NeuroNest (BETA)

NeuroNest Release Notes

Release Date: May 2026
Platform:

• macOS (Apple Silicon (M1/M2/M3/M4) + Intel x86 64-bit)
• Windows (x86 64-bit (Intel/AMD), ARM 64-bit, Universal/x64 ),
• Linux (x86 64-bit (Intel/AMD), ARM 64-bit, x86 64-bit (Intel/AMD), ARM 64-bit, x86 64-bit (Intel/AMD))

---

Security Hardening

7-Layer Defense-in-Depth

• Firewall Engine — 4-tier regex-based scanning: sanitization, prompt injection detection, secrets scanning, policy enforcement with 18 default rules
• Enhanced Firewall — Hybrid regex + LLM semantic analysis with sophistication assessment, confidence scoring, and advanced PII redaction (email, phone, SSN, credit card, API keys, passwords)
• Action Security Analyzer — Pre-execution risk classification with pattern analyzer (shell threat signatures), policy rail analyzer (composed threats like curl piped to bash), and ensemble aggregation
• Runtime Protection — Production-only anti-tamper: DevTools detection, debugger flag detection, file integrity verification (SHA-256), environment anomaly detection, keyboard shortcut blocking
• Secure Communication — HTTPS enforcement, certificate pinning (TOFU), HMAC-SHA256 request signing with timestamps and nonces, replay protection
• Native C++ Module — Compiled Node-API addon for cryptographic operations (SHA-256, HMAC verification, secure random, file hashing, timestamp validation) with JS fallback
• Edit Lock — Directory-scoped file edit restrictions during debugging sessions

Firewall Configuration

• 4 policy presets: Permissive, Balanced, Strict, Enterprise
• Per-agent and per-project policy overrides
• LLM tier management with connection testing and latency thresholds
• Redaction configuration with format preservation
• Configuration import/export

Knowledge Graph

• Enhanced graph analysis with god node filtering (excludes file-level hubs and method stubs)
• Surprising connections with composite scoring: confidence weight, cross file-type bonus, cross-directory bonus, cross-community bonus, peripheral-to-hub detection
• Human-readable reasoning for each surprising connection
• Suggested questions generated from graph structure: god node coupling, cross-community bridges, community imbalance, dependency blast radius, isolated nodes
• Community distribution chart with color-coded stacked bar and legend
• Edge confidence breakdown (EXTRACTED / INFERRED / AMBIGUOUS) with percentages
• Themed node and edge detail panels (centered, adapts to selected theme)
• Suggested question click shows confirmation dialog, navigates to project search on confirm
• Collapsible All Nodes list in graph report

AI Features

• Brainstorm Mode — Auto-detects feature requests in chat and asks 3-5 clarifying questions before coding. Generates design summary from Q&A session.
• Cross-Model Second Opinion — Sends same query to two different configured LLM providers, compares responses with agreement scoring, shows overlapping vs unique findings
• Crash Recovery — Auto-saves session state (decisions, remaining work, failed approaches) to SQLite every 30 seconds. Offers resume on restart.
• Directory-Scoped Edit Lock — Locks agent edits to a specific directory during debugging. Prevents drive-by changes to unrelated code.
• Project Learning Memory — Per-project pattern storage (patterns, preferences, pitfalls, conventions, dependencies) that compounds over sessions. Confidence decays 5% per week.

All features are project-specific and displayed in the Task Management view under "Features".

Action Event Stream

• Typed Action/Observation event architecture for the AI pipeline
• Actions: shell_command, file_write, file_read, tool_call, agent_delegate, code_execute
• Observations: command_output, file_content, error, status_update, tool_result
• Pub/sub with topic filtering
• Query API with filters (agent, session, type, time range)
• Dashboard stats: total events, actions, observations, success rate, average latency

Docker Sandbox Runtime

• Secure Docker-based code execution with read-only filesystem, no-new-privileges, memory/CPU limits, network isolation
• Project mounted read-only at /workspace with tmpfs scratch space
• Command execution with timeout and kill support
• File operations (read/write/delete/list) inside containers
• Session management with port allocation
• Docker availability check with install button (Homebrew-based, with confirmation dialog and manual install option)
• Sandbox controls in Runtime Debugger via 🐳 button

UI Improvements

User Profile

• Plan labels standardized to COMMUNITY, PROFESSIONAL, ENTERPRISE across all views
• Device ID shows HWID (SHA-256 hash) matching the value sent to API
• License key displayed in profile details
• First name and last name populated from license data
• License email mismatch panel with logout option on passkey login

Settings

• Local AI providers (Ollama, llama.cpp, OpenMythos) hidden from Add Provider dropdown when not installed
• llama.cpp icon updated to llama silhouette SVG
• OpenMythos panel loads from cache for instant display, refreshes in background

Dashboard

• Action Event Stream stats section (5 metric cards)
• Docker Sandbox status indicator with install button

Knowledge Graph

• 6 stat cards: Nodes, Edges, Communities, God Nodes, Surprises, Files Analyzed
• Community Distribution chart
• Graph Report with god nodes (edge counts), surprising connections (reasoning), communities (member previews), edge confidence breakdown
• Themed node/edge detail panels (centered, theme-aware)
• Suggested questions with search integration

Provider Updates

• Bearer token updated for neuronest.cc license API
• API response field mapping fixed (snake_case: license_key, user_email, expiry_date, first_name, last_name)
• Features normalized to lowercase for consistent plan derivation
• "Code already used" validation response treated as valid for returning users
• markUsed API call updated to include appId parameter

Technical Details

• 117 AI agents across 13 departments
• 11 LLM provider integrations (OpenAI, Anthropic, Gemini, DeepSeek, Grok, Mistral, NVIDIA, Groq, Ollama, llama.cpp, OpenMythos)
• 6 messaging channels (WhatsApp, Telegram, Discord, Slack, Email, GitHub)
• 198 skills with 2,275 agent-skill assignments
• SQLite with WAL mode, 10 migrations
• 53 license-related tests (property-based + unit)
• macOS only (Apple Silicon + Intel)