NV Digital Open Operator System (NVoOS)

Version: 1.1.97 Release Date: 2026-10-06

See § Release History for the 12 most recent releases; every older release is in the Previous Releases table. Full detail: CHANGELOG.md.

🆕 v1.1.97 Highlights: ECC-inspired agent-harness enhancements, OOS parity gaps closed, and gateway protocol negotiation. Harness enhancements (#6912, Proposal 056 — all inert by default) — cascade model routing on both engines (the framework-core CascadeRouter + the legacy WP_MCP_AI_Cascade_Executor gate with Pro Jev classifier wiring), the run_assistant_eval + scan_assistant_security base tools, hook profiles (minimal/standard/strict gating), the session distiller's canonical wp_mcp_ai_memory_stored event, and the Pro suggest_workflows_from_history trace-store miner. OOS parity (#6913) — the destructive-ops gate no longer becomes an HTTP 500 on the OOS path (canonical 428/429 envelopes), the Output Guardrail + Citation Verifier run at the final payload on all three chat surfaces, and bin/check-chat-parity.php (35 → 38 features) gains a parity-check CI job. Fixed (#6914) — the MCP Gateway negotiates protocol versions (2024-11-05 fallback), so Zed and other strict MCP clients connect. npm (#6907/#6908) — all 23 packages/ are live on the public registry and the bridge's bare npx form resolves. Tool count: ~352 base + ~1,313 Pro (~1,665 total; +2 base +1 Pro). Stale 1.1.95 build ZIPs removed (30 files).

MCP Specification: 2026-07-28 (Stateless Core, Full Compliance)
Maintained by NV Digital
License: GPLv3 or later
Requires: WordPress 6.0+, PHP 7.4+
Patent Status: Patent Pending (Application #19/410,504)
Documentation: Grade A (95/100) — 1,617 files across 12 directories, 108 admin screenshots, 100% feature coverage

🔍 For Reviewers & Auditors

New to this repo? Start here → docs/project/FOR_REVIEWERS.md

That document answers every common question in one place: what the project is, current security posture, what's production vs experimental, PHP version requirements, AI development methodology, compliance status, and scoping advice for a limited-budget review.

Quick links for reviewers:

• Addon Inventory — what each of 30 addons does and its status
• Security Posture — current state of all 50 audit findings
• Compliance Traceability — every .org rejection reason → commit → verification command
• AI-Assisted Development — methodology, transparency, and what to scrutinize
• Architecture Overview — component diagram and data flow

📑 Table of Contents

Getting Started

• 🧩 Overview
• 🎯 Our Mission
• 🛡️ Active Security Monitoring
• ⚠️ Warranty & Safe Use
• 🏗 System Architecture
• 🚀 Features
• 📦 Installation
  • 🌱 Try It on Your PC
• 🔌 What You Lose Without Third-Party Plugins
• ⚙️ Configuration Checklist
• 📚 Documentation

Core Functionality

• 🧠 Memory & Tool Stack Overview
• 🛠 Built-in tools & automations
• 🗨️ Front-end chat surfaces
• 💬 Frontend Shortcode

Addons & Extensions

• 🏗 System Architecture (covers addon ecosystem)

Orchestration & AI Features

• 🏗 System Architecture (covers orchestration, harnessing, MCP servers, memory bridge)

AI Providers & Integration

• 🧠 Language Model Providers
• 🧱 ChatKit Integration
• 🌐 Crawl4AI Integration
• 📡 Job Notification System
• 🧊 Elementor Widgets

Performance & Optimization

• ⚡ Message Bundling
• 🎯 Agentic Loop Token Management
• 🔄 Chat Performance Optimizations
• 🌐 Mesh Compute Routing
• 🔗 Federation & Discovery System

Remote MCP Setup

• 🔒 MCP Server Authentication
• 🌐 Connecting Remote MCP Clients
• 🛰 REST API Endpoints
• 🌊 SSE Streaming Support
• 📝 MCP JSON-RPC 2.0 Endpoint
• 🔑 Assistant API Credentials
• 🎫 Token Management UI

