Telegram bot as universal deal intake (five input types, one conversation)

One Telegram DM bot accepts an X URL, raw text pitch, meeting transcript URL, deck PDF attachment, signature-envelope link, or deck-share URL (DocSend, Pitch, Notion, Gamma, Figma, Canva). It detects type, runs the matching ingestor, classifies content availability, runs deep research, pushes a fully-populated task into the team's tracker, and DMs the task link back. Forwarded decks record the scout's handle so attribution is preserved.

Replaces "forward this to me, I'll look at it next week" with "forward to bot, get a researched task link in ~90 seconds". Deals stop getting lost to inbox triage.

1. Create a bot via BotFather; set TELEGRAM_BOT_TOKEN. Only one process should long-poll.
2. URL detection by host substring (x.com, granola.ai, docusign, deck-share hosts).
3. PDF detection by mime + filename suffix.
4. Each type routes to ingestors/<type>.py:ingest() returning a DealContext (founder, company, signal_text, attachments).
5. L1 content-availability gate: if signal_text < 200 chars and no attachment, ask the user for more info and set status awaiting_user_input. On the next message, merge old + new context.
6. L3 research (Opus + WebSearch + Bash) emits a structured memo.
7. Create the tracker task (two-step if your tracker has templates — see #10).
8. DM the task URL back to the sender; record scout attribution on forwarded inputs.

• Privacy-hidden Telegram forwards strip the original sender — fall back to a "who forwarded this?" follow-up question instead of silently losing attribution.

• Family-office variant. Same architecture, applied to a family office's CRM intake. Operator drops in: voice memos from meetings, screenshots of chat pitches, forwarded emails, scanned business cards, photographs of slides. Bot normalises into one deal record per surface. The variant ships with a daily 12-agent news scan attached (see [new-N13-07]) so each new deal-card is enriched with same-day news hooks for the relationship manager.