Adds the AI-driven content pipeline so portal content is never manually
created — the system proposes it, admins review and approve.
Worker:
- `tower.event.extract` queue + Event Extractor processor: LLM extracts
event title/date/time/location from messages with #event hashtag,
date/time patterns, RSVP phrases, or location terms; seva action items
within event messages become separate SEVA_TASK drafts
- `tower.faq.candidate` queue + FAQ Curator processor: LLM extracts Q&A
pairs from messages with question patterns + meaningful answer length
- `classify.processor.ts`: `isEventCandidate()` and `isFaqCandidate()`
run after rule matching on all non-spam, non-DNC messages; candidates
are enqueued to the extraction lanes in parallel with normal routing
- `match-rules.ts`: add P1 to TenantRuleRow action union so P1 rules
are fully typed end-to-end
- `NormalizedMessage`: add `quotedPlatformMsgId` field
- `IngestJobData.effectiveAction`: add `'P1'` to the union
Schema:
- `ContentDraft` model with `DraftType` (EVENT|SEVA_TASK|FAQ_ITEM) and
`DraftStatus` (PENDING_REVIEW|APPROVED|REJECTED); carries `proposedJson`
from the LLM, `confidence` score, and `publishedId` after approval
- Migration: `20260617270000_add_content_drafts`
API:
- `DraftsModule`: `GET /admin/drafts`, `POST /admin/drafts/:id/approve`,
`POST /admin/drafts/:id/reject`; approve creates the actual
Event/SevaOpportunity/KnowledgeItem and marks the draft APPROVED;
reject marks it REJECTED — both audit-logged
Web:
- `/admin/drafts` and `/drafts` — tabbed view by type (All/Events/Seva/FAQ)
- `DraftCard` — shows source message excerpt, AI proposed content, and
approve/reject buttons; approve/reject immediately refresh the list
- "AI Drafts" added to both super-admin and tenant-admin sidebars
- "AI Content Drafts" button added to admin dashboard
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
