SMRT CRM Strategic Roadmap

Two-LLM session: openclaw owned DEV patches 068-075, Claude Sonnet ran PROD MAXXL migration (076-078). Full backup before. Zero rebuilds.

✅ openclaw on DEV — 8 patches (068-075)

• 068 SERVER mp-routes.js — schedule btn now finds published contacts (status filter fix)
• 069 BUNDLE MAXXL_PROPER api path /api/mp → /skynet-api/api/mp
• 070 SERVER apollo-integration.js — added enrichPerson method (was crashing)
• 071 SOURCE skynet.tsx — 8 fifaSelected copy-paste bugs fixed in TS source (queued for next rebuild)
• 072 BUNDLE Skynet TM tab crash (fetchTmStats undefined) + tmTab refresh
• 073 BUNDLE Soccerdonna verify mapping (.name→.full_name) + polling array
• 074 SERVER Nightly + Soccerdonna import tolerance (drop bad ids → fallback to all approved)
• 075 SERVER Agent import tolerance (same pattern)

✅ Claude Sonnet on PROD — promote 068-075 (skip 071) + MAXXL migration 076-078

• 068-070+074-075 applied to PROD via openclaw with chattr unlock/relock — backend healthy after restart
• 072-073 applied to PROD CRM bundle
• 071 SKIPPED — needs rebuild, queued (would lose MAXXL IIFEs that aren't yet codified)
• 076 NOARG Replicated 15 mp_* tables DEV → PROD skynet.db (idempotent IF NOT EXISTS)
• 077 NOARG Copied 5 backend files (mp-routes, mp-helpers-v2, mp-crm-bridge, mp-send-worker, contact-identity), fixed DB path /skynet-dev → /skynet, mounted /api/mp + send-worker in PROD server.js
• 078 BUNDLE Extracted 90KB MAXXL IIFE block from DEV bundle (UI_V5 + PROPER + HIDE_CMP_BTN), appended to BOTH PROD bundle filenames (B5BSXjE7 + Dtu0_nu2), chattr +i relocked

✅ PROD verification (curl-based)

• https://crm.pros11.com/ → 200 ✓
• https://crm.pros11.com/skynet → 200 ✓
• https://crm.pros11.com/assets/index-Dtu0_nu2.js → 200, 2,557,979 bytes (was 2,467,221 + 90KB MAXXL) ✓
• https://crm.pros11.com/skynet-api/api/mp/campaigns → 200 [] (isolated PROD db) ✓
• https://crm.pros11.com/skynet-api/api/mp/sender-mailboxes → 200 ✓
• MAXXL strings verified in served bundle: __MAXXL_UI_V5__ + __MAXXL_PROPER_V1__ + __MAXXL_PROPER_API_PATH_V1__ + __MAXXL_HIDE_CMP_BTN__ + maxxl-skynet-page ✓

📦 Pre-flight backup

• /backups/FULL-CRM-SKYNET-20260428_145639/ — 3.9GB, 24 files (PG dumps both DBs, SQLite gz, all code dirs, ledger, configs, openclaw workspace)
• Rollback ready: extract Smrt-CRM-FINAL.tar.gz + restore skynet.db + pm2 restart

⚠️ Known caveats / what may still need fix

• PROD MAXXL UI = pure IIFE — neither DEV nor PROD skynet.tsx React source has MAXXL tab. The MAXXL_UI_V5 IIFE injects everything (tab button, body class, panel). Fragile if React re-renders nav too aggressively.
• Source backport queued — patches 071 (Skynet TS bugs) + MAXXL React component still pending in /Smrt-CRM-FINAL/client/src/pages/skynet.tsx. Next rebuild needs both BEFORE running, or features lost.
• PROD send-worker no SMTP creds — pm2 logs show "channels: resend=false brevo=false hostinger=false". Add RESEND_API_KEY/BREVO_API_KEY to PROD env (pm2 id 7 / skynet-backend) before users send real campaigns.
• 3 truly impossible bugs (minified vars: fifaSelected→nightlySelected/soccerdonnaSelected/agentSelected) → backend tolerance patches 074+075 mask them. Real fix = rebuild after 071 source.
• MAXXL bug 8 (no retry on AI draft) + bug 10 (template toggle DOM sync) still unfixed — embedded in template literals / dual React state. Defer to source backport.
• Roadmap HTTPS — DNS A record missing for roadmap.pros11.com → 187.77.74.76. HTTP-only nginx vhost in place.
• Mimo v2.5 fallback to gemini in OpenClaw — provider-catalog patch applied but still not selected. Needs deeper debug.