Assistant Management

• 🛠 Assistant Editor Overview
• 📊 Assistant Storage: CPT vs CCT
• ⚡ Assistant Tool Shortcuts
• 🧠 Agent Skills
• 👔 Professional & Team Layers
• 🧵 REST Chat Payloads & Attachments

Development

• 🐳 Local Development with Docker
• 🧑‍💻 Development Tooling
• 📦 NPM Packages
• 🧪 Testing & QA
• 🧩 Hooks & Filters
• 🧰 WP-CLI Commands

Reference

• 🔐 JetEngine Capability Reference
• 🛰 JetEngine REST API Reference
• 🧮 Usage Tracking
• 🧷 Attachment MIME Controls
• 🧾 Logging
• 🧾 JetEngine REST Endpoint Report Helper
• 🔌 Optional Tools & Dependencies
• ✅ Manual QA Scenarios

Changelog

• 📜 Release History
• 📋 Previous Releases

🗺 Repository Map

🧩 Overview

Real-time AI Orchestration Toolkit / Harness for Wordpress - NV oOS is a modular AI framework (Object-Oriented System) for WordPress that connects your site's data with 18 language-model providers: OpenAI, Gemini, Anthropic, DeepSeek, OpenRouter, Baseten, Kimi (Moonshot), Z.AI (GLM), DigitalOcean, NVIDIA NIM, Cloudflare Worker AI, Azure, Ollama, LM Studio, Hugging Face, and TypeSafe Jev (decision-only), plus local WebLLM and embedded engines. It allows you to create and manage AI Assistants that can interact with users, access WordPress data, and perform custom tool functions.

✨ What's New at a Glance (v1.1.97)

• 🤖 Agent-harness enhancements (PR #6912, Proposal 056 — all inert by default). Cascade model routing on both engines (the framework-core CascadeRouter extends ProviderRouter and injects into the OOS engine via WordPress CascadeClassifier/CascadeValidator adapters; WP_MCP_AI_Cascade_Executor gates the legacy create_chat_completion() path — one wp_mcp_ai_cascade_enabled switch + Jev classifier govern both, with post-call validation + escalation completing the previously dormant tier machinery); the run_assistant_eval base tool (three trajectory judges + deterministic checks, +1 base); hook profiles (minimal/standard/strict gating, meta → filter → option → constant resolution, master switch + kill list); the scan_assistant_security base tool (OWASP-informed config audit, masked secret evidence, posture score, +1 base); the session distiller (session-end transcript → canonical wp_mcp_ai_memory_stored event, zero provider cost by default); and the suggest_workflows_from_history Pro tool (harness trace-store mining, frequency × recency × success, +1 Pro). ECC MIT attribution in CREDITS.md.
• 🛡 OOS parity gaps closed (PR #6913). The four Proposal 029 G2 guardrail gaps that blocked promoting the OOS engine switch: the destructive-ops gate no longer becomes an HTTP 500 on the OOS path (typed catch → canonical 428/429 envelopes via WP_MCP_AI_REST::translate_gate_exception()); apply_pre_response_render() runs the Output Guardrail + Citation Verifier at the final payload on all three chat surfaces; the OOS bridge fires wp_mcp_ai_agentic_iteration_complete; the before-tool-execute waterfall bridge gains regression tests; and bin/check-chat-parity.php now tracks all four surfaces (35 → 38 features, 100% parity) with a parity-check CI job.
• 🔧 Gateway protocol negotiation fixed (PR #6914). The MCP Gateway's initialize now negotiates the protocol version (negotiateProtocolVersion(), 2024-11-05 fallback) — Zed and other strict MCP clients stop aborting with "Unsupported protocol version" against the live fleet endpoint.
• 📦 npm publish wave (PRs #6898, #6904–#6909, #6911). All 23 packages/ are live on the public npm registry (0.1.0-alpha.3; nvoos-api/nvoos-sse-client 0.1.0-alpha.4 after the TypeScript was stripped from their shipped JS; nvoos-mcp-bridge latest = 0.1.0-alpha.3 — the v1.1.96-deferred publish executed). GitHub Packages parity + the new dry-run-by-default npm-maintenance.yml.
• 📦 Versioning — bumped to 1.1.97 across all version-bearing files. Pro addon: 1.1.97. MCP Gateway: 0.1.1 (0.1.0 → 0.1.1 post-window via #6925 — the non-force mirror sync; the negotiation fix shipped on the 0.1.0 line). nvoos-mcp-bridge: 0.1.0-alpha.3 (published, latest). Media Worker: 3.4.0 (unchanged). Media Studio: 0.6.1 (unchanged). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). ChatGPT Plugin addon: 0.1.0 (unchanged). Fleet Operator: 1.0.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — zero plugins/ diff in-window). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.10.03 (unchanged). Tool count: ~352 base + ~1,313 Pro (~1,665 total — +2 base +1 Pro) — run_assistant_eval + scan_assistant_security (base) and suggest_workflows_from_history (Pro) from #6912. Provider count: 18 chat providers (unchanged). Addon count: 30 (unchanged). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 62 (unchanged). Stale build ZIPs removed: the 1.1.95 oOS set (9 root + 2 optional-components + 19 toolkit-addons = 30 files).

