Software Updates

repair schema drift trinity + harden API routes against missing columns

• Three production bugs caused the symptoms (meeting list spinner-of-death,
• 'instant meeting' opens dead tab, join-route 500s):
• 1. MeetingInvite column-name drift
• create/schedule/list routes wrote inviteToken/guestEmail, but the actual
• table has token/email/guestName/joinCount/createdByEmail. Every invite
• INSERT was silently 500ing.
• 2. Meeting.consentAINotesAcceptedAt was never migrated
• Four routes (join, guest consent, yen-advisor, list) reference this
• column but no migration ever added it. SELECT blew up -> join route
• 500'd -> meeting tab appeared dead.
• 3. Meeting.kind casing inconsistency
• Legacy default was lowercase 'internal'; new code expects 'INTERNAL'.

open meeting tab synchronously to bypass popup blocker; persistent error toasts in room client

• startMeeting(): window.open('about:blank') BEFORE the await fetch() so browsers don't treat it as a programmer-initiated popup and block it. Navigate tab to meeting URL on success; close it on failure. This fixes the silent 'can't start meeting' symptom where the room was being created but the tab was blocked.
• meeting-room-client: replace toast.error() with errorToast() for the LiveKit connection-timeout and translation errors so they persist until dismissed, giving enough time to open a support ticket.

remove duplicate function body in schedule/route.ts (parse error)

Meetings overhaul: 3-category cards, instant attendees, user-scoped list, CSP fix, calendar on schedule

• CSP: add static.ads-twitter.com to script-src (clears console noise)
• meetings/list: filter to meetings where viewer is host, participant, or invited (MeetingInvite) — each user sees only their own meetings
• meetings/list: surface errors in UI (AlertCircle + Retry) instead of silently showing empty state on auth/DB errors
• meetings/create: accept kind + attendeeEmails; creates MeetingInvite per attendee + sends 'join now' email (fire-and-forget)
• meetings/schedule: calendar event created for host + Katura team member attendees (get-or-create default calendar); external attendees get email only; migrated to shared invite-email.ts helper
• src/lib/meetings/invite-email.ts: shared sendInstantInviteEmail + sendScheduledInviteEmail extracted from schedule/route.ts
• UI: replace single 'Start an instant meeting' card with 3 side-by- side category cards (Internal / Client Consultation / Vendor-Partner), each with 'Start instantly' and 'Schedule' buttons
• Instant launch modal: title + attendee chips + AI notes toggle, opens pre-keyed to selected kind
• Schedule modal: openScheduleModal(kind) pre-selects kind before open

Persistent error toast + support ticket system

• errorToast() replaces toast.error() across admin — duration: Infinity, never auto-dismisses, user must click X or open a support ticket
• Support ticket modal: slides in from right, pre-fills error + page URL + timestamp + user-agent via window custom event (no prop drilling)
• POST /api/admin/support/ticket — requireKaturaAdmin auth, sends Brevo email to support inbox with full error context + user's description
• SupportTicketModal mounted once in admin/layout.tsx
• All 8 toast.error() sites in meetings-client.tsx migrated to errorToast() with detail (raw API error) and source (meetings/create|end|schedule)

Remove orphaned code fragment in meetings-client (parse error at line 219)

Fix two build-breaking parse errors

• providers.ts:55 — two statements on one line (missing newline between const requested and const rawOverride)
• meetings-client.tsx:77 — unescaped double quotes inside double-quoted string (noUpcoming label), switched outer to single quotes

Swap OpenAI Realtime → DeepL for live meeting subtitles

• translate.ts: replace openai gpt-4o-mini with DeepL REST API (~50-150 ms vs 300-700 ms; purpose-built; no prompt engineering needed)
• free-trial keys (:fx suffix) auto-routed to api-free.deepl.com
• pro keys routed to api.deepl.com
• preserve_formatting:true keeps brand names, SKUs, numbers intact
• BCP-47 → DeepL lang code conversion (EN-US, PT-BR, ZH-HANS defaults)
• soft-fail on missing key or API error (shows source text, never crashes)
• recordSpend updated: model='deepl', cost at $25/M chars rate
• all three cache layers (process LRU + DB) untouched — cache hits are free
• use-realtime-translation.ts: replaced with no-op stub (same interface, zero behavior) — meeting-room-client compiles without changes
• realtime-session/route.ts: returns 410 Gone with explanation
• The SpeechRecognition → translateCaption → DeepL path that was already running in parallel now handles all subtitle translation. No duplicate audio WebSocket path, no PCM16 streaming, no ephemeral token overhead.
• OPENAI_API_KEY downgraded to warn (still needed for AI minutes)
• DEEPL_API_KEY added as a required check with free/pro tier detection

Fix transcript gap + track translation cost + AI usage page under Treasury

• meeting-room-client.tsx: remote caption.final now fire-and-forgets POST to /api/admin/meetings/[id]/transcript so guest speech reaches the DB before generateMinutes() runs — single root cause of blank AI notes
• lib/meetings/translate.ts: wire recordSpend() after every live OpenAI gpt-4o-mini call; source='meeting-translate' for per-area breakdown
• treasury/ai-usage: new page at /admin/treasury/ai-usage with budget gauge, sparkline, by-source/model/task breakdown, raw spend log, historical month nav, blind-spot notice for Realtime sessions

Inquiry Triage Agent (Priority 1 — Virgil's build order)