🧪 Manual browser verification needed

• Open https://crm.pros11.com/skynet in fresh browser (or Cmd+Shift+R hard refresh)
• Look for "⚡ MAXXL" tab in Skynet nav (10th tab) — IIFE adds it on load
• Click → should see empty campaigns list (PROD db isolated, no test data)
• Try create campaign → add contacts → verify dedup works
• If MAXXL tab missing: check browser console for IIFE errors

All P0 items from previous handoff shipped to DEV. Awaiting browser test + your call on prod backport.

✅ Patches landed (DEV only)

• 052 contact_identity table + indexes (Postgres)
• 053 backfill 120,060 identities from existing CRM contacts
• 054 + 055 dedup helper wired into MAXXL contacts-add route
• 056 MAXXL→CRM bridge: auto-create company + contact when MAXXL adds new lead
• 057 + 059 fetchLeadsForCampaign fallback fixed (returns real CRM contacts when Apify capped)
• 058 + 060 send worker (60s nodemailer cron, drains mp_campaign_messages)
• 061 contact edit/delete + company block (mp_blocked_companies table + 5 endpoints + auto-skip wireup)
• 062 UI v6: Company column + ✏️/✖/🚫 buttons + Blocked Companies floating panel
• 063 free DuckDuckGo search (web hints, no email — last-resort lead suggestor)
• 064 tone-matching backend (sender style ingest endpoints)
• 065 news/funding signals via Tavily — recentSignals[] in research dossier
• 066 reply→Calendly: booking_intent classifier + auto-reply cron + UI v7 settings button
• polish: send-worker mirror to email_communications, sparse research auto-fetch from PG, prompt() → modal block dialog, MANIFEST entries 052-066

🧪 Verified end-to-end

• 9 real CRM football contacts inserted in MAXXL campaign 4 ("My Test"), each w/ 11-section DeepSeek dossier
• Tavily signals returned 3 real Manchester United news bullets in test dossier
• Send worker test: msg sent live via Hostinger SMTP (cleaned up after)
• Bridge test: MAXXL contact auto-created CRM company "Bridge Test FC" + contact (cleaned up after)

⚠️ Open from this session

• Source backport — RULE #4: dist patches MUST land in /Smrt-CRM-FINAL/. 15 patches not yet backported. Separate session.
• PROD schema apply — contact_identity table only on smrt_crm_dev. PROD smrt_crm awaits user OK.
• is_dup policy — currently allows known CRM contacts into MAXXL. Tighten if double-outreach concern.
• Per-contact sender round-robin from mp_sender_mailboxes, not just SMTP_USER fallback.
• UI v6 polish — verify Company column appears + buttons fire after hard-refresh in browser.

⚡ MAXXL = MoneyPrinter clone, lives in Skynet (port 3010 dev, 3006 prod). Branded MAXXL not MoneyPrinter (legal).

🌐 URLs / paths

• UI: https://dev.pros11.com/skynet → 10th tab "⚡ MAXXL"
• Backend: https://dev.pros11.com/skynet-api/api/mp/* (proxied to localhost:3010/api/mp/*)
• Skynet routes file: /skynet-dev/backend/mp-routes.js (~900 lines)
• Helpers file: /skynet-dev/backend/mp-helpers-v2.js
• Mount in server.js: line ~3994 area, marker /*__MP_ROUTES_MOUNT_V1__*/
• Frontend overlay: in /Smrt-CRM-DEV/dist/public/assets/index-Dtu0_nu2.js at end of bundle, marker /*__MAXXL_UI_V5__*/ + inline patches
• SQLite tables: 12 tables prefixed mp_* in /skynet-dev/backend/skynet.db
• PROD: NOT YET PROMOTED. PROD bundle still chattr +i on its old state.

