Software Updates

universal translation cache + multi-module translation support

• New ContentTranslation table (content_translations) — polymorphic cache for any entity type: email_message, chat_message, task, task_comment, lead, crm_note, meeting_note
• Raw SQL migration: prisma/migrations/add-content-translations.sql
• Single @@unique(entityType, entityId, fieldName, targetLanguage) ensures one cached row per entity per language pair, shared across ALL users
• src/lib/translation-service.ts — cache-first translation service
• Checks DB cache before calling AI (zero cost on re-opens)
• Persists every new translation to DB automatically
• getLanguageStats() for analytics
• POST /api/admin/translate — universal endpoint for all entity types
• GET /api/admin/translate/stats — language distribution analytics
• POST /api/admin/email/[id]/translate — refactored to use shared service
• src/hooks/useTranslation.ts — cache-aware hook, auto-translates on mount
• src/components/admin/translation/TranslationSideBySide.tsx
• TranslationSideBySide: two-column layout with original + translated
• TranslationBadge: pill badge for headers (shows 'cached' indicator)
• LanguagePill: small non-interactive language tag for list views
• src/components/admin/translation/LanguageAnalyticsWidget.tsx
• Bar chart of languages by volume + breakdown by module
• thread-view.tsx: refactored to use useTranslation hook (cleaner)
• 'cached' label shown when translation served from DB
• thread-list.tsx: LanguagePill shown on inbound foreign-language rows
• server-data.ts: joins content_translations to populate language data in email list rows at zero extra AI cost
• MessageBubble: hover-reveal Translate button on all inbound messages
• Inline translated card below bubble (WhatsApp/Telegram style)
• Cache-first via universal translate endpoint
• TaskDrawer: Translate button next to Description label
• TaskTranslateButton component: toggle translation for descriptions AND individual task comments
• LanguageAnalyticsWidget ready to drop into /admin/analytics

add missing closing parenthesis in meeting feedback page

• Fixed syntax error in thank-you screen conditional return
• Meeting feedback page now builds successfully

auto-translate emails in threads with side-by-side layout

• On new email threads, auto-detect language of inbound messages
• If foreign language detected, show translated version side-by-side
• Left column: original email in detected language
• Right column: English translation with identical message card format
• Translation badge in subject header with quick dismiss (X)
• Two new reply buttons under translated column:
• 'Reply in English': sends English reply auto-translated to sender's language
• 'Reply all in English': same for all recipients
• Maintains existing manual Translate toggle in toolbar
• Graceful fallback: silently ignores if already English
• Uses existing translate API endpoint with auto-translation on mount

Fix 4 critical issues from testing session

• 1. VOICE COMMANDS: firePendingQuery phase-2 timer never called setYenChatOpen(true)
• — YenChatPanel's initialQuery auto-submit only fires when open=true, so voice
• commands were silently dropped. Adding setYenChatOpen(true) restores full Hey Yen.
• 2. TRANSLATION PICKER: Remove disabled={lang.code===sttLang} restriction.
• User wants 'Don't Translate' vs 'Translate to English' to be distinct options
• even when their spoken language IS English (e.g. see translated French → English).
• Also disable useRealtimeTranslation for same-language pairs (en→en no-op).
• 3. AUTO-SCROLL: Add postStage + streaming to useEffect deps so chat scrolls
• to bottom when post-response buttons appear, not just when thread changes.
• 4. EMAIL CONFIRMATION: After send_email succeeds, append an in-thread Yen
• message confirming delivery + recipient, not just a transient toast.
• Bottom input padding trimmed (pb-3→pb-2), hint text removed, sticky
• action bar padding tightened (py-2.5→py-2).

Meeting Display Tray: share interactive 3D CAD models with all participants

• Bottom-left Layers button (desktop) opens category menu
• CAD Files category: jewelry projects ordered newest first, filtered to STL files
• Pick project → pick file version → pick metal → 'Display for everyone'
• Broadcasts display_open via LiveKit data channel to all participants
• Each participant gets their own independent MeetingCadViewer overlay
• Orbit controls: drag to rotate, scroll/pinch to zoom
• Metal picker: Yellow/Rose/White Gold, Platinum, Silver, Wax
• Close button (local only) + Stop Sharing button (host only)
• Broadcasts display_close when host stops sharing
• Guest room client also handles display_open/display_close
• New API: GET /api/admin/meetings/[id]/display-assets

Uniform toolbar buttons, fix tab close on exit, independent private/public Yen, feedback close-tab button, fix /admin/yen logo blend mode