• RATIONALE (Virgil): 'Every inbound inquiry gets classified: hot lead, warm
• lead, cold lead, support issue, spam, urgent, needs Stella, needs Rosser.'
• Estimated return: $3,000–$10,000/month from faster response + fewer lost leads.
• AI cost: ~$0.001 per inquiry (Haiku ROUTING). At 100 inquiries/month = ~$0.10.
• Effectively free.
• NEW prisma/migrations/add-inquiry-triage.sql
• · InboundInquiry table — captures any raw inbound contact regardless of source
• · Fields: source, rawText, sender metadata, triage lifecycle, Yen's full
• classification output (priority/urgency/category/dealReadiness/estimatedBudget/
• summary/suggestedReply/suggestedAction/routeTo/openQuestions/triageNotes)
• · Indexes on tenantId+status, priority+createdAt, senderEmail
• · Idempotent — safe to re-run
• NEW src/lib/yen/agents/inquiry-triage.ts
• · triageInquiry(input) — Haiku agent, structured JSON output
• · Classifies: priority (HOT/WARM/COLD/SUPPORT/SPAM), urgency, category,
• dealReadiness, estimatedBudget, summary, suggestedReply, suggestedAction,
• routeTo (stella/rosser/both), openQuestions, triageNotes
• · Fallback defaults on every field — partial response still saves, inquiry
• is never lost due to model returning bad JSON
• · Cost constant: TRIAGE_COST_ESTIMATE_CENTS = 0.09 (< 1/10th of a cent)
• NEW src/app/api/admin/yen/triage/route.ts
• · POST — submit raw inquiry, creates DB row immediately (PENDING → TRIAGING →
• TRIAGED), then runs agent. If agent fails, row stays PENDING with triageError.
• Never loses an inquiry.
• · GET — list inbox ordered HOT→WARM→SUPPORT→COLD→SPAM, then by createdAt.
• Filters: status, priority. Returns priority counts for sidebar badges.
• · PATCH — mark ACTIONED (with sentReply) or ARCHIVED
• · recordSpend wired (fire-and-forget)
• NEW src/app/admin/yen/triage/page.tsx
• · Two-panel inbox layout: list (left) + detail (right)
• · Priority badges with color coding: Flame (red/HOT), ThermometerSun
• (orange/WARM), Snowflake (sky/COLD), Wrench (violet/SUPPORT), Ban (spam)
• · Detail panel: summary card, budget signals, deal readiness, suggested action,
• open questions, collapsible raw text, collapsible triage notes,
• editable draft reply textarea, Mark actioned + Archive buttons
• · Submit modal: source picker, sender fields, paste textarea, triage on submit
• · Status filter tabs: Triaged / Pending / Actioned / All
• · Refresh button, empty states, loading states
• UPDATED docs/YEN-CONSULTATION-MODE-SCOPING.md
• · Inquiry Triage Agent marked shipped

Virgil ROI model + corrected build order

• NEW docs/YEN-ROI-MODEL-VIRGIL.md
• · Full ROI analysis from Virgil — verbatim reasoning preserved
• · Conservative monthly profit model: $11,500–$60,000 gross, $15k planning number
• · Per-source breakdown: outreach, inquiry handling, follow-up recovery,
• consultation prep, admin savings
• · Human-AI split doctrine: Yen prepares/drafts/follows up, Rosser/Stella close
• · Corrected build order: Inquiry Triage → Follow-Up → Prep → Intelligence → Slices
• · Risk table: robotic replies, over-automation, spam, budget, non-use
• UPDATED docs/KATURA-OPERATING-PLAN-2026-05.md
• · Section 4 rewritten with corrected priority order (4a–4f)
• · 4a: Inquiry Triage Agent — highest ROI, $3k–$10k/month, costs ~$1–2/month AI
• · 4b: Follow-Up Agent — cash register, $2k–$15k/month, daily cron
• · 4c: Consultation Prep Agent — higher close rate, $3k–$20k/month
• · 4d: Meeting Intelligence Agent — closes memory loop (was previously #1)
• · 4e: Slices 3–7 — consultation product features
• · 4f: Marketing spend — fund from AI budget headroom
• · Summary updated with Virgil's 13x–33x ROI target

operating plan 2026-05 + update consultation scoping status

• NEW docs/KATURA-OPERATING-PLAN-2026-05.md
• · Revenue sources in order of immediacy
• · Full sprint summary: what was built and why (money logic for each)
• · Real AI cost numbers per consultation (Haiku advisor: ~$0.05-0.20/session)
• · Next build priorities in revenue-impact order: Meeting Intelligence Agent,
• Consultation Prep Agent, then Slices 3-7
• · Marketing budget framing: headroom below $750/month → traffic + lead capture
• · Technical debt and risks table with severity ratings
• · Decision log: locked decisions + rationale (don't revisit without reason)
• · Migration runbook for add-memory-graph.sql + Vercel env vars
• · One-paragraph summary
• UPDATED docs/YEN-CONSULTATION-MODE-SCOPING.md
• · Status table: Slices 0/1/2 + memory arch marked shipped with commit hashes
• · Slices 3-7 + agents marked as pending with correct order

