chatgpt-sync

OBJECTIVE: Pull ChatGPT conversations updated since the vault's newest note and fold them into the Obsidian vault (notes + MOC + SQLite), on demand. Idempotent. IDENTICAL pipeline to the scheduled ChatGPT Nightly Sync; the only difference is this fires when the operator asks.

CONTEXT (canon: memory [[chatgpt-export-pipeline]]):

• ChatGPT chats are NOT on disk - pulled via /backend-api with a bearer token in $IMPORTS_ROOT/chatgpt/secrets/bearer.txt. The cookie/bearer lives ~1-2 weeks; there is NO refresh-token, so it must be hand-refreshed periodically (this is why the flow is semi-auto, not a blind cron).
• ⚠️ GRABLI: the detail endpoint /conversation/{id} 403s with the bearer alone - incremental_pull.py already sends the required headers (OAI-Device-Id, oai-language, Referer/Origin). The LIST endpoint works without them.
• Single source of truth = $IMPORTS_ROOT/chatgpt/nightly_sync.py - it chains: vault_backup (pre) -> incremental_pull.py CUTOFF -> chatgpt_export_to_vault.py --zip -> copy staging->vault (NEVER overwrites a concept-enriched note, only related_concepts: []) -> build_chatgpt_moc.py -> vault_backup (post). CUTOFF = newest vault note date - 1 day (so a missed day leaves no gap).
• Vault home: $OBSIDIAN_VAULT/01-Conversations/ChatGPT/conversations/ + _ChatGPT-MOC.md. SQLite: _imports\chatgpt\chatgpt_conversations.db. Freshness heartbeat: _imports\chatgpt\_freshness.json.

STEPS:

1. Run the orchestrator (it does backup + pull + fold + MOC + backup itself): cmd.exe /c "$IMPORTS_ROOT/chatgpt/nightly_sync.cmd" (or python $IMPORTS_ROOT/chatgpt/nightly_sync.py). Run it foreground; it logs to _imports\chatgpt\_nightly_sync.log.
2. Read the tail of $IMPORTS_ROOT/chatgpt/_nightly_sync.log for the result line(s): copy: +N new, ~M refreshed, K enriched-preserved and nightly_sync DONE: +N new chats.
3. Interpret the exit code:
   • 0 = success (even +0 new is fine - nothing new to pull).
   • 7 = PULL FAILED, token dead/expired. nightly_sync.py now self-heals this automatically (calls token_heal.py on exit 7 → re-mints the bearer from the long-lived session cookie → retries the pull), so a bare exit 7 in the log usually means the automatic heal ALSO failed. DEFAULT ACTION: run skill /chatgpt-token-heal (root-cause fix: session cookie → fresh bearer, 0 browser, 0 LLM; falls back to Chrome cookie re-harvest only if the cookie itself died). Do NOT hand-do the old Chrome Blob-download dance — that is the fallback the skill handles. ESCALATE to the operator (D-type) ONLY if /chatgpt-token-heal reaches L3 (Chrome logged out of chatgpt.com on the hub).
   • 8 = staging build failed (rare) → report the log tail, do not guess.
4. (optional) Reindex for RAG so new notes are searchable now: python $IMPORTS_ROOT/brain_embed_update.py --wait-gpu 10 - OR just rely on the nightly Brain Reindex @04:00 (skip if the GPU is busy with a parallel session).
5. Report: how many new chats folded, the newest note date, freshness FRESH/STALE (read _freshness.json), and BROKEN-link status if a validate ran. End with a 🧒 In plain words recap (messages TO the operator only).

CONSTRAINTS:

• ANTI-RECENTS: when hunting for ONE specific chat by hand on chatgpt.com, never conclude "the chat does not exist" from the recents list. Use the built-in SEARCH by keyword + check Projects/archive + confirm the active account by email. The scripted path (/backend-api LIST via incremental_pull.py) pulls ALL chats and is more reliable than eyeballing the page. Canon: memory [[web-ui-search-not-recents]] (incident 2026-07-23).
• DON'T duplicate logic - ALWAYS go through nightly_sync.py; never re-implement the pull/convert here. If incremental_pull.py or the converter needs a change, edit THOSE files (single source), not this skill.
• AK-47: this skill is just this SKILL.md (a procedure). No new server/DB/service.
• WINDOWS cp1252: the log may show mojibake for Cyrillic titles - cosmetic only; the vault notes are correct UTF-8. Don't "fix" by re-encoding the notes.
• Token is a secret: never print it, never paste it into a message/vault/always-loaded layer (credential-store anti-leak boundary).
• Backup-before-write is already inside nightly_sync.py (vault_backup pre+post) - don't double-run it.

RELATION (do not duplicate):

• Scheduled twin: Windows task ChatGPT Nightly Sync (~02:00) + watchdog ChatGPT Vault Freshness (Daily) (~10:00, pings the operator on STALE via Telegram). Same scripts. Live on the always-on HUB HUB-1 since 2026-06-25 (moved off the laptop HP17 so sync no longer depends on the laptop waking; may still be duplicated on HP17 as backup — idempotent, a double run just sees +0).
• Canon + gotchas + token how-to: memory [[chatgpt-export-pipeline]].
• Sibling sync skills: [[health-sync]], [[faaa-sync]], [[claudeai-sync]], [[whatsapp-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.