🎯 Mission: Modernizing Small to Medium Business Websites

NV oOS is specifically designed to help small to medium-sized businesses fast-track their outdated, stale, or insecure company websites to modern technology standards—without the need to add yet another wrapper around API calls. Instead, we're trying to peel back decades of API wrappers with the help of AI, providing:

• Direct AI Integration - No middleware required. Connect directly to OpenAI, Gemini, Anthropic, Hugging Face, Cloudflare Worker AI, Ollama, LM Studio, OpenRouter, and DeepSeek without custom development
• Security-First Architecture - Built-in protection against nefarious usage with active monitoring and prevention systems
• Enterprise-Grade Features - Access to capabilities typically requiring expensive custom development
• Compliance & Audit Tools - Comprehensive logging, rate limiting, and usage tracking built-in
• Zero Technical Debt - Modern codebase following WordPress standards, ready for current technology stacks

🛡 Active Security Monitoring

NV oOS actively prevents and monitors against nefarious behavior. The plugin includes:

• Nefarious Usage Monitor - Real-time detection of suspicious patterns and automatic emergency shutdown capabilities【F:includes/class-wp-mcp-ai-nefarious-usage-monitor.php†L1-L676】
• Root Security Key - Optional emergency authentication layer to prevent unauthorized reactivation after security incidents【F:docs/features/security/root-security-key.md†L1-L511】
• Granular Capability Controls - Every tool and API endpoint enforces WordPress capabilities to prevent unauthorized access
• Rate Limiting - Built-in protection against abuse with configurable limits per user, model, and time period
• Comprehensive Audit Logging - Track all API calls, tool executions, and security events for compliance and forensic analysis
• Input Sanitization & Output Escaping - All user input sanitized, all output escaped following WordPress security best practices

This is not a tool for circumventing security or promoting bad practices. Every feature is designed with security, transparency, and responsible AI usage as core principles. The plugin actively works to stop and prevent misuse before it happens.

Latest audit: See docs/operations/compliance/SECURITY_AUDIT_2026_04.md — the published summary of the April 2026 security & compliance code review (no Critical findings; 5 High items, 3 Fixed and 2 Partially Fixed). Full deliverables under docs/project/audits/2026-04/.

WordPress.org compliance hardening (May 9, 2026): docs/operations/compliance/WORDPRESS_ORG_COMPLIANCE_2026_05_09.md — findings B3, B8, B10, B13, and production vendor remap all resolved.

⚠ Warranty & Safe Use

We make every effort to keep NV oOS safe and secure — but by design, it can be destructive and resource-intensive when not properly configured.

NV oOS grants AI assistants access to powerful WordPress operations. The same capability that automates real work can cause irreversible harm if misconfigured:

• Destructive tools — bulk content deletion, user management, file writes, mass email, WP-CLI, direct database operations
• API billing exposure — uncapped AI provider calls can exhaust quotas and trigger unexpected charges
• Server resource exhaustion — concurrent agentic loops and SSE streams can saturate CPU/memory on shared hosting

Before going live: test on staging, take verified backups, apply least-privilege tool permissions, enable rate limiting, and review the system prompt of every public-facing assistant.