wire cost guard + model router into Yen chat; revised budget thresholds

• REVISED cost-guard.ts — 3-tier thresholds (per Rosser/Virgil budget review)
• · Soft warning: $650/month (65%) — logs to console + dashboard flag
• · Approval required: $850/month (85%) — BudgetStatus.approvalRequired = true
• · Hard stop: $1,000/month (100%) — AI calls blocked, 402 response
• · Normal target: $750/month (leave headroom for marketing spend)
• · BudgetStatus now includes approvalRequired boolean
• WIRED yen/chat/route.ts
• · Imports checkMonthlyBudget, recordSpend, getModelForTask
• · Budget check before every agentic loop — returns 402 with plain-English
• message if exhausted (tells Rosser to raise YEN_MONTHLY_BUDGET_CENTS)
• · recordSpend() fire-and-forget after each turn — taskClass=CORE_ASSISTANT
• · effectiveModel fallback for Anthropic now uses getModelForTask('CORE_ASSISTANT')
• → claude-sonnet-4-5 by default, not the raw provider.model string
• UPDATED docs/YEN-MEMORY-ARCHITECTURE.md
• · Revised budget table: $250–400 Yen core, $50–125 extraction,
• $100–250 translation, $25–75 embeddings, $100–200 agents, $100–150 reserve
• · Per-agent cost limits table (Consultation Prep $2/run, Summary $1.50,
• Curator $0.50, Invoice $0.25, Deep Strategy = approval required)
• · Framing: AI as leverage not vanity — extra budget → traffic + lead capture

Yen memory architecture — model router, cost guard, graph schema

• Based on Virgil's architecture review. Budget constraint: ≤ $1,000/month.
• NEW src/lib/yen/model-router.ts
• · TaskClass enum: EXTRACTION | SUMMARY | ROUTING | CORE_ASSISTANT |
• CLIENT_WRITING | HIGH_STAKES
• · EXTRACTION/ROUTING/SUMMARY → Haiku (cheap + fast)
• · CORE_ASSISTANT/CLIENT_WRITING → Sonnet (best balance)
• · HIGH_STAKES → Sonnet by default; Opus only with YEN_ALLOW_OPUS=true
• · Per-task cost constants; estimateTaskCost() helper
• · Model name overrides via YEN_MODEL_HAIKU / _SONNET / _OPUS env vars
• NEW src/lib/yen/cost-guard.ts
• · Hard monthly spend ceiling: YEN_MONTHLY_BUDGET_CENTS (default 100000 = $1,000)
• · checkMonthlyBudget(tenantId, estimatedCents) — blocks if over limit
• · recordSpend(opts) — append-only ledger write after each AI call
• · getMonthlySpendSummary() — dashboard-ready breakdown by task class
• · Alert logs at 75% and 90% of budget
• · Fail-open if YenTokenSpend table doesn't exist yet (migration safety)
• NEW prisma/migrations/add-memory-graph.sql
• · MemoryNode — extracted memory facts (scope/type/confidence/importance/
• accessLevel/expiresAt/embedding slot ready for pgvector)
• · MemorySource — source traceability; every node must cite its origin
• · MemoryEdge — typed graph (supports/contradicts/depends_on/supersedes/
• must_not_violate + 7 more relationship types)
• · MemorySummary — rolling summaries per user/project/client/consultation
• · YenTokenSpend — append-only AI spend ledger (cost-guard reads this)
• FIXED src/app/api/admin/meetings/[id]/yen-advisor/route.ts
• · Wrong callAnthropic() signature (was passing messages/tools/prompt in
• wrong order — endpoint would have thrown on first real call)
• · Now uses getModelForTask('EXTRACTION') → Haiku model (saves ~12x vs Sonnet)
• NEW docs/YEN-MEMORY-ARCHITECTURE.md
• · Virgil's full 4-layer system captured as living reference
• · Model allocation table, cost estimates, tenant isolation rules,
• agent hierarchy, canary memory guidance, priority build order

Slice 2 — Yen Advisor Sidebar

• MODIFIED GET /api/admin/meetings/[id]/join · SQL now returns meetingKind (COALESCE kind, 'INTERNAL') and consentAINotesGranted (consentAINotesAcceptedAt IS NOT NULL) · Response includes meetingKind + consentAINotesGranted
• NEW POST /api/admin/meetings/[id]/yen-advisor · READ_ONLY: no writes, no tool calls · Guards: requireKaturaAdmin + kind === CONSULTATION + consentAINotesAcceptedAt set · Sends transcript to Claude with structured extraction prompt · Returns { stones, budget, style, openQuestions, summary } · Graceful JSON parse fallback — never throws 500 on bad model output
• MODIFIED meeting-room-client.tsx · JoinData: +meetingKind, +consentAINotesGranted · RoomInner props: +meetingKind, +consentAINotesGranted · State: +showYenSidebar · Sparkles button in control bar — CONSULTATION-only, disabled if no consent (toast explains why rather than silently doing nothing) · YenAdvisorSidebar component: polls every 12 s while open, renders summary card + stone prefs + budget signal + style + still-to-ask · Footer: timestamp + 'Not visible to client' notice · Transcript toggle closes Yen sidebar; Yen toggle closes transcript

Slice 1 — consultation consent gate

• NEW PATCH /api/guest/meetings/[token]/consent · Validates invite token, checks meeting.kind = CONSULTATION · Writes consentRecordingAcceptedAt / consentTranscriptAcceptedAt / consentAINotesAcceptedAt as COALESCE timestamps (idempotent) · INTERNAL/VENDOR meetings return ok without writing
• MODIFIED GET /api/guest/meetings/[token]/join · Returns meetingKind + consentAINotesGranted bool · Fixes column reference: i.inviteToken → i.token
• MODIFIED guest-room-client.tsx · Stage union adds 'consent' · State: meetingKind, consentRecording/Transcript/AiNotes, submittingConsent · useEffect routes CONSULTATION guests without AI consent to consent stage · submitConsent: PATCHes endpoint then advances to prejoin (non-fatal on error) · Consent interstitial: 3 independent checkbox cards + Accept/Skip actions