Fix privacy leak, 3x Yen logo, faster control hide, tool timeouts, sticky confirm bar, tile query wrap

implement proper locale resolution with browser language detection

• Replace router.refresh() with window.location.reload() in locale switcher (router.refresh() doesn't trigger server component re-renders with new cookies)
• Add 3-tier locale precedence in resolveLocale(): 1. k99_locale cookie (explicit user choice) 2. Accept-Language header parsing (browser auto-detection) 3. English default (fallback)
• Auto-set k99_locale cookie when browser language is detected on first visit
• Confirm fallback merge order in i18n/request.ts (English base + translated override)
• Users in non-English locales now auto-detect on first visit
• Language switcher now immediately changes page content via full reload

Add post-meeting feedback page with thumbs up/down ratings and token reward incentive

API platform: per-service billing controls, custom SVG icons, i18n

• New ApiServiceDefault table — platform-wide defaults with conservative caps (yen: 20k/mo + 1k/day, email: 10k/mo + 5k/day, meetings: 50k + 200/day, etc.)
• New ApiServiceLimit table — per-tenant overrides for instant fine-grained control
• service-limits.ts module: checkServiceLimit(), incrementServiceUsage(), disableServiceForTenant(), setMonthlySpendCap() admin helpers
• credits.ts now gates every debit through service limits BEFORE touching balance
• Three new error codes: service_disabled, spend_cap_exceeded, daily_cap_exceeded
• errors.ts: serviceDisabled(), spendCapExceeded(), dailyCapExceeded() helpers
• catalog-icons.tsx: 25 unique stroke SVG icons, one per API product (email=plane, inbox=tray, bell, yen=Y-spark, brain, camera, globe, signature=pen, files, workflow=nodes, shield, calendar, tasks=checkbox, support=headset, crm=network, invoice=receipt, payroll=group+$, webhook=lightning, domain=radial, flag, import=upload, pulse=heartbeat, product=box, bag=orders, chart=bars)
• catalog.ts: all emoji strings replaced with icon ID strings
• Landing page: icons render in styled bordered squares with hover effects
• ApiPlatform namespace added to messages/en.json (all UI strings)
• i18n/request.ts: English fallback merge so new namespaces work in all 24 locales
• Landing page reads locale from k99_locale cookie (set by main site or footer)
• ApiPlatformLocaleSwitcher component: dropdown in footer, sets cookie + refreshes
• All 12 major languages available in the switcher
• generateMetadata() is now translated

Separate CC button from Brain (AI Notes), auto-translate meeting minutes to viewer account language

Style API platform landing page with Katura branding

• Add white Katura logo SVG in header (inverted for dark background)
• Update main heading to use Playfair Display font matching 'TIMELESS ELEGANCE'
• Add gold dividers above and below hero section for elegant styling
• Improve typography: uppercase tracking, refined button styling
• Update header with proper logo and navigation spacing
• Match homepage design language while maintaining dark API platform aesthetic

Fix 7 meeting bugs: Yen logo blend, non-overlapping layout, 2-phase voice, guest Hey Yen stale closure, missing notes route, rounded YenPanel, allowGuestYenRef

Fix build error: rename branding.ts to branding.tsx (file contains JSX)

Fix overlapping UI: subtitle-aware offsets for Yen+chat panels; white Yen logo; context-aware invocation; block same-lang translation; meetings open in new tab

Fix in-meeting chat sidebar layout: flex-col container so messages scroll and input pins to bottom

Logo refresh + meeting UX: circular Yen logo, subtitle overlay, invite dual-action

• Added public/yen-logo.png (black cursive Yen script on transparent bg)
• admin/yen page: all 4 logo instances → circular dark navy (#0b1527) circle with Yen logo (rounded-full + overflow-hidden + ring-1) Empty-state hero circle is now h-20 w-20 (larger than before)
• YenBubble FAB: gradient rose→indigo replaced with dark navy circle + Yen logo. Panel header + message avatars + empty-state all updated.
• Meeting toolbar Yen button: Sparkles icon replaced with Yen logo PNG (filter:invert so black→white to match other toolbar icons, 70% opacity when inactive, 100% when active)