📄 Full details: WARRANTY.md — security commitment, "AS IS" disclaimer, destructive-operations table, resource-consumption guide, and mitigation checklist aligned with OWASP, NIST SP 800-53, ISO/IEC 27001, and the WordPress Plugin Developer Handbook.

Patent Pending

NV oOS is the subject of a pending patent application for its novel System and Method for Dynamic AI Orchestration Layer with Real-Time Capability Gating and Resource Budgeting.

Application Number: 19/410,504

The patent covers NV oOS's innovative approach to implementing sophisticated AI orchestration in WordPress's request-based PHP architecture—a platform not designed for real-time streaming, asynchronous operations, or persistent state management. This technical achievement enables enterprise-grade AI capabilities on WordPress by recreating event-driven behavior within PHP's synchronous execution model.

Key Innovations Covered:

• Dynamic resource budget allocation during streaming operations
• Capability-based access control for AI tool execution
• Registry-state-based scheduling in stateless environments
• Metrics-driven budget adjustment for real-time optimization
• Persistent-behavior illusion in request-based architectures

The orchestration layer makes NV oOS unique in the WordPress ecosystem by solving fundamental architectural limitations that prevent traditional WordPress plugins from supporting advanced AI features. See the System Architecture section below for technical details on how these innovations work together.

🏗 System Architecture

NV oOS implements a comprehensive orchestration layer for managing AI operations during real-time streaming events. The system architecture comprises:

• 18 language-model providers — OpenAI, Gemini, Anthropic, DeepSeek, OpenRouter, Baseten, Kimi (Moonshot), Z.AI (GLM), DigitalOcean, NVIDIA NIM, Cloudflare Worker AI, Azure, Ollama, LM Studio, Hugging Face, TypeSafe Jev (decision-only), plus local WebLLM + embedded engines
• ~1,665 tool classes (~352 base + ~1,313 Pro; live count via WP_MCP_AI_Tool_Registry::get_tools() is authoritative) registered through a singleton Tool Registry
• 36 REST controllers (16 base + 20 pro) under the mcp-ai/v1 namespace
• 64 service classes powering orchestration, budgets, and workflows
• 5 authentication methods — WordPress nonce, assistant credentials, mesh keys, Auth0 JWT, guest tokens
• Toolkit MCP servers — per-toolkit JSON-RPC 2.0 servers exposed under /wp-json/mcp-ai-pro/v1/mcp/{slug}; discoverable at /.well-known/mcp
• 8 inline-async-tick consumers — cooperative tick-lock pattern eliminates WP-Cron startup latency for background jobs (transcript mining, async tool executor, SaaS Apply, Crawl4AI, Docs Hub rebuild, Graphify reindex, Harness eval, Gemini Veo polling)
• 7 LLM Harness layers (+ 1 Pro) — opt-in epistemic layers A–H activated per-assistant via the LLM Harness metabox
• Orchestration Phases 1–7 — HITL approval queue, prompt-injection detector, structured output, OTel exporter, DAG builder, durable runs, triggers/webhooks, sub-agents

📖 For a detailed explanation of how NV oOS extends standard SSE and MCP protocols with novel orchestration features, see ORCHESTRATION-LAYER-ARCHITECTURE.md

Core Orchestration Layer: Overcoming PHP's Limitations

Critical Context: Most real-time AI streaming systems are built with Node.js, Python FastAPI, or Go — platforms designed for asynchronous, event-driven operations. These platforms natively support:

• Long-lived connections and persistent state
• Non-blocking I/O and parallel execution
• Event loops and asynchronous callbacks
• WebSocket protocols and SSE streaming

NV oOS achieves the same capabilities in PHP/WordPress — an environment fundamentally not designed for these patterns — through a sophisticated orchestration layer that creates a "persistent-behavior illusion":

