local-chatgpt-token-heal

OBJECTIVE: Get the ChatGPT nightly sync back to exit=0 by refreshing the dead bearer token — the correct, root-cause way (session cookie → fresh bearer), falling back to Chrome only when the cookie itself has died.

The root cause (why this skill exists)

The secrets\bearer.txt accessToken expires every ~5-9 days — that is OpenAI's design, there is no refresh-token for it. BUT the login session cookie (__Secure-next-auth.session-token, stored in secrets\session_token.txt) lives ~3 months and chatgpt.com/api/auth/session mints a fresh bearer from it on every call. So the fix is deterministic: cookie → new bearer, no browser, no LLM. The cookie also silently rotates (NextAuth rolling sessions) and token_heal.py captures the rotated one, so routine use keeps the cookie fresh indefinitely.

⚠️ GRABLI: a bare curl -H "User-Agent: Mozilla/5.0" to /backend-api/... returns 403 for BOTH dead AND live tokens — Cloudflare bot-blocks the curl UA before the auth layer. That curl is a USELESS token test. The authoritative check is token_heal.py --verify-only (urllib + full browser headers, reaches the real auth layer → 200/401).

LAYERS (do them in order, stop at the first that reaches exit=0)

L1 — deterministic re-mint (default; no browser, no LLM)

python "%IMPORTS_ROOT%\chatgpt\token_heal.py"     # PowerShell: use %IMPORTS%\chatgpt\token_heal.py

• exit 0 → fresh bearer written + VERIFIED (HTTP 200), cookie rotated if server rotated it. DONE — go to "Finish".
• exit 5 → session cookie dead/absent → do L2.
• exit 6 → transient network/Cloudflare → wait a few min, retry L1 once; still 6 → note it and stop (not a token problem).

Note: nightly_sync.py already calls L1 automatically on pull exit 7 and retries the pull. So most of the time you never run this by hand — this skill is for when L1 itself returns 5 (cookie dead) and a human-in-the-loop Chrome step is needed, or when Anton runs /chatgpt-token-heal directly.

L2 — Chrome re-harvest of the session cookie (only when L1 exits 5)

Needs a Chrome logged into chatgpt.com on account owner.work@example.com + the Claude-in-Chrome MCP (load via ToolSearch if deferred). Do it ON THE HUB.

1. list_connected_browsers → confirm a local Chrome. navigate a tab to https://chatgpt.com/api/auth/session.
2. get_page_text on that tab → JSON. (On the hub this reads the token directly; the old Blob-download dance is NOT needed here.)
3. From that JSON take BOTH fields and write them to secrets (single line, no trailing newline, back up the old first):
   • accessToken → secrets\bearer.txt
   • sessionToken → secrets\session_token.txt ← this is the long-lived cookie; saving it is what makes future heals deterministic
4. Navigate the tab away from the session page (hygiene) so the token isn't left on screen.
5. Re-run L1 (token_heal.py) to VERIFY + capture any rotation → expect exit 0.

If the page shows an empty {} or a login screen → Chrome is logged out → L3.

L3 — human (only if Chrome is logged out of chatgpt.com)

Escalate to Anton via the approval rail (D-type "needs his hands"): ask him to log into chatgpt.com in Chrome on the hub (account a2), then re-run this skill. This is the ONLY genuine human blocker.

Finish (after any layer reaches a fresh bearer)

cmd /c "%IMPORTS%\chatgpt\nightly_sync.cmd"

then tail %IMPORTS%\chatgpt\_nightly_sync-HUB-1.log → confirm exit=0. Report: which layer healed it, session-cookie expiry date, +N new chats. NEVER print the token/cookie value.

Files

• Engine: %IMPORTS%\chatgpt\token_heal.py (L1 + --verify-only)
• Secrets: secrets\bearer.txt (short-lived, ~5-9d) · secrets\session_token.txt (long-lived cookie, ~3mo) · _token_heal_last.json (heal stamp)
• Auto-heal loop: nightly_sync.py calls L1 on pull exit 7, retries pull, and bus-TASKs the hub (L2) if L1 exits 5.
• Canon: memory [[chatgpt-export-pipeline]], [[credential-store]]; sibling skill /chatgpt-sync.

Like this skill? It is one of 100 in second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.

Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).

Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_

Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.