🔥 Sub-agent finding (CRITICAL — invalidates earlier assumptions)

• MoneyPrinter does NOT use Apollo/Apify. Real stack: OpenClaw + Bright Data residential proxies (live LinkedIn scraping). YC product previous name "ClawGTM" = "OpenClaw for Sales" — smoking gun. github.com/brightdata/openclaw-plugin ships 47 scrapers (LinkedIn/Crunchbase/X/Reddit). Privacy policy lists ONLY Google Workspace as third party — no PDL/Coresignal/Apollo disclosed.
• MoneyPrinter LLM = Claude Sonnet (NOT DeepSeek). User dossier quality matches Claude prose. Founder's viral LinkedIn: "emails that don't sound like ChatGPT" = Claude tell. OpenClaw playbook recommends Claude 3.5 Sonnet. Currently MAXXL uses DeepSeek-v4-flash everywhere → quality gap explained.
• Apify cap on user account until 2026-05-09. Pipeline Labs scraper returns "Monthly usage hard limit exceeded". Fallback to existing CRM contacts wired but not firing because early-return path skips it.

📦 What's BUILT on DEV (verified working)

• SQLite schema: 12 mp_ tables (campaigns, sequences, campaign_contacts, contact_research, contact_variables, campaign_messages, sender_mailboxes, campaign_blacklist, campaign_folders, assistant_chats, campaign_analyses, linkedin_accounts)
• Backend endpoints: /api/mp/{campaigns, sequences, contacts, /:id/sequences/draft, /:id/contacts/:cid/publish, /:id/publish-all, /:id/contacts/:cid/stop, /:id/schedule, /sequences/:id/edit-with-ai, /campaigns/:id/blacklist, /assistant/chat, /folders, /sender-mailboxes, /conversations, /scraper/by-company, /webhook-stats, /linkedin/connect, /auto-add/tick}
• AI sequence drafter (DeepSeek v4-flash, var-aware: pulls extracted vars first, drafter only uses known keys)
• Per-contact research dossier v2: 11 sections matching MoneyPrinter exactly (contactInfo with roleDescription / summary / companyInfo with whyThisCompany / fundingInfo / researchFindings { contactProfile / relevanceToOurCompany / companyOverview / painPoints / ourCompanyValueProps / proofPointsToMention / personalizationAngles } / templateVariables)
• Custom variable extractor (in research output)
• Schedule with var resolution + auto signature append (pulls users.email_signature)
• Per-step inline manual edit + Edit-with-AI inline popover + Delete + delay number stepper + Regenerate-all
• Per-campaign Analyze + Per-campaign Blacklist UI (working — verified via curl)
• AI Assistant tool-using chat (CREATE_CAMPAIGN auto-navigates, LIST_CAMPAIGNS, GET_ANALYTICS) — branded MAXXL on DeepSeek (verified)
• Sub-page nav inside MAXXL: 🏠 Home · 📋 Campaigns · 📨 Inbox · 🎯 Find Leads (5 modes) · 📧 Senders (with Google OAuth button) · 💬 Assistant
• Inbox = unified replies port from main CRM email_replies
• Create Campaign modal — 2 paths (AI / Manual 4-step wizard) exact MoneyPrinter field mapping (Campaign / Targeting / Sequence / Launch)
• Per-contact lifecycle: pending → generating → draft → published → active → sent (statuses + colors)
• 📤 Publish per-contact + 🚀 Publish All Draft + ⏸ Stop button
• 🔄 Auto-Add Now manual trigger button
• Auto-add cron worker every 5min on active campaigns (LLM extracts Apollo filters from target_company_text + target_role_text)
• 🔬 Research drilldown UI = full slide-out panel matching MoneyPrinter (all 11 sections, color-coded headers)
• Horizontal Skynet tablist + bulletproof sibling-panel hide
• 5s blink poll = KILLED (was breaking textarea focus); replaced with 🔄 Refresh button

⚠️ KNOWN BROKEN / STUBBED (must fix next session)