1. Real-Time Budget Enforcement - Monitors token/memory usage during streaming, prevents exhaustion through predictive allocation
2. Capability-Based Tool Gating - WordPress role-based access control for AI tool execution
3. Predictive Optimization - Analyzes usage patterns to prevent resource overruns before they occur
4. Distributed Orchestration - Multi-provider support with policy-aware routing
5. Auditability & Compliance - Complete governance layer with logging and rate limiting
6. Cron-Based Task Orchestration - Extends orchestration to async operations with budget inheritance

Multi-Agent Orchestration Enhancement (DeepSeek V4-Inspired)

Added: January 2026 (v1.1.0)

Building upon the core orchestration layer, NV oOS now includes a sophisticated multi-agent coordination framework inspired by DeepSeek V4's orchestration patterns:

Key Components:

• Agent Role System - Four specialized roles (Planner, Executor, Critic, Specialist) with role-specific capabilities
• Team Composition - Automated team assembly based on task requirements and profession expertise
• Coordinated Workflows - Multi-step workflows with agent delegation, result aggregation, and validation
• Team CPT Integration - Persistent team configurations with orchestration modes (single/sequential/parallel/swarm)
• Profession-Based Discovery - 296 professions auto-assigned agent roles via intelligent seeding across 17 knowledge bases

Example Multi-Agent Workflow:

// 1. Compose research team (planner + executors + critic)
$orchestrator = new WP_MCP_AI_Agent_Team_Orchestrator();
$team = $orchestrator->compose_team( array( 'task_type' => 'research' ) );

// 2. Execute coordinated workflow
// Planner decomposes task → Executors research subtasks → 
// Communication service aggregates → Critic validates quality
$result = $orchestrator->execute_team_workflow( $team, $task, $context );

Documentation:

• See Multi-Agent Orchestration for complete technical details
• See DEEPSEEK-V4-README.md for documentation suite overview
• See DEEPSEEK-V4-USAGE-GUIDE.md for practical examples

Why This Architecture Is Novel: Overcoming PHP's Limitations

• Event loops and background workers

PHP/WordPress, by contrast, is fundamentally request-based:

• Every HTTP request spawns a new process that dies after responding
• I/O operations block execution
• No persistent memory between requests
• No native event loop or async coordination

NV oOS solves this by implementing an orchestration layer that creates a "persistent-behavior illusion" — effectively recreating Node.js's event loop behavior within WordPress's synchronous, request-based architecture. This architectural compensation is the system's core technical innovation:

This makes NV oOS patent-worthy as a technical workaround — it achieves sophisticated AI orchestration in an environment specifically not designed for such patterns. See ORCHESTRATION-LAYER-ARCHITECTURE.md for the complete technical analysis.

Computer-Implemented Resource Management

The system operates as a computer-implemented method executing on a processor with memory, performing:

1. Dynamic Resource Budget Allocation: The orchestration layer dynamically allocates token and memory budgets to tool execution requests based on real-time system capacity and operation requirements. The WP_MCP_AI_Resource_Manager continuously monitors server resources (PHP memory limits, execution time constraints) and automatically adjusts operational parameters.

2. Capability-Based Access Control: Tool execution endpoints enforce granular capability-based access controls. Each tool in the registry declares required WordPress capabilities, and the REST API controller validates user permissions before allowing execution. This ensures secure, policy-driven access to all operations.

3. Registry-State-Based Scheduling: The WP_MCP_AI_Tool_Registry maintains tool availability state and schedules execution based on policy constraints. Tools are loaded conditionally based on dependency availability, and execution is scheduled according to assistant configuration and user permissions.

4. Metrics-Driven Budget Adjustment: The system continuously monitors execution metrics (memory usage, API response times, token consumption) and adjusts resource budgets in response to prevent resource exhaustion and reduce latency. The WP_MCP_AI_Token_Budget_Manager implements safety margins and dynamic chunking to prevent API limit overruns.

System Components

The system comprises a processor and memory storing instructions that:

• Monitor real-time resource availability through PHP runtime introspection
• Enforce capability checks at REST endpoint boundaries
• Schedule tool execution through a centralized registry
• Adjust token and memory budgets based on detected system metrics
• Maintain operation logs for audit and optimization

This architecture is embodied in non-transitory computer-readable media (PHP source files) that, when executed by a web server processor, cause the system to perform the complete resource management workflow. The implementation prioritizes stability, security, and efficient resource utilization across diverse hosting environments.