Slice 0 — scheduled meetings, kind picker, Upcoming tab

• ALTER Meeting: add kind (INTERNAL|CONSULTATION|VENDOR), customerId FK, scheduledStartAt/End, actualStartAt, three consentXxx timestamps, 3 new indexes; backfill kind='INTERNAL' for all existing rows
• CREATE ConsultationSession: meetingId (unique), customerId, detectedRequirements, shortlistJson, selectedDiamondId, cadBriefJson, personalizationNotes, sensitiveNotes, proposedInvoiceId, outcome
• CREATE ConsultationArtifact: sessionId FK, kind, payload (client-safe), internalPayload (never sent to client), audit columns (approvedByEmail, approvedAt, broadcastAt, rejectedAt/Reason)
• NEW POST /api/admin/meetings/schedule
• Accepts kind, scheduledStartAt/End, title, customerId?, attendeeEmails[]
• Validates future date, validates customerId against User table
• Inserts Meeting with status=SCHEDULED, inserts ConsultationSession if kind=CONSULTATION, inserts host MeetingParticipant
• For each attendee: creates MeetingInvite + sends Brevo email with .ics calendar attachment (fire-and-forget, non-blocking)
• UPDATED GET /api/admin/meetings/list
• Returns kind, scheduledStartAt/End, customerId, customerName, customerEmail
• JOINs User on customerId for customer display
• ORDER BY: LIVE first, SCHEDULED second, then by time DESC
• 'Schedule meeting' button (top-right) opens scheduling modal
• Modal: kind picker (3 options with icon + description), title, date/time pickers, customer search (debounced, only for CONSULTATION, auto-adds email to attendees on select), attendee chip input (Enter to add)
• Upcoming / Recent tabs with count badge on Upcoming
• Upcoming: SCHEDULED + LIVE meetings; links to /prep page for scheduled
• Recent: ENDED/ARCHIVED meetings; auto-switches if no upcoming
• KindBadge component: violet for CONSULTATION, sky for VENDOR, none for INTERNAL
• Customer name shown inline on meeting cards for consultations
• timeUntil() helper (in Xm / Xh / Xd) for scheduled meeting countdown
• i18n: all new strings in both EN + PT

harden /admin/meetings security+perf, add Yen risk-class taxonomy

• B.1 CRITICAL: lock down /api/guest/translate
• require inviteToken + validate against MeetingInvite (expiry + status)
• per-token + per-IP rate limits
• CORS lock to katura1999.com / *.vercel.app / dev localhost
• use real invite.meetingId for cache bucketing (kill 'guest_shared' hack)
• B.3: escape HTML in invite email body (guestName/hostName/title/email)
• B.4: realtime-session route now binds to [id] meeting, validates active, attaches metadata (meetingId/adminEmail), audit-logs to MeetingTranslationSession
• B.5: cap MeetingInvite joins at 5; force-expire all invites on End meeting
• B.7: random per-session LiveKit identity (randomBytes 8) — no inviteId leak
• B.8: hashed guest chat identity via XOR-fold (no token prefix exposure)
• B.9: tight Permissions-Policy on /g/:token* in next.config.ts (camera/mic/display-capture self only; X-Frame DENY; noindex)
• B.10: rate-limit invite GET endpoint per-IP
• B.11: Brevo emails tagged [meeting:id, invite:id] for searchable audit
• A.2: SR interimResults=true + interim caption render (instant feedback)
• A.7: skip MeetingTranslationCache write for utterances <= 120 chars
• New: prisma/migrations/add-meeting-translation-session.sql
• New: src/lib/security/escape-html.ts (shared escaper)
• src/lib/rate-limit.ts: 5 new RATE_LIMITS entries for meetings
• src/lib/yen/risk-classes.ts: 6-class taxonomy (READ_ONLY / REVERSIBLE_INTERNAL_WRITE / EXTERNAL_NOISY / FINANCIAL / IDENTITY / DESTRUCTIVE) + CLASS_POLICY + TOOL_RISK_CLASS for every shipped + Tier-1 tool + SUBAGENT_DEFAULT_BUDGET (canCreateSubAgents:false hardcoded)
• src/lib/yen/sensitivity.ts: tag() / redactForModel() / redactForAdmin() for field-level visibility labels (public/internal_only/never_model)
• src/lib/yen/outbound-scrubber.ts: pre-approval pattern scanner (SSN, Luhn-valid cards, API keys, internal markers, cost-basis terms, supplier refs, HR notes) — scrubOutboundMessage + scrubArtifact
• src/lib/yen/guardrails.ts: evaluateGuard now falls through to CLASS_POLICY default (BLOCK/GATE/ALLOW) after explicit SOFT_GATED check
• YEN-CLAUDE-CAPABILITIES.md (9-section capabilities + roadmap doc)
• MEETINGS-TRANSLATION-AUDIT.md (perf A.1-A.11 + security B.1-B.12)
• YEN-VIRGIL-REVIEW-RESPONSE.md (point-by-point response with code evidence)