1. fetchLeadsForCampaign() Apollo path — works code-wise but Apify account capped till 2026-05-09. Switch to Bright Data + OpenClaw per sub-agent finding.
2. Fallback to existing CRM contacts — code present but not executing on early-return path (no runId case). Bug fix: move fallback BEFORE return.
3. No send worker. Messages sit in mp_campaign_messages with status=awaiting_review forever. Need worker that picks up published+pending → calls Brevo or Gmail OAuth → updates status=sent. Bridge to existing main-CRM email_queue path.
4. NO global dedup logic across CRM/Skynet/MAXXL. Currently 3 separate stores (Postgres contacts table, Skynet agent_contacts SQLite, MAXXL mp_campaign_contacts SQLite). Same person can be added 3× with different IDs. NEEDS NEW SHARED DEDUP TABLE + API.
5. NO import-to-CRM flow. When MAXXL adds a contact (via auto-add or Find Leads), it goes into mp_campaign_contacts only. Should ALSO upsert to main CRM contacts table (Postgres) so it appears in Skynet/CRM contact lists. Mirror the existing Skynet "import to CRM" button pattern.
6. Pending AI Reviews on Skynet Overview shows empty on DEV (works on PROD). Pre-existing UI query mismatch — agent_contacts has 6,426 pending data identical to PROD. Standalone fix ~20min.
7. Campaign blacklist UI — backend/UI verified working via curl, but user reported it as "mock" — possibly tested before v5 patch took effect. Re-verify next session.
8. Lead Scraper modes 4+5 = stub bridges only; need wire-up to Skynet Pipeline Labs CSV upload.
9. Bright Data + OpenClaw integration = TBD whole new infrastructure; 1-2 days work.
10. Tone-matching from sent emails not built (high-impact feature).
11. News/funding signal layer (Crunchbase + news scrape per lead) not built.
12. Multi-channel sequence orchestration with branching not built (LI accept → wait 2d → DM).
13. Reply classifier → auto-Calendly not built (we have classifier from earlier UD work but not wired to MAXXL+Calendly).
14. Daily performance brief email not built.
15. Email signature auto-append shipped backend, not yet visually tested in delivered email.

✅ USER DECISIONS LOCKED 2026-04-25 (pre-compact)

• LLM: KEEP DeepSeek v4-flash for now. Skip Claude Sonnet swap.
• Scraper: Try FREE fallbacks first (raw Playwright on our server, free Phantombuster trial, Apify alt actors). Bright Data = roadmap option, not immediate spend.
• Build order next session: #1 Global dedup → #2 Import-to-CRM → #5 Send worker (F) → #6 Tone-matching from sent emails (B) → #7 News/funding signals (C) → #4 Free scraper fallback → #9 Multi-step LI+Email branching (D — depends on LinkedIn channel) → #8 Reply classifier → auto-Calendly (E).
• Skip for now: #3 Anthropic Claude swap (deepseek first), #10 LinkedIn channel until basics done, #11 Daily brief, #12 Pending AI Reviews fix.
• Reply classifier → auto-Calendly explained: sentiment classifier (already shipped UD-P3) + NEW booking-intent classifier → if positive + intent → auto-reply with Calendly link + read user's Google Calendar availability → on prospect pick → write meeting back via Google Calendar API + Slack/email notify.

🎯 P0 NEXT-SESSION PRIORITIES (in suggested order)