Symfony Process Integration (December 2025)

NV oOS Pro addon integrates the Symfony Process component for secure external command execution. This modern framework replaces direct exec() calls in 6 Pro tools and 2 supporting services, providing:

• Enhanced Security: Proper argument escaping and command validation
• Timeout Management: Configurable timeouts with graceful handling
• Better Error Handling: Comprehensive exception catching and WordPress-friendly error reporting
• Process Control: Real-time output streaming and cancellation support

Migrated Tools & Services:

• FFmpeg operations (video frame extraction, metadata reading)
• Python rembg (background removal)
• WP-CLI execution
• Meta AI Jukebox (music generation)
• Supporting services for video and audio processing

The Process Service (WP_MCP_AI_Process_Service) provides WordPress-friendly wrappers with WP_Error integration, making external process execution consistent with WordPress coding standards.【F:includes/services/class-wp-mcp-ai-process-service.php†L1-L220】【F:docs/history/2025/implementations/symfony-phases/SYMFONY_PHASE2B_PROCESS_INTEGRATION.md†L1-L100】

📜 Release History

The 12 most recent releases are documented in full below. Every older release has a one-line summary in the Previous Releases table, and complete per-release detail lives in CHANGELOG.md.

v1.1.97 — October 6, 2026

ECC-Inspired Harness Enhancements, OOS Parity Closure & Gateway Protocol Negotiation

• 🤖 Agent-harness enhancements (PR #6912, Proposal 056 — all inert by default). Cascade model routing on both engines — the framework-core CascadeRouter (extends ProviderRouter, injected into the OOS engine via the WordPress CascadeClassifier/CascadeValidator adapters) and the live legacy base+Pro path (WP_MCP_AI_Cascade_Executor gating WP_MCP_AI_Language_Model_Router::create_chat_completion(), with Pro Jev classifier wiring via the proposal-045 routing signal); one wp_mcp_ai_cascade_enabled switch + classifier/validator seam govern both paths, completing the previously dormant tier machinery with post-call validation + escalation. run_assistant_eval (+1 base) — three trajectory judges (tool selection / use / response quality) + deterministic checks + the wp_mcp_ai_eval_judge seam. Hook profiles — minimal/standard/strict gating, meta → filter → option → constant resolution, master switch + kill list (includes/hooks/). scan_assistant_security (+1 base) — OWASP-informed config audit with masked secret evidence + posture score. Session distiller — session-end transcript → canonical wp_mcp_ai_memory_stored event (MemPalace/CCT/Graphify/recall hydrate; zero provider cost by default). suggest_workflows_from_history (+1 Pro) — mines the harness trace store for recurring tool chains (frequency × recency × success). ECC MIT attribution in CREDITS.md + @credit tags.
• 🛡 OOS parity gaps closed (PR #6913). The four Proposal 029 G2 guardrail gaps that blocked promoting the OOS engine switch: the destructive-ops gate no longer becomes an HTTP 500 on the OOS path (typed catch → canonical 428 confirmation / 429 limits envelopes via WP_MCP_AI_REST::translate_gate_exception()); apply_pre_response_render() runs the Output Guardrail + Citation Verifier at the final payload on all three surfaces (legacy non-streaming, legacy streaming's final SSE event — blocked → error event, and the OOS handler); the OOS bridge fires wp_mcp_ai_agentic_iteration_complete with the same (iteration, assistant_id) pair; the before-tool-execute waterfall bridge gains regression tests. bin/check-chat-parity.php now tracks all four surfaces (35 → 38 features, 100% parity) with a parity-check CI job.
• 🔧 Gateway protocol negotiation fixed (PR #6914). The MCP Gateway's initialize no longer hardcodes 2026-07-28 — negotiateProtocolVersion() echoes the highest mutually supported version (2026-07-28 → 2024-11-05), reads protocolVersion + supportedProtocolVersions, and defaults to 2024-11-05 — so Zed and other strict MCP clients stop aborting with "Unsupported protocol version".
• 📦 npm publish wave (PRs #6898, #6904–#6909, #6911). The v1.1.96-deferred public npm publish is executed: all 23 packages/ live on the public registry at 0.1.0-alpha.3 (the bridge's latest = 0.1.0-alpha.3 with the bare-npx bin alias; nvoos-api/nvoos-sse-client republished at 0.1.0-alpha.4 after the TypeScript was stripped from their shipped JS). The publish workflow loops over all 23 packages with an optional subset input, GitHub Packages covers all 23, and the new npm-maintenance.yml adds dry-run-by-default unpublish/dist-tag-add.
• 📦 Versioning — bumped to 1.1.97 across all version-bearing files. Pro addon: 1.1.97. MCP Gateway: 0.1.1 (0.1.0 → 0.1.1 post-window via #6925 — the non-force mirror sync; the negotiation fix shipped on the 0.1.0 line). nvoos-mcp-bridge: 0.1.0-alpha.3 (published on the public npm registry, latest; nvoos-api/nvoos-sse-client at 0.1.0-alpha.4; the other 21 packages at 0.1.0-alpha.3). Media Worker: 3.4.0 (unchanged). Media Studio: 0.6.1 (unchanged). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). ChatGPT Plugin addon: 0.1.0 (unchanged). Fleet Operator: 1.0.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — zero plugins/ diff in-window). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.10.03 (unchanged — zero catalog diff in-window). Tool count: ~352 base + ~1,313 Pro (~1,665 total — +2 base +1 Pro) — run_assistant_eval + scan_assistant_security (base) and suggest_workflows_from_history (Pro) from #6912 (live registry authoritative; the dual-layer lib/core CascadeRouter is never counted). Provider count: 18 chat providers (unchanged). Addon count: 30 (unchanged). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 62 (unchanged). Stale build ZIPs removed: the 1.1.95 oOS set (9 root + 2 optional-components + 19 toolkit-addons = 30 files).

