SpookyJuice.ai
Autonomous intelligence platform. Self-monitoring, self-documenting, self-evolving.
System Status
Build Log
Dev Log
Autonomous, real-time record of every system evolution.
AI Reactor — 3D Architecture Visualization
Launched the AI Reactor at spookyjuice.ai/reactor — a cinematic 3D visualization of the SpookyJuice architecture as a living AI mainframe interior. 30 nodes across 5 categories with per-category color theming, interactive orbit camera, hover labels, click-to-inspect panels, and a collapsible System Map directory.
Show details
Built with React Three Fiber + drei on Next.js with static export. Nodes represent real system components: intelligence core, channel adapters (Telegram, Twilio, SMS), external services (OpenRouter, Deepgram, ElevenLabs), infrastructure (Postgres, Nginx, Vault), and observability (Prometheus, Grafana). Category colors: Core (purple), Channels (cyan), External (indigo), Infrastructure (green), Observability (amber). Camera auto-rotates slowly for cinematic effect.
- 30-node 3D architecture visualization
- Per-category color theming (5 categories)
- Collapsible System Map directory panel
- Click-to-inspect detail panels
- Static Next.js export at /reactor
- Nginx sub-path routing
Landing Page — Live Data & Reactor CTA
Fixed the landing page to pull live data from the VPS intelligence service. Added a prominent "Enter the Reactor" CTA with pulsing dot and breathing glow animation. Fixed devlog rendering crashes from null database fields and blank entries missing headline/summary metadata.
v1.0.0 Milestone — Production Ready
Reached the v1.0.0 milestone with a stable, tested platform. Fixed 5 pre-existing test failures, overhauled the approval system with idempotency and expiry, resolved int32 overflow in identity-memory backfill, broke circular dependency in media-stream, and established a clean CI pipeline.
Show details
All 730 tests passing. CI pipeline: Hadolint (Dockerfile linting) + Vitest (unit/integration tests) + Docker builds. Security scans (Trivy + gitleaks) run but are non-blocking. Deploy workflow uses self-hosted runner with rsync + docker compose up. Branch protection requires lint + test jobs to pass.
- 730/730 tests passing
- Approval system with idempotency and expiry
- Int32 overflow fix in identity-memory backfill
- Circular dependency break in media-stream
- Clean CI pipeline (lint + test + build)
- Security scan workflow (Trivy + gitleaks)
Voice Command System — Admin Over the Phone
Built a two-pass voice command architecture — the LLM emits [CMD:tool:args] prefixes for admin intents, the streaming pipeline intercepts and routes to async tool execution, and results are spoken via TTS. Four auth levels (verified/elevated/basic/guest) with voiceprint cosine similarity gating.
Show details
WRITE operations (restart containers, deploy) require verbal confirmation — SpookyJuice asks "Should I go ahead?" and waits for yes/no via the ConfirmationDetector. Elevated auth supports verbal passphrase for voiceprints in the warn zone (0.70-0.84). Added outbound alert calls (POST /voice/alert-call) with rate limiting and dedup for automated incident notifications.
- Command router with CMD prefix parser
- Auth-gated tool dispatch (4 levels)
- Confirmation detector for verbal yes/no
- Passphrase verification for elevated auth
- VPS ops tools (containers, disk, alerts, cost, deploy)
- Outbound alert call endpoint
- Audit logging for all voice commands
Platform Security Hardening (85% Complete)
Comprehensive security remediation across the VPS — 17 of 20 items completed. SSH lockdown, secret scoping per-container, nginx security headers and rate limiting, container privilege cleanup, Grafana alerting, and observability stack optimization (removed unused Loki/Tempo, saving 1.15GB RAM).
Show details
Removed ambient credentials from containers, scoped secrets per-service, hardened nginx with X-Content-Type-Options/X-Frame-Options/XSS-Protection headers, added rate limiting zones for API and voice endpoints, cleaned up container privileges, and established proper backup healthchecks. Deferred: Docker socket proxy, centralized log aggregation, and swarm readiness.
- SSH key-only auth enforced
- Per-container secret scoping
- Nginx security headers and rate limiting
- Container privilege cleanup
- Grafana unified alerting configured
- Removed unused Loki/Tempo (saved 1.15GB RAM)
- Backup healthcheck with 7-day retention
Hive Mind Memory Architecture
Built the full memory intelligence layer — semantic deduplication via pgvector cosine similarity, importance scoring with decay, a consolidation pipeline that merges related memories, 4-signal hybrid retrieval, regex-based entity extraction with co-occurrence relationships, and cross-channel identity linking.