Fix 400 error + chat panel height + toggle + sentence case

• FIX 1 — 400 'Invalid value: function_call' (CRITICAL — Yen broken):
• The llm-adapter was using the OpenAI Responses API for ALL OpenAI-
• compatible providers, including xAI. But the Responses API is only
• supported at api.openai.com — xAI (api.x.ai) and all other compat
• hosts only speak Chat Completions. When failover or primary hit xAI,
• the adapter sent Responses API format (input[] with type:function_call
• items), which xAI rejected with 400 'Invalid value: function_call'.
• Fix: added apiStyle: 'responses_api' | 'chat_completions' to
• ModelAssignment. OpenAI assignments get responses_api; xAI gets
• chat_completions. Added full Chat Completions path in llm-adapter:
• callOpenAIChatCompletions() + streamOpenAIChatCompletions(). Dispatch
• in callLLM/streamLLM routes to correct path based on apiStyle.
• FIX 2 — Yen chat panel opened as small window (height not constrained):
• YenChatPanel is 'absolute top-0 bottom-0 left-0' — but for top/bottom
• to constrain height, the element needs a positioned ancestor (relative/
• absolute/fixed). The left column div was position:static, so the panel
• had no containing block and height was auto (content-only). The panel
• appeared as a small box the height of its header + input.
• Fix: added 'relative' class to the left column div. YenChatPanel now
• fills the full height of the column (top:0 to bottom:0).
• FIX 3 — Toggle switch knob moved outside the pill boundary:
• Both subtitle toggles used 'top-1 w-4 h-4 translate-x-1/translate-x-6'.
• The container is w-11 h-6 (44x24px). With translate-x-6 (24px) and
• a 16px knob, the right edge was at 40px — technically inside 44px but
• visually hugging the rounded edge. Changed to 'top-0.5 left-0.5 w-5 h-5
• translate-x-0/translate-x-5': 2px padding on all sides, visually clean.
• FIX 4 — Subtitle text forced lowercase (Deepgram returns all-lowercase):
• subtitle-panel.tsx had textTransform:'lowercase' as the default (non-
• caps) case, making all-caps mode the only way to see capitalised text.
• Fix: changed default to textTransform:'none' and added sentenceCase()
• helper that capitalises the first letter of the string and after . ! ?
• Applied to all transcript text and translations.

Fix 3 persistent bugs: tile dismiss, shipping state inference, lookup resilience

• BUG 1 — Yen tile X still not clickable (stacking context):
• The tile lives inside 'absolute inset-0 flex gap-1' which had NO
• z-index, so it defaulted to z-auto (0). The title bar is a SIBLING
• at z-30 — it painted over the entire tile including the X button.
• The previous fix (z-10→z-40 inside the tile) was irrelevant because
• z-40 only applied within the tile's local stacking context.
• Fix: (a) Moved dismiss button OUT of YenMeetingTile, rendering it as
• a sibling element in the parent at z-40 — same stacking context as
• the z-30 title bar, so it wins. (b) Added z-0 to the video container
• so it participates in the parent stacking context explicitly.
• BUG 3 — Shipping fails for Miami / certain ZIPs:
• ShipStation carriers require toState for reliable rate lookups. When
• the model omitted toState (common — it only passes ZIP), carriers
• returned errors silently. Added zipPrefixToState() that auto-infers
• the US state from the first 3 digits of any ZIP code. Also added
• toCity to the tool schema + API call so carriers get full address
• context. Tool description now says ALWAYS provide toState.
• BUG 4 — Stella lookup still failing (unguarded queries):
• workerProfile.findMany() and emailContact.findMany() had NO
• .catch(() => []) — if either table/migration doesn't exist in prod,
• the exception killed the ENTIRE execLookupClient function, hiding
• results from buckets that DID succeed (User, JewelryProject).
• Fix: added .catch(() => []) to both queries, matching the pattern
• already used on the other two buckets.

Fix 4 Yen/meeting bugs: tile dismiss, chat input, shipping origin, lookup

• BUG 1 — Yen tile X button never worked:
• Dismiss button was z-10 inside a stacking context where the title bar
• is z-30. When user hovers (which keeps the title bar visible), the bar
• ate the click. Fixed: z-10 → z-40 in yen-meeting-tile.tsx.
• BUG 2 — No text input/confirm in Yen chat panel after voice trigger:
• publicMode=true caused the panel to 'return null' before rendering,
• leaving the host with no way to type follow-ups or confirm actions.
• Fixed: remove the early null-return. Panel now always renders when
• open; the 'Broadcasting' badge still communicates public mode.
• BUG 3 — Shipping rates used wrong origin (Cartersville/Austin, not user-specified):
• Both the meeting yen-chat get_shipping_options tool (ShipStation) and
• the general yen chat getShippingRates tool (ShipEngine) had hardcoded
• ship-from with no origin parameters in the schema. The model couldn't
• pass 'from Greenville SC' even when the user said it explicitly.
• Fixed: added fromCity/fromState/fromZip params to both tool schemas
• and handlers. Origin defaults to studio but overrides when provided.
• BUG 4 — Stella's email not found (lookup silently failed):
• execLookupClient was querying User.firstName, User.lastName,
• User.jobTitle, User.department — columns that don't exist in the DB
• schema. Prisma threw 'column does not exist', which was caught and
• returned as 'Could not search contacts: ...' — model then asked for
• the email. Fixed: User query now only uses name + email (guaranteed).
• WorkerProfile + JewelryProject buckets get individual .catch(() => [])
• so one missing table never kills the whole lookup.