1. GLOBAL DEDUP system (user explicit ask 2026-04-25). New shared table contact_identity in Postgres with: email_lower (unique), linkedin_url_normalized (unique), full_name_lower + company_name_lower composite. CRM, Skynet, MAXXL ALL must check this before insert. Add API POST /api/contact-identity/upsert + GET /api/contact-identity/lookup?email=&linkedin=. Migration: backfill from existing Postgres contacts + Skynet agent_contacts.
2. Import-to-CRM flow from MAXXL (user explicit ask). When MAXXL adds contact via auto-add or Find Leads → mirror to Postgres contacts + companies tables via POST /api/v1/contacts (already exposed). Use shared dedup from #1.
3. Anthropic Claude Sonnet wiring (sub-agent finding: MoneyPrinter = Claude not DeepSeek). User said previously Anthropic key wired in env but couldn't find it — check /Smrt-CRM-FINAL/.env and gmail-poller env first; if missing, ask user for ANTHROPIC_API_KEY. Then add provider switching in mp-helpers: callLLM(prompt, opts) routes by env AI_PROVIDER=claude|deepseek|mimo. Use Claude for: research dossier, sequence drafter, campaign analyze, AI assistant chat. Keep DeepSeek for: cheap classification, edit-with-AI quick rewrites.
4. Bright Data + OpenClaw scraper as Apify replacement. Sign up Bright Data (~$50/mo Web Unlocker tier). Install github.com/brightdata/openclaw-plugin. Replace fetchLeadsForCampaign() path: Bright Data → LinkedIn people search → email-find via secondary call. Handle rate limits properly. Until then: fix the existing CRM-contacts fallback.
5. Send worker. Cron in Skynet every 60s: SELECT from mp_campaign_messages WHERE status='pending' AND scheduled_at < NOW() LIMIT 10 → call Gmail OAuth send (preferred — already wired in main CRM gmail-poller path) OR Brevo → update status=sent + log to main CRM email_communications. THIS unblocks actual outreach.
6. Tone-matching from sent emails. Per-user: ingest last N sent emails from Gmail (already have OAuth) → embed style → condition every drafter prompt with style profile. Endpoint POST /api/mp/users/:uid/learn-tone.
7. Hiring/news/funding signal layer. Per-lead at research time: scrape company's recent job postings (LinkedIn jobs / Greenhouse / Lever public) + Crunchbase news → include in dossier "Recent Signals" section + use in personalization vars.
8. Reply classifier → auto-Calendly. We already have sentiment classifier (shipped UD-P3). Add: positive sentiment + has booking-intent → auto-reply with Calendly link + write meeting back to user calendar via Google Calendar API.
9. Multi-step LI+Email orchestration with branching (depends on LinkedIn channel #10).
10. LinkedIn channel — Playwright per-user, burner-first, free LI account, hard cap 8/day. 10-14 days. See dedicated LinkedIn section below.
11. Daily performance brief email at 9PM (S effort).
12. Pending AI Reviews UI fix on Skynet Overview (~20min).

🔑 ENV / CREDENTIALS STATE

• DEEPSEEK_API_KEY — set in /Smrt-CRM-FINAL/.env + /skynet-dev/backend/.env ✅
• MIMO_API_KEY — set ✅
• ANTHROPIC_API_KEY — NOT FOUND in env files. User MEMORY says wired. Need to locate or get from user.
• APIFY_API_KEY = apify_api_3O5BEtvS4ySb7J5qceoWbpEmO279bE4wr5hf — capped till 2026-05-09
• GMAIL_CLIENT_ID + GMAIL_CLIENT_SECRET + GMAIL_REDIRECT_URI — set on smrt-crm-dev ✅ (added 2026-04-25)
• Need to acquire: Bright Data API token (~$50/mo signup) for OpenClaw scraping
• LinkedIn: nothing (will use FREE account + browser cookie capture, no Premium)

📋 MAXXL FILE INDEX

• /skynet-dev/backend/mp-routes.js — all backend endpoints
• /skynet-dev/backend/mp-helpers-v2.js — drafter v2, sig fetch
• /skynet-dev/backend/skynet.db — 12 mp_ tables
• /skynet-dev/backend/server.js — line ~3994: mounts /api/mp
• /Smrt-CRM-DEV/dist/public/assets/index-Dtu0_nu2.js — UI overlay v5
• /Smrt-CRM-FINAL/.env — main CRM secrets
• /skynet-dev/backend/.env — Skynet DEV secrets (DeepSeek + Mimo)
• /Smrt-CRM-DEV/.env — DEV CRM secrets
• /root/crm-patches/ — all ledger patch scripts

⚙️ HOW TO TEST MAXXL ON DEV (quick start)