v1.1.96 — October 5, 2026

NV oOS MCP Bridge npx Package, MCP Gateway Addon & Per-Request Memory Cut

• 🔌 nvoos-mcp-bridge npx package (PR #6897, Proposals 054/055). @nvdigitalsolutions/nvoos-mcp-bridge — a zero-dependency stdio↔HTTP relay npm package (nvoos-mcp + nvoos-mcp-ssh bins) that connects Zed / Claude Desktop / Cursor / Codex to any NV oOS site with one npx command; byte-identity CI drift gate against the repo bin/, hermetic node:test suites (14/14), opt-in Docker E2E, and the public-npm publish workflow. Fleet Operator config generators — generate_zed_json() / generate_claude_json() emit copy-paste context_servers / mcpServers blocks through the npx package (the one-time operator result + admin page offer Zed / VS Code / Claude Desktop alongside the Hermes YAML). The public npm publish itself is deferred on the scope claim + NPM_TOKEN.
• 🚪 MCP Gateway addon (PR #6897, Proposal 055 — addon 0.1.0, inventory #31). addons/mcp-gateway/ exposes the whole fleet as one public MCP endpoint — a media-worker-style Express service speaking stateless streamable HTTP (MCP 2026-07-28) with public API-key auth + rotation, per-key rate limits, <site-slug>.<tool> tool namespacing, graceful per-site degradation, and fail-closed env config (39/39 node:test suites; subtree-mirror workflow; Velocity deployment guide). The mirror repo, Velocity deployment, and directory submissions are deferred on maintainer credentials.
• 🧠 Per-request memory cut + bootstrap hardening (PR #6896). Plain front-end requests no longer load ~350 tool classes — the registry defers defaults until first tool access (wp_mcp_ai_is_plugin_runtime_context(): REST/admin/AJAX/cron/CLI load eagerly; third-party tools win slug conflicts via ensure_default_tools_loaded()), the OOS engine pre-warm (12 adapters + 12 provider clients + ~140 core tools) is gated off page views, and the model-catalog migration reads the ~250–400 KB JSON only on filemtime change. Pro blob options (wp_mcp_ai_pro_remote_sites, workflows, per-workflow execution logs, Ezuite SKUs, approvals, media-worker secrets, pricing buffers capped at 500, slash-command history, license key) write autoload=false with the one-time wp_mcp_ai_pro_repair_option_autoload() repair. A bootstrap-integrity guard + fail-soft loader chain (wp_mcp_ai_require_if_exists()) make missing files after a partial update degrade instead of fataling — skipped files recorded in a non-autoloaded option, one-time admin notice, self-healing prune; updater VERIFY_FILES +17 paths; new Site Health wp_mcp_ai_file_integrity test. Nine dead FF/Yahoo registry entries removed and the Inline_Async_Tick_Trait double-declaration activation fatal fixed (trait_exists() guard). A/B probe: 42 MB deferred vs 74 MB eager per front-end render (~32 MB saved).
• 🔧 Media Worker runtime version strings fixed (PR #6895). A deployed 3.4.0 worker no longer reports 3.3.0 — both /api/health version: fields and the WORKER_VERSION heartbeat constant now match package.json (the five-spot version-sync checklist lands in .context/media-worker.md).
• 📦 Versioning — bumped to 1.1.96 across all version-bearing files. Pro addon: 1.1.96. Media Worker: 3.4.0 (unchanged). Fleet Operator: 1.0.0 (unchanged — inventory-row correction from the stale 0.1.0). MCP Gateway: 0.1.0 (new — inventory #31). Media Studio: 0.6.1 (unchanged). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). ChatGPT Plugin addon: 0.1.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — zero plugins/ diff in-window). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.10.03 (unchanged). Tool count: ~350 base + ~1,312 Pro (~1,662 total — unchanged) — no registrations in-window (live registry authoritative). Provider count: 18 chat providers (unchanged). Addon count: 30 (+1 — the MCP Gateway addon). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 62 (unchanged). Stale build ZIPs removed: the 1.1.94 oOS set (9 root + 2 optional-components + 19 toolkit-addons = 30 files).