Add @mention autocomplete to live meeting invite panel

• Typing '@' in the invite name field triggers the existing suggest-participants API (team / past guests / clients)
• Debounced 180ms — won't fire on every keystroke
• Dropdown shows avatar initial, full name, email, and colored kind badge (blue=team, teal=client, amber=guest)
• Arrow keys + Enter for keyboard navigation; Escape to dismiss
• Selecting a suggestion prefills both name AND email fields
• X clear button resets the field and re-focuses
• @-less input: plain text mode, no suggestions (new guest names)
• Panel close + 'Invite another' both clear autocomplete state
• Zero changes to invite/join/token flow — guests still get full LiveKit permissions via the signed token at /g/[token]

Remove Opus entirely — Sonnet is the hard ceiling

• OPUS_MODEL constant removed from model-router.ts
• isOpusAllowed() helper removed
• Opus removed from HIGH_STAKES failover chain
• Opus confirmation wall removed from yen/chat route
• YEN_MODEL_OPUS env var no longer referenced anywhere
• Pricing table in header updated — Opus entry removed
• Header comment updated: 'Sonnet is the hard ceiling. No Katura use case justifies Opus.'

Demote Opus from automatic primary to last-resort confirmation wall

• Opus is no longer the primary model for HIGH_STAKES when YEN_ALLOW_OPUS=true. Sonnet is now the ceiling for all automatic routing.
• In the failover chain, Opus moves to dead last — only reachable if every other provider (Sonnet, Grok-4, GPT) fails AND YEN_ALLOW_OPUS=true.
• New Opus pre-flight in yen/chat route: if the chain contains an Opus assignment and opusApproved != true, returns HTTP 402 with cost warning instead of calling the model. UI must re-POST with opusApproved: true.
• Cost warning includes estimated $/call for both Opus and Sonnet so the user can make an informed decision.

Wire general yen/chat through failover + add ensemble mode for HIGH_STAKES

• yen/chat (general sidebar) now routes anthropic/openai through callLLMWithFailover — automatic xAI/Grok fallback on 5xx/429/timeout
• Classifier runs once per request; spend records against actual served model + classified task class (no more hardcoded CORE_ASSISTANT)
• Groq/Ollama keep legacy single-shot path (local/free dev only)
• New callLLMEnsemble() runs the whole chain in parallel for HIGH_STAKES, picks consensus by tool-call agreement → text centrality → first success
• Opt-in via YEN_ENSEMBLE_MODE=true (~3x cost, ~90¢/call)
• Strategy doc updated with ensemble section + new env flag

xAI provider + failover + deterministic classifier

• Add 'xai' provider (XAI_API_KEY, https://api.x.ai/v1)
• Add grok-4-fast ($0.20/$0.50) and grok-4 ($3/$15) models
• New getFailoverChain(task) returns ordered fallback list
• ROUTING/EXTRACTION now prefer Grok-4-fast (cheapest output anywhere)
• HIGH_STAKES chain: Sonnet → Grok-4 → GPT-5.4 (third opinion)
• CLIENT_WRITING unchanged: Claude is still the velvet glove
• OpenAI clients now use assignment.baseUrl (xAI works as drop-in)
• xAI dispatches through OpenAI Responses API path (compatible)
• New callLLMWithFailover() and streamLLMWithFailover()
• isRetryableError() classifies 5xx/429/timeout/connection errors
• onFailover callback for observability + cost tracking
• classifyRequest({query, context, hint}) — zero-cost, microseconds
• Risk catalogs: legal, financial, compliance, PII keywords
• Auto-escalates to HIGH_STAKES on any risk signal
• Verb-based natural classification (writing/summary/routing)
• Length-based escalation: >20K tokens requires standard tier
• Caller hint as floor — never downgrades below requested tier
• Classifies every query before model selection
• Uses streamLLMWithFailover with full chain
• Records spend against actually-served model (post-failover)
• recordSpend now uses classification.taskClass (was hardcoded)
• Logs escalation reasons for audit trail
• Adds 'xai' to PROVIDER_COLORS (orange)
• Provider detection now recognizes grok-* models
• Adds 'import type React' to fix React.ElementType reference
• Full provider matrix + failover behavior + classifier rules
• Pricing reference, env vars, end-to-end flow diagram
• Cost & output wins per optimization
• Future work: ensemble voting, A/B testing, per-tenant routing

