profile only.
4. Treat "the call returned" as success — verify with the result JSON / screenshot / received message.

Host tools

• Use the host's shell tool to run node <SKILL_ROOT>/session.js ... and, when needed, launch the dedicated Chrome.
• Use the host's file and image-reading tools to inspect result JSON and screenshots.
• Use the host's normal confirmation mechanism before sending, publishing, paying, deleting, or making another irreversible change.