Claude/Anthropic provider, tracking dialog, hotkey spinner, email help button

• Yen: Add Anthropic (Claude) as a first-class provider
• New 'anthropic' case in providers.ts (claude-sonnet-4-5 default)
• src/lib/yen/anthropic-adapter.ts — full bidirectional translation between OpenAI-compat format used internally and Anthropic Messages API format (tools, tool_use blocks, tool_result, image parts, system prompt)
• chat/route.ts: branches on provider.name==='anthropic' → callAnthropic() instead of fetch to /chat/completions; all guardrails/approval gating unchanged
• Validates ANTHROPIC_API_KEY at request time with clear error message
• Set YEN_PROVIDER=anthropic in Vercel env to activate
• System prompt updated: identifies as Claude-powered
• Email: Wire up tracking details dialog (T-TRACK-01)
• thread-view.tsx: import EmailTrackingDialog; add trackingMsgId state; pass onOpenTracking to each MessageCard; render EmailTrackingDialog at bottom
• MessageCard: accepts onOpenTracking prop; shows EmailStatusBadge with onClick next to timestamp — click badge to open full timeline (opens, unique opens, forwards, IP, device, location, time-after-send)
• Cmd+/ support hotkey: instant open with screenshot loading indicator
• Modal now opens IMMEDIATELY on keypress (setOpen(true) before capture)
• Added 'capturing' state; shows animated spinner 'Capturing screenshot…' in the modal header while html-to-image renders in the background
• Spinner disappears automatically when screenshot arrives or times out
• Email: Add '?' help button explaining all features
• email-shell.tsx: HelpCircle button in top search bar → opens Dialog
• Covers: AI Summary, Translate, Smart Replies, To Task, Tracking, Print, Compose (C shortcut), Schedule Send, Undo Send, Search
• HelpSection component at bottom of file (icon + title + description)
• Yen page: sends pageContext (url, kind, label) with every chat message

AI email translation — inbound + outbound (T-FE54K)

• POST /api/admin/email/[messageId]/translate Detects language of last inbound message and translates to English (or any target language) via gpt-4o-mini. Single call handles both detection and translation. Returns detectedLanguage, detectedLanguageName, translatedText, isAlreadyTargetLanguage.
• POST /api/admin/email/translate-outbound Translates an admin-written reply from English to a target language before sending. Supports pt-BR vs pt-PT distinction. Returns translatedHtml and translatedSubject.
• thread-view.tsx: Translate button in toolbar (Globe icon, next to AI Summary). Clicking fetches translation and renders a two-column panel: original on the left, English on the right. Below the panel, two reply buttons: 'English (they receive English)' and '[Language] (they receive [Language])'. The latter opens compose with translateToLanguage set.
• compose-dialog.tsx: ComposePrefill gains translateToLanguage field. When set, handleSend calls translate-outbound before /email/send, swaps body+subject with translated versions, and shows a 'Translating…' spinner + '🌐 Sending in Portuguese' badge in the footer so the user always knows which language will actually be delivered.

build error + meeting cam/mic auto-activate + signatures use-case doc

• fix(build): correct import path in saturday-request/route.ts sendBrevoTransactionalEmail lives in @/lib/marketing/brevo-campaigns, not @/lib/brevo — was causing Turbopack build failure for last 4 deploys
• fix(meetings): auto-activate camera+mic when permission previously granted (T-29244) Two root causes: 1. Permissions-Policy: 'camera=(), microphone=()' blocks getUserMedia in Firefox at the top-level document (Chrome ignores this; Firefox enforces it). Changed to 'camera=(self), microphone=(self)' — allows same-origin access, still blocks cross-origin iframes. 2. Admin <LiveKitRoom> was missing 'video' and 'audio' props. Guest room always had them. Added both — browser auto-activates silently if permission was previously granted, prompts once if not.
• docs: add KATURA-SIGNATURES-USE-CASES.md — 29 extended use cases for Katura Signatures with full implementation details and priority matrix

Katura Signatures — internal e-sign + Stripe payment system

• SQL schema: SignatureTemplate, SignatureDocument, SignatureField, SignatureAuditLog with 10 seeded templates (custom deposit, engagement ring, lab diamond, natural stone release, employee W-4, contractor W-9, bench jeweler, casting house, wholesale, NDA)
• Domain layer: types, tokens, repo, Stripe integration
• Public API: GET/POST /api/signatures/[token]/{sign,decline,pay}
• Admin API: GET/POST /api/admin/signatures + [id]/send + /templates
• Public UI: /signatures/[token] — single-page sign + pay with embedded Stripe Payment Element and TYPED/DRAWN/UPLOADED signature methods
• Admin UI: /admin/signatures list/new/detail with audit log
• Webhook: extends /api/webhooks/stripe to route signature payments via metadata.katura_signature_id (falls through to existing handlers if not a sig)

CC icon, gear settings panel, 7 fonts, draggable PiP, iOS subtitles fix

• Replace Brain icon with Captions (CC) icon from lucide-react
• Add Settings gear button (visible while subtitles active)
• Subtitle settings panel: all-caps toggle (default ON), 7 font options (Katura Display, Sans, Serif, Mono, Rounded, Condensed, Humanist), 5 size options (S → 2XL) — styles applied imperatively via useEffect
• Draggable self-view PiP: FaceTime-style pointer-drag with corner snapping (BR → BL → TR → TL) using setPointerCapture + onPointerUp release detection
• Fix iOS subtitles: remove disabled state from CC button on iOS; toggleAiNotes now enables subtitle display on iOS (remote captions still work via LiveKit data packets) without attempting to start unsupported local STT

Phase 1 invoice engine — commercial spine (no Stripe)