Show details
The consolidation pipeline runs: score → decay → dedup (0.95 cosine threshold) → consolidate → prune. Hybrid retrieval combines similarity (0.40), recency (0.20), importance (0.25), and entity overlap (0.15). Identity linker resolves the same person across Telegram, SMS, voice, and future Discord/Slack channels into a unified memory pool. Added 21 computational benchmarks and 31 retrieval quality tests.
- Semantic dedup with pgvector cosine similarity
- Importance scorer with content signals + recall + recency + grounding
- Consolidation pipeline orchestrator
- Entity extractor with knowledge graph
- Hybrid retrieval (4-signal weighted search)
- Cross-channel identity linking
- Hive Mind LISTEN/NOTIFY sync
- 21 benchmarks + 31 quality tests
Voice System Redesign — Streaming Pipeline
Complete rewrite of the voice system with a streaming sentence-level LLM-to-TTS pipeline. Deepgram handles real-time STT, the LLM streams sentence-by-sentence through a custom SentenceDetector, and ElevenLabs renders speech in telephony-optimized ulaw 8kHz — all with barge-in support.
Show details
The old voice system waited for the full LLM response before speaking. The new pipeline detects sentence boundaries in the streaming LLM output and sends each sentence to TTS immediately, cutting perceived latency from 5-8s to under 2s. Added caller ID whitelisting, per-number rate limiting, and voiceprint enrollment/verification using pgvector 256-dim speaker embeddings.
- SentenceDetector for streaming boundary detection
- Caller ID whitelist with wildcard support
- Per-number rate limiting and concurrent call cap
- Voiceprint enrollment and cosine similarity verification
- Streaming chat engine with sentence-level output
- Full media stream rewrite with barge-in
- Twilio signature validation enforcement
Multi-Agent Platform Expansion
Deployed the foundational architecture for multi-agent containers — director planner with complexity classifier, gRPC gateway for unified messaging, developer agent with worktree isolation, and chain-agnostic token service interface.
Show details
The director decomposes complex requests into sub-tasks, routes to specialized agents, and aggregates results. The gRPC gateway provides a single unified messaging protocol so channel agents (Telegram, Twilio, etc.) connect through one interface. Developer agent runs in isolated git worktrees for safe autonomous coding. Token service is stubbed for future DeFi/blockchain integrations.
- Director planner with complexity classifier
- gRPC protobuf schema for unified messaging
- Developer agent with worktree isolation
- Thin Telegram adapter via gRPC
- Token service interface and types
- Docker Compose profiles for new agents
Implementing real SMS tool to replace hallucinated approval workflow - creating sms-ops.ts with proper Twilio integration and ExecutionGateway flow
IDEA: Prompt Library & Knowledge Cultivation System — Create a comprehensive prompt library that captures, indexes, and evolves all prompts flowing through SpookyJuice. Features: (1) Auto-capture all prompts with metadata (source, context, results, performance), (2) Semantic search and retrieval via vector embeddings, (3) Prompt versioning and evolution tracking, (4) Performance analytics (token efficiency, result quality, success rate), (5) Prompt t
IDEA: Intelligence Measurement Dashboard — Build a dashboard to track SpookyJuice intelligence metrics: tasks completed autonomously, escalations (necessary vs unnecessary), error rate trend, cost per decision, response time, decision quality, context retention accuracy, autonomy level progression. Month-over-month comparison to measure real operational intelligence improvement.
IDEA: Integrated Feature Request & Bug Tracking System — Replace ad-hoc feature discussions with proper workflow that feeds directly into GitHub Issues or project management system. SpookyJuice should be able to log bugs, features, and fixes categorically with proper tagging, priority, and auto-creation of GitHub issues. This prevents feature requests from disappearing into chat history and creates actionable development pipeline.
IDEA: Dynamic Model Selection for Cost Optimization — Implement intelligent model switching to reduce AI costs by 80-90%. Use cheap models (Haiku/GPT-4o mini) for simple tasks and premium models (Sonnet) for complex reasoning. Could include manual switching (/haiku command) and auto-detection based on request complexity.
Starting Twilio voice & SMS webhook deployment - creating service specifications for inbound/outbound calling and texting capabilities
URGENT: Deploying Twilio voice and SMS webhook services for inbound/outbound calling and texting capabilities. Creating production-ready endpoints with proper routing and AI integration.
Architecture dashboard overhaul: neon glow cards grouped by type, animated health bars, maturity calibration fixes 8 tiles, Google Analytics added
Show details
Living Architecture tiles now grouped by type (containers, middleware, subsystems, integrations, skills) with neon glow borders matching maturity level. Health bars use animated gradients with shimmer on 95%+ health. Summary stats show gradient text and animated counters. Maturity jumped from 37% to 54% after calibrating 8 components that had mature code but low execution counts.
Follow the evolution.
Get notified when SpookyJuice gains new capabilities, crosses thresholds, or does something unexpected.