1. https://dev.pros11.com/skynet → click "⚡ MAXXL" tab (10th)
2. 📋 Campaigns → "+ New Campaign" → 2-path modal opens (AI / Manual)
3. Click Manual → 4-step wizard fills all MoneyPrinter fields (name/instructions/Email/LI toggles/seq counts/LI note → Targeting w/ AI generate buttons → Sequence preview AUTO-DRAFTED with vars → Launch toggles)
4. Existing test campaign: id=4 "My Test" status=active. Click 🔄 Auto-Add Now → currently returns 0 contacts (Apify cap; need #1+#2+#4 P0 fixes)
5. OR add test contact via API: curl POST /api/mp/contacts/test-id/research with body fields
6. Click 🔬 on any contact → full research drilldown panel slides in
7. Per-step ✨ Edit with AI → inline popover (no more prompt())

12 deep-dive items captured:

1. Per-step Edit pencil = "Edit with AI" popover (textarea: "What changes would you like to make?" + Generate button + voice mic). AI rewrites that step from prompt.
2. Per-step delay edit = inline number stepper (−/+ days, ✕/✓). 0 = Immediately.
3. Email signature auto-appended to every generated email (full Smrt Stats signature with phone/disclaimer/legal footer).
4. Campaign Analyze modal = AI reviews entire campaign + suggests changes. Tabs: Current Analysis / History. Footer: "Apply N Changes" review-and-approve.
5. Campaign Insights = Outreach Overview slide-up — per-campaign chart of initial vs follow-ups by sender, channel filter (Email/LinkedIn), 7d/30d range.
6. Research Report (per-contact) = AI-generated dossier with: Contact Info (Name/Role/Headline/Location/Connections/LinkedIn/Why-this-person) + Summary (long-form 1st-person career narrative) + Company Info (Name/Website/Founded/Employees/Industries/About/Key Roles) + Funding (Total Investment) + Research Findings (Why-Tomasz-and-PZPN-Are-a-Good-Fit-for-Smrt-Stats — AI's reasoning + pain points listed).
7. Add Contacts modal = picker from existing CRM contacts. Tags filter + search + per-row Add/Added.
8. 3-dot campaign menu: Duplicate Campaign / Disable Campaign (red).
9. 3-dot per-contact menu (in campaign): Stop Sequence / Delete Contact / Delete and Blacklist Company.
10. AI Assistant conversational campaign builder — caught full chat: user asks "3 high-ROI ideas", AI proposes (pro-club priced-out-of-Wyscout + agencies + federations), iterates ("worldwide football federations except Swiss/German"), refines ("direct federation use vs federations enabling clubs"), creates the actual campaign as proposal user can approve.
11. Conversations filters: Channel / Date / Response Type. Inbox / All tabs.
12. Lead Scraper mode breakdown: Mode 1 (find people at company) = standalone modal w/ Company Name + Website OR LinkedIn URL + Role description + N contacts (1-20) + 0.1 cred/contact. Mode 2 (ICP) + Mode 3 (jobs) = redirect to Create Campaign (filters live in campaign editor's Targeting). Mode 4 (people CSV upload) = Enable Contact Enrichment toggle + minimal CSV (full_name + email/linkedin_url/website) + AI auto-enriches missing + optional NL filter ("Only extract executive founders") + 0.1 cred/lead. Mode 5 (company CSV) = name + linkedin_url/website, max 500/upload, optional NL filter ("Don't extract from recruiting firms").
13. Advanced Campaign Settings = per-campaign blacklisted companies (vs global blacklist in Org settings).
14. Sender selector dropdown = mailbox rotation per contact (fernandez@smrtstats vs ivan@smrtstats).
15. Contact GENERATING state (yellow badge) — AI personalization happens async per contact after they're added to campaign.
16. Notifications bell: empty popover ("You're all caught up!").

Architectural shift: all clone features go into Skynet (port 3010 dev / 3006 prod), not main CRM bundle. Skynet is the intelligence layer; CRM is the system of record. Org Knowledge Base (already shipped today as MPC-12 backend) lives in CRM since it's shared metadata.

Where: dev.pros11.com/skynet → click "⚡ MAXXL" tab. Backend: Skynet DEV port 3010 at /api/mp/*. Frontend: IIFE in main CRM bundle, scoped to body.maxxl-skynet-page.

Architecture: All MAXXL features in Skynet (intelligence layer). Reads org_knowledge from main CRM postgres directly. CRM = system of record.