• prisma/add-sales-invoices.sql: 5 tables (SalesInvoice, SalesInvoiceLineItem, SalesInvoiceAgreement, SalesInvoicePayment, SalesAgreementTemplate), 5 enums, global sequence sales_invoice_number_seq, 7 seeded EN agreement templates
• types.ts: enums, status colors, computeTotals (bp-based tax), formatMoney, STATUS_TRANSITIONS state machine
• numbering.ts: KAT-YYYY-NNNNN via Postgres sequence
• repo.ts: createInvoice, getInvoiceById, listInvoices, listLineItems, listPayments, recordManualPayment, markSent, markViewed, voidInvoice, getActiveAgreementTemplate
• GET/POST /api/admin/invoices
• GET/DELETE /api/admin/invoices/:id
• POST /api/admin/invoices/:id/payments (manual; $25K approval gate)
• POST /api/admin/invoices/:id/send
• /admin/invoices (list + KPI strip: outstanding, paid this month, drafts, overdue)
• /admin/invoices/new (60-second builder: type, customer, line items, tax rate, notes — save draft / save & send)
• /admin/invoices/:id (detail: line items, payment history, record manual payment form, void/send actions)
• /admin/invoices/:id/print (print-stylesheet view, Cmd-P)
• /[locale]/account/invoices
• /[locale]/account/invoices/:id (email-gated: owner only)
• 'Invoices' entry under Sales & CRM group
• 'New Invoice' quick action in command palette
• admin-i18n-provider: salesInvoices + newSalesInvoice keys
• docs/invoicing-architecture-2026-05-10.md (Virgil memo, 4-phase plan)
• docs/sales-priority-consultation-positioning-2026-05-10.md

consultation engine, sales OS, meetings reliability+latency

• Sales region routing with GDPR consent + 2-hr lead time
• Slot engine: Sunday closed, Saturday request-only, priority slot (1/day for 3 days, 45min, $100 surcharge — 'Priority Consultation' scarcity, not fake-busy)
• Inspiration upload to Supabase, ICS calendar invites, reminders cron
• Consultation cancel/reschedule by token, Saturday-request endpoint
• /admin/sales hub + regions CRUD + API
• /admin/meetings/[id]/prep brief, /admin/meetings/health page
• Branded force-end AlertDialog (no more browser confirm)
• Interim flush via navigator.sendBeacon on rec.onend — fixes 'no notes were taken' bug when host force-ends mid-utterance
• Empty-transcript fallback explains the cause + chat-only recap
• /api/admin/meetings/[id]/regenerate-minutes route
• TranscriptHealth admin diagnostic panel on /minutes page
• LiveKit data-packet caption broadcast (lossy interim, reliable final) so REMOTE viewers see captions in real time, bypassing Postgres
• Debounced + token-cancelled interim translation per viewer
• In-memory LRU translation cache + cross-meeting DB lookup
• Direct browser→OpenAI realtime translation WebSocket via ephemeral token (audio never traverses Vercel)
• /api/admin/people/search backend with employees > contractors > clients ranking, @username + prefix matching, multi-channel hints
• meetings-latency-deep-dive-2026-05-10.md
• meetings-invite-search-2026-05-10.md
• MEETINGS-STRATEGY.md updated

timestamps+timeline, smart end/leave, PiP, screen share picker, grace period rejoin, MEETINGS-STRATEGY.md

camera OverconstrainedError on Chrome, minutes from transcript regardless of aiNotesEnabled flag, Brain button disabled in Firefox

6 live-call bugs — camera zoom, iOS STT, name contrast, end-call lag, chat Realtime SQL, fire-and-forget end API

post-closure reply bell, support@ sender fix, address table migration, mailbox signature

FaceTime PiP layout, C-SPAN captions, visible control buttons

autofill fixes, postal code lookup, country-first form

• SavedAddressesSection: when user has no saved addresses and clicks 'Add address', the form now fetches /api/user/recent-order-address and pre-fills all fields (firstName, lastName, address, city, state, ZIP, country, phone). Label defaults to 'Primary' and isDefault=true.
• Empty state message updated with 'Add your first address — we'll try to pre-fill it from your order history' link text.
• New API route: GET /api/user/recent-order-address
• Checks the most recent Order.shippingAddress JSON (normalizes across multiple field-name conventions: address1/addressLine1/line1, etc.)
• Falls back to User model denormalized fields (addressLine1, city, provinceCode, countryCode, zipCode, addressPhone) — these are populated by Shopify import + default-address saves.
• Returns { suggestion: {...} | null }, always 200 (never breaks the form)
• prefillAddress() was reading addr.address / addr.apartment / addr.zip but /api/user/addresses returns addressLine1 / addressLine2 / postalCode.
• Corrected all three field names. Addresses now actually fill the form.
• session useEffect now splits session.user.name into firstName / lastName and fills the first/last name inputs when they're empty.
• Works even when the user has no saved addresses (Shopify customers, etc.)
• prefillAddress() now has a second fetch to /api/user/profile when no saved addresses exist, using firstName, lastName, addressLine1, addressLine2, city, provinceCode, countryCode, zipCode, addressPhone.
• /api/user/profile GET extended to return all address fields.
• Phone is set from saved address, then profile phone, then addressPhone.
• The currency-based setCountry useEffect is now guarded by sessionStatus: if the user is authenticated, we skip the currency guess entirely and let prefillAddress() set the country from their saved address.
• For unauthenticated guests, BRL→BR, MXN→MX, EUR+pt→PT, etc. still apply.
• Country selector moved to the TOP of the shipping address form (above name fields) so the form adapts immediately when country changes.
• For BR, US, and CA the ZIP/CEP input is shown ABOVE the address lines with an autofill note ('Cidade e estado serão preenchidos...' for PT, English equivalent for US/CA).
• On blur of the postal code input, handleZipLookup() fires:
• Brazil: calls ViaCEP (viacep.com.br/ws/{CEP}/json/) — fills city (localidade), state (uf), and address line 1 (logradouro) if empty.
• USA: calls Zippopotam (api.zippopotam.us/us/{zip}) — fills city and state abbreviation.
• Canada: calls Zippopotam (/ca/{postal}) — fills city and province.
• A green ✓ Autofilled badge appears when lookup succeeds.
• For other countries, the ZIP input stays in its traditional position (city | ZIP row) with no autofill behaviour.
• State/province select now rendered inline in the city row (not a separate row), keeping the form compact.