enterprise-grade AI usage dashboard

• Add byTaskClass breakdown (EXTRACTION/ROUTING/SUMMARY/etc.)
• Add byProvider aggregation (OpenAI vs Anthropic split)
• Add previousPeriodTotals + trend % (cost/calls/tokens vs prior period)
• Add avgCostPerCall to byModel and byTaskClass
• Add provider field to byModel rows
• Add memoryHealth from memory-graph.ts getMemoryHealth()
• Add routingConfig snapshot showing active model assignments
• Add approvalRequired + fraction to budget response
• Fix BigInt literal compat (BigInt(0) vs 0n)
• Daily sparkline now includes totalIn/totalOut per day
• Period selector tabs (Today / This Week / This Month)
• KPI cards with trend indicators vs previous period
• Budget progress bar with color thresholds + approval warning
• Provider split stacked bar (emerald=OpenAI, violet=Anthropic)
• Task class table with unit economics (avg cost/call)
• Enhanced By Model table with provider dot + avg/call
• By Source table now shows task class labels
• Memory Graph Health card (nodes, types, durability, confidence)
• Active Routing Config table (task → model → tier badge)
• Token Rate Limits with per-user and team-wide progress bars
• Loading spinner on period change (non-blocking)

wire streaming loop through streamLLM adapter

• Add streamLLM / LLMMessage / LLMToolDef imports
• Add LLM_TOOLS constant mapping Anthropic.Tool[] → LLMToolDef[]
• Convert thread messages to LLMMessage[] format
• Replace agentic loop with StreamEvent consumer
• Tool execution functions and SSE surface unchanged
• Remove direct Anthropic client instantiation from handler

multi-model routing: OpenAI + Claude per task class

• Routes by provider + model per task class
• EXTRACTION/ROUTING → OpenAI nano ($0.20/$1.25)
• SUMMARY/CORE_ASSISTANT → OpenAI mini ($0.75/$4.50)
• CLIENT_WRITING/HIGH_STAKES → Claude Sonnet ($3/$15)
• Falls back to Anthropic-only if no OPENAI_API_KEY
• Updated pricing: Haiku $1/$5, Sonnet $3/$15, Opus $5/$25
• Unified callLLM() + streamLLM() over both providers
• OpenAI: Responses API (not legacy Chat Completions)
• Anthropic: Messages API via SDK
• Normalized tool definitions, responses, streaming events
• Extraction now uses EXTRACTION task class (was SUMMARY)
• Both extraction + summary wired through llm-adapter
• Provider-agnostic: routes to nano/mini when OpenAI available
• Added checkMonthlyBudget() before streaming (was missing)
• Added recordSpend() after response with token tracking
• Memory writeback now passes clientId/projectId
• Hoisted meeting lookup IDs for both retrieval and writeback
• Applied Virgil corrections (Responses API, advisory classifier)
• Applied Codex audit findings (6 bugs fixed)
• Updated implementation status

add multi-model routing plan (OpenAI + Claude)

wire memory + validation into live meeting Yen

• Load KV memories (loadRelevantMemories) + graph memories (buildMemoryContext) before every Claude call in yen-chat route
• Inject client/project-scoped memory into system prompt
• Add 6-point validation checklist (names, live data, fabrication guard)
• Post-response memory writeback via processTranscriptIntoMemory
• Update YEN-CAPABILITY-AUDIT.md with implementation status

Yen capability audit — strengths, weaknesses, and roadmap to agent

no eager autocomplete + Yen system prompt overhaul + weather dispatch

• Remove eager load of all contacts when instant meeting modal opens
• Suggestions only appear when user has typed ≥1 character
• onFocus no longer fetches empty query
• Replaced rule-list style with a real PA persona: proactive, autonomous, acts-then-confirms rather than asks-permission-to-try
• 'Use the transcript' — explicitly instructed to scan meeting context for names/emails mentioned earlier in the call
• Lookup before email: silently call lookup_client without announcing it; only ask for email if lookup AND transcript both yield nothing
• 'Incomplete sentences are okay' — Yen should infer intent from context rather than waiting for a grammatically complete request
• Tools described by trigger pattern, not as a checklist
• Removed bureaucratic preamble ('Great question!', 'Certainly!')