/gemini-pull — extract a shared Gemini conversation

One command. Do not re-derive this; every naive approach fails for a different reason (list below).

Steps

1. Run the bundled script (playwright chromium; ~30–60s):

   HOME=/home/paul node ~/.claude/skills/gemini-pull/gemini-pull.js "<share-url>" /tmp/gemini-out.md

   Either URL form works — the script follows the share.gemini.google 301 to gemini.google.com/share/<hex>?skid=….

2. Verify: grep -c '(no response text)' out.md must print 0. Non-zero means responses didn't render — re-run once; if it persists, the share page structure changed (see trap 5).

Why nothing else works (do not retry these)

1. WebFetch dies with Parse Error: Header overflow on Google's oversized response headers. Always.
2. curl / raw fetch returns only the JS shell (title "Gemini - direct access to Google AI", no conversation, no AF_initDataCallback blobs). Content loads client-side via RPC.
3. Headless rendering without cookies blocks forever on Google's consent wall ("Before you continue to Google"). The script pre-sets cookie SOCS=CAISHAgBEhJnd3NfMjAyNDA4MDYtMF9SQzIaAmVuIAEaBgiAo_CmBg on .google.com, which skips it.
4. chromium-shell --headless=new --virtual-time-budget --dump-dom hangs until killed (the page never settles). The script instead waits for the conversation selector, scrolls until the turn count is stable, then extracts.
5. Share pages ≠ chat UI: turns are <share-turn-viewer> elements containing <user-query-content> + <message-content>. There is NO <model-response> on share pages — waiting on it yields user turns only. Bare <user-query> innerText duplicates text via an a11y "You said …" prefix; use user-query-content.

Requirements

Playwright's node package + chromium (present on prime at /home/paul/e2e-probe/node_modules/playwright with browsers in /home/paul/.cache/ms-playwright — hence HOME=/home/paul). On another host: npm i playwright && npx playwright install chromium, then edit the require() path at the top of the script.