Built (v5):

• ✅ Schema (12 mp_ tables in skynet.db)
• ✅ AI sequence drafter (DeepSeek, var-aware: pulls extracted vars first, drafter only uses known keys)
• ✅ Per-contact research dossier (Smrt-Stats-aware via org_knowledge)
• ✅ Variable extractor + handshake fix
• ✅ Schedule with var resolution + auto signature append (users.email_signature)
• ✅ Per-step manual edit (textarea auto-save) + Edit-with-AI inline popover (no more prompt) + Delete + delay editor
• ✅ Per-campaign Analyze (AI score+issues+suggestions) + Per-campaign Blacklist UI
• ✅ AI Assistant tool-using chat (CREATE_CAMPAIGN, LIST, ANALYTICS)
• ✅ Sub-pages: 🏠 Home / 📋 Campaigns / 📨 Inbox / 🎯 Find Leads / 📧 Senders / 💬 Assistant
• ✅ Create Campaign modal — 2 paths (AI chat / Manual 4-step wizard exactly matching MoneyPrinter fields)
• ✅ Senders w/ Google OAuth button (reuses gmail-poller /api/auth/gmail/start)
• ✅ Inbox = unified replies (cross-campaign from main CRM email_replies)
• ✅ Find Leads 5-mode UI
• ✅ GENERATING worker (research-async flips status pending→generating→active)
• ✅ Horizontal Skynet tablist (no vertical wrap) + bulletproof sibling-panel hide
• ✅ Conversations port + Lead Scraper bridge endpoints + folders/senders/blacklist/analyses backend

⚠️ Known issues:

• Pending AI Reviews shows "0" on DEV (UI component query mismatch — DEV agent_contacts has 6,426 pending, identical to PROD; pre-existing bug, not MAXXL)
• Lead Scraper modes 4+5 (CSV uploads) = stub bridges, need wire-up to existing Skynet Pipeline Labs flow
• Lead Scraper mode 1 calls /api/skynet/scrape-company which doesn't exist yet — need Skynet route wrapping Apify Pipeline Labs Apollo actor

Data sources for MoneyPrinter: Bundle minified, vendor names stripped. Industry knowledge: most likely Apollo.io (300M+ profile DB, $0.04-0.10/contact), People Data Labs (enrichment), Hunter.io/Snov.io (email verification), Apify Sales Navigator scraper (LinkedIn profile data). SMRT CRM already has all wired in Skynet — MAXXL Lead Scraper bridges to existing Pipeline Labs flow.

Why deferred: single-session can't safely deliver multi-day project. Wrong shortcut = ban Ivan's professional LinkedIn account. Backend table mp_linkedin_accounts ready + endpoint stubs at /api/mp/linkedin/* placed.

Why build vs buy: Heyreach/Expandi/Dripify all use SAME browser-automation approach (no official LinkedIn API for messaging). Buying = same ban risk + $39-300/mo. Building = full control + free + integrates directly with MAXXL DB.

Plan (next session, 10-14 days):

1. Install Playwright + Chromium on srv1373075 (~500MB, headless via Xvfb)
2. Cookie capture: user pastes li_at from LI DevTools → encrypted in mp_linkedin_accounts.session_cookie_encrypted
3. Use FREE LI account (no $99/mo Premium needed — 12 connections/day cap is normal)
4. Burner-first: fresh @gmail account → 2 weeks testing → if survives, onboard real account
5. Worker microservice (Node + Playwright, separate PM2 process)
6. Hard cap 8/day connections (less than LI's 12 soft cap = safety margin)
7. Anti-detection: random 15-90s delays, mouse jitter, viewport randomization
8. Connection-state poller (/mynetwork/invitation-manager) every 30min
9. Inbox poller (/messaging/) every 15min → ingest to email_replies (source='linkedin')
10. Kill switch: detect "Are you a robot?" → freeze + alert
11. Multi-account team: each member adds own LI → combined 8/day × N capacity
12. 2 full weeks burner testing before promoting to real accounts

Open source / cheap alternatives (rejected):