in-room chat panel, AI notes toggle, smart subtitle defaults

• New reusable MeetingChatPanel component (src/components/meetings/meeting-chat-panel.tsx)
• Right-side collapsible panel with unread badge on the floating bubble
• Supabase Realtime (postgres_changes INSERT) for live message delivery
• Read receipts: marks messages read on panel open (admin rooms only)
• File uploads to Supabase Storage bucket 'meeting-files', images shown inline
• Link auto-detection: URLs rendered with external-link icon
• Enter to send, Shift+Enter for newline
• New MeetingChatMessage Prisma model (prisma/schema/meetings.prisma)
• Fields: id, meetingId, senderIdentity, senderName, type (text|file|link), content, fileName, fileUrl, fileSize, fileMime, readBy[], createdAt
• SQL migration ready at prisma/add-meeting-chat.sql
• New API routes:
• GET/POST /api/admin/meetings/[id]/chat — admin auth via requireKaturaAdmin
• POST /api/admin/meetings/[id]/chat/read — marks messages read by identity
• GET/POST /api/guest/meetings/[token]/chat — guest auth via invite token
• Wired into admin room (meeting-room-client.tsx) with admin chat endpoints
• Wired into guest room (guest-room-client.tsx) with guest chat endpoints
• Guest identity = 'guest_<first8charsOfToken>', read receipts disabled
• meetings-client.tsx: new 'Enable AI notes' checkbox (default: true)
• When disabled, notice text updates to explain no transcript/summary generated
• end/route.ts: respects aiNotesEnabled flag from Meeting row
• If disabled, skips generateMinutes() and returns a placeholder summary
• If enabled, includes chat messages in the AI prompt under '## In-Meeting Chat'
• summarize.ts: generateMinutes() now accepts optional chatMessages param
• Chat messages formatted with timestamp, sender, and type label ([file]/[link])
• Appended to transcript prompt so Claude/GPT-4o summarises chat in minutes
• Admin room: subtitle auto-enable logic now uses host's defaultLanguage from join API response (previously hardcoded 'en' comparison)
• join/route.ts: now returns defaultLanguage alongside aiNotesEnabled
• Guest join API: now returns meetingId in POST response (needed by chat panel)
• Comment updated to explain the per-participant vs host-language comparison
• Run against production: psql "$DATABASE_URL" -f prisma/add-meeting-chat.sql or execute the SQL in Supabase dashboard SQL editor

resolution email dark theme, logo URL hardened, share dialog wider

• Outer background: beige (#f5f1eb) → white (#ffffff)
• Card background: white (#ffffff) → Katura navy (#0B1527), matching site footer
• All body copy, headings, reference line, divider, original message block updated to white text at varying opacities (rgba(255,255,255,0.85/0.60/0.45))
• 'Note from our team' block: dark (#fafafa) → translucent white (rgba overlay) with white text to match the dark card
• Logo src hardcoded to https://katura1999.com/brand/header-logo.png (absolute) instead of interpolating BASE_URL, which could resolve to localhost or a misconfigured env var and break image loading in email clients
• LOGO_URL hardcoded to https://katura1999.com/brand/header-logo.png (same fix — was dynamically built from BASE_URL which is unreliable at send time)
• DialogContent max-w-3xl → max-w-5xl so all form fields, toggles, and the share link table are visible without horizontal scrolling

eagerly create LiveKit room before issuing tokens to fix silent connect failures

stats i18n+currency, remove TrustBar, lock ticket priority, remove High option

• live-stats.tsx: add useTranslations('SoftwareUpdates') + useCurrency().formatPrice()
• all 12 stat labels/hints now use t() keys (en + pt)
• live firm cost now converts to user's active currency via formatPrice()
• methodology footnote translated
• messages/en.json + messages/pt.json: add SoftwareUpdates namespace (28 keys)
• [locale]/layout.tsx: remove TrustBar import + render (Lifetime Warranty bar gone)
• ticket-actions.tsx: remove editable priority select → read-only text display
• removed PRIORITY_OPTIONS const, setPriority state
• shows 'Set by reporter — cannot be changed' note
• platform/admin/support/page.tsx: remove 'High' SelectItem (Low/Normal/Urgent only)

repair corrupted meetings-client + polished support emails

• File was corrupted: JSX fragment injected into an i18n comment block at line 9, causing 'Unterminated regexp literal' parse error in Vercel
• Rewrote entire file with correct structure: all imports at the top, i18n dictionary and getAdminLocaleMeetings() in the module scope, then the MeetingsClient component — identical logic, no behaviour change
• All outbound support emails (new ticket, resolution notice, urgent IT alert) now hardcode sender as support@katura1999.com rather than inheriting from the SUPPORT_EMAIL env var (which was set to jewelryadvisor@katura1999.com in Vercel, causing emails to appear to come from the wrong address)
• Full polished HTML template replacing plain <div> layout
• Katura header-logo.png at top (public/brand/header-logo.png)
• Katura beige/cream card layout matching brand aesthetic
• Message block moved to top of card, directly below heading — as requested
• Metadata table (From, Source, Page URL, IP, Geo, Browser, OS, Device, Locale, Viewport, Submitted) below the message
• Muted warm-tone footer with reply-to note and admin link
• Full polished HTML template replacing one-liner htmlContent string
• Same Katura header-logo.png and card layout
• Replaces emoji (🚨) in subject line with plain [URGENT] prefix
• Red URGENT banner box at top of card (no emojis) with 20-min escalation note
• Message shown prominently below heading, metadata table below
• Sender changed from 'Katura IT Alerts' to 'Katura Support'
• Replaced text-only 'KATURA' wordmark with header-logo.png image
• Sender hardcoded to support@katura1999.com