• Linked Helper $15/mo desktop — runs on user's PC, lower ban risk, but can't integrate with MAXXL DB
• Apify linkedin-sales-navigator-scraper $0.10/profile — read-only LEAD SOURCING, no messaging
• tomquirk/linkedin-api Python free — unofficial private endpoints, breaks frequently, highest ban risk, NOT recommended
• Dripify $39-89/mo — same automation, same risk, costs money

What it is: AI SDR / outbound sales engine. Generates per-lead personalized email + LinkedIn + cold-call sequences. Human-review-and-approve gate by default. Built on top of: lead scraper (300M+ profiles claimed), Google Workspace mailbox marketplace ($3/mo + $1/mo warming), LinkedIn account linking (12 req/day per user), parallel dialer with AI live teleprompter, conversational AI assistant.
User loves: per-contact AI personalization with custom variables ({{Federation Name}}, {{regionorspecificcompetition}}) extracted from research; Email + LinkedIn coordinated cadence; Edit-with-AI / Regenerate buttons on every sequence; AI chat side panel inside campaign editor; org-level "Detailed Analysis" knowledge base feeds personalization; intent signals (job postings, competitor followers).
Pricing tiers: Startup $250 (375 leads/mo, 1125 sequences) → Basic $800 (1200 / 3600 + AI research per lead) → Standard $2000 (3750 / 11250 + auto enrich + 3 competitors) → Premium $3000 (6000 / 18000 + 5 competitors + Slack support) → Enterprise (custom).
Stack: frontend Next.js App Router on Vercel · backend Express on Azure App Service · Segment analytics · custom auth (Bearer rgr_live_*) · public API POST /apis/v1/campaigns for contact ingest only · HubSpot integration native.

Verdict: SMRT CRM already has ~50% of the core data model (contacts, deals, email_communications, email_replies, brevo_events, gmail_oauth_tokens, campaigns, pipeline_rules, sentiment classifier just shipped). The two BIG gaps are: (A) LinkedIn outreach channel and (B) AI sequence personalization with custom-extracted variables. Phone dialer (MP-7) and managed-mailbox marketplace (MP-8) are out of scope (heavy infra). HubSpot integration (MP-14b) is low-priority since SMRT IS the CRM.
Effort scale: S = 1-2 days · M = 3-7 days · L = 1-3 weeks · XL = 1-2 months · XXL = quarter+. All built ledger-patch style on DEV first.

Recommended build order (90-day plan):

1. Week 1-2: MPC-12 Org knowledge base (small, unblocks AI personalization)
2. Week 1-2 parallel: MPC-15 Edit-with-AI buttons everywhere (immediate user-visible win, low risk)
3. Week 2-3: MPC-2 Campaign folders + multi-channel stats refactor
4. Week 3-6: MPC-5 AI personalization with extracted variables (the core magic — biggest piece, builds on MPC-12)
5. Week 6-9: MPC-3 + MPC-4 Campaign editor wizard + run view (depends on MPC-5)
6. Week 9-12: MPC-13 AI Assistant chat (global + side-panel)
7. Week 6-13 parallel track: MPC-9 LinkedIn channel (Playwright per-user) — slowest, riskiest, can ship in parallel
8. Week 12-14: MPC-10 Lead scraper consolidation (mostly UI on existing Skynet)
9. Week 14: MPC-14 public ingest API + final polish

Skipped: MPC-7 Parallel dialer (cold calling not in SMRT scope) · MPC-8 mailbox marketplace (SMRT not multi-tenant SaaS) · HubSpot bidirectional sync (SMRT replaces HubSpot for user). All on DEV first per RULE #1.

Final state PROD: Email Dashboard 100% real data. 9 tabs single row no flicker. Reply classifier active (Brevo + cron paths). Sentiment-aware rules auto-move deals. Negative replies auto-suppress contact email.

Skipped: BPMN process designer · Service ticketing module · Marketing campaign designer · Quote/contract module · Marketplace · Mobile native app · Account hierarchy · Field-level encryption · IP restrictions · E-signature. Reasons: vertical mismatch with football scouting CRM, low ROI, or existing equivalent (RTK / Brevo / DocuSign elsewhere).