roomCreate token grant + onError handler + full en/pt i18n

• livekit.ts: roomCreate: true added to token grant — silently required on some LiveKit Cloud plans; missing it caused all connection failures (NOT an env var issue — env vars were already correct in Vercel)
• MeetingRoomClient: onError on <LiveKitRoom> captures real error message into lkError state, shows immediately on diagnostic overlay (no 15s wait)
• lkError also rendered on the pre-join error screen
• MEETING_STRINGS: ~80 keys × en + pt
• getAdminLocale() reads admin-locale cookie (AdminI18nProvider sets this)
• RoomInner accepts lang prop; t() shorthand for all string lookups
• MeetingRoomClient passes lang to RoomInner
• Covers: overlays, toasts, title bar, invite panel, transcript panel, language pills, all button titles/aria-labels
• MEETINGS_STRINGS: en + pt — page header, feature pills, card, input, buttons, Live/Minutes-ready badges, recent list, loading/empty states, confirm dialog, all toast messages
• timeAgo() moved inside component to use lang-aware relative strings

loading skeleton + sidebar INP fix + mobile meeting room optimization

• Add src/app/admin/loading.tsx: skeleton shown instantly on nav click, eliminates the 2,200ms INP block (UI now responds immediately while server data loads in the background)
• Fix sidebar collapsible groups: replace inline maxHeight style with CSS Grid grid-rows-[0fr/1fr] trick — no more lint warning
• meeting-room-client.tsx mobile optimizations:
• h-dvh instead of h-screen (iOS Safari viewport fix)
• Detect isIOS / isMobile at module level
• Hide screen-share button on iOS (getDisplayMedia not supported)
• Controls auto-hide: 6s on mobile vs 4s on desktop
• Touch targets: 48×48px min (was fixed 56×56px regardless of screen)
• Icons scale sm:h-6 (slightly smaller on tight mobile screens)
• Transcript sidebar: full-width on mobile (w-full sm:w-80)
• Invite panel: bottom-sheet on mobile, top-right dropdown on desktop
• Control bar: pb-[max(env(safe-area-inset-bottom),2rem)] for iOS notch
• Control buttons wrap to rows on very small screens
• meeting room page.tsx: add Viewport export (device-width, no zoom, viewportFit=cover) — prevents accidental pinch-zoom in video calls

close tab on end/disconnect + connection timeout diagnostic

• On meeting end (red button) or unexpected disconnect: window.close() is called so the new tab closes itself. Falls back to location.href if window.close() is blocked (manually-opened tab).
• Added 15s connection timeout: if ConnectionState never reaches Connected, a full-screen overlay explains the most common cause (missing Vercel env vars) and shows a Close tab button. Also fires a toast immediately.
• Added serverUrl guard: if /join returns an empty or non-wss:// serverUrl, the client shows a clear error explaining which Vercel env vars are missing rather than silently handing an empty string to LiveKit.
• Connecting overlay now shows during initial connect phase (not just Reconnecting), so users see spinner instead of a blank black screen.

full-screen room, robust camera/mic, broadcast subtitles, per-viewer minutes translation

• /admin/meetings/[id] now bypasses the admin shell (no sidebar/header squeezing the room). Minutes pages still render with the shell.
• 'Start now' opens the meeting in a NEW TAB so the operator's admin context isn't lost. List rows: LIVE → new tab; ENDED → /minutes inline.
• Camera/mic/screen toggles now wait for ConnectionState.Connected before publishing, and surface specific errors: NotAllowed → 'Permission denied. Check browser settings.' NotReadable → 'Device in use by another app.' NotFound → 'No camera/mic found.'
• Fixed translation response key mismatch (res.translated → res.text) so live subtitles actually appear in the room.
• HOST-BROADCAST SUBTITLES: when any participant clicks the Languages button, a LiveKit data-channel message tells every other participant to enable their own per-viewer subtitle translation. Each client translates into THEIR OWN preferredLanguage — nobody loses context.
• Summary always renders in the host's (meeting.defaultLanguage) tongue.
• New <MinutesTranslator/> client widget detects when the viewer's browser locale differs from the host's, and offers a one-click 'Translate for me' that translates summary + decisions + action items + open questions + risks + follow-ups in parallel.
• Repo is private — copy fixed: 'public GitHub repository' → 'private GitHub repository'. JSDoc updated to require a fine-grained PAT with read-only Contents access.
• 'Updated every 5 minutes' → 'Refreshed continuously'. ISR cache bumped from 300s to 3600s now we no longer need the unauth IP limit.
• Verified line-count: regenerated repo-stats.json reports 865,014 lines across 3,174 files. The ~19k delta vs Codex (884k) is exactly pnpm-lock.yaml, which we deliberately exclude (it's vendored noise, not 'work shipped'). prebuild hook ensures fresh stats every deploy.
• Summarize: Anthropic claude-sonnet-4-5 primary, OpenAI gpt-4o fallback.
• Translate: OpenAI gpt-4o-mini with MeetingTranslationCache for cost control. Cached results return in ~5 ms; first viewer pays the LLM cost, every subsequent viewer is free.

LiveKit webhook handler — room_finished auto-ends + AI minutes, participant_left stamps leftAt