v1.1.95 — October 4, 2026

Toolkit Hardening Wave, Real Image Processing, Toolkit Labels & SSE Keepalives

• 🖼 Real image processing (PR #6881, issue #6877 — Media Worker 3.3.0 → 3.4.0). The four placeholder image tools now run real processing through the media worker sidecar: Wave 1 Sharp-native POST /api/image/enhance + /api/image/upscale (factor 2/4/8, lanczos3, 8192px cap) behind the new Pro WP_MCP_AI_Sharp_Image_Processing trait drives enhance_image_quality/upscale_image_ai (honest wp_mcp_ai_sharp_unavailable fallback); Wave 2 POST /api/image/edit (colorize | style_transfer, Gemini → OpenAI → Replicate auto-resolution) behind the new WP_MCP_AI_Provider_Image_Edit trait drives colorize_image/apply_artistic_style — success meta written only on real success. All six ported files mirrored byte-identical to CG Pro; Wave 3 (Real-ESRGAN) deferred per the plan.
• 🔨 Nine-toolkit hardening wave (PRs #6875–#6893). Confirmed production fatals fixed across ten clusters under the new mcp-ai-wpoos-toolkit-audit skill: document-generation (unguarded shell_exec(), the nonexistent Gemini generate_content(), array-content flattening, the repaired fenced-JSON regex, the Anthropic OCR provider); image-production (prompt-optimizer provider bypass, Stability/sidecar guards); CRM (the Gemini trim() fatal, the dead wp_mcp_ai_chat_completion() AI path); ECA management (the research_eca preg_match() fatal); ecommerce (unguarded exec() → WP_MCP_AI_ALLOW_SHELL_TOOLS + Process Service); media (nested WP_Error reads, the year_month directory traversal closed, the unreferenced-media meta fail-safe); healthcare (HIPAA audit-trail rewire + OpenMed guards); shopify-sync (the count(null) failure-path fatal + MCP-server descriptor wiring); orchestration (envelope + parser flattening, cross-tool is_wp_error() guards); and the loose incident tools + shared services (never-loadable execute() signatures, the nonexistent Paper Store maybe_commit(), Qdrant/gmail/roboflow/data-quality guards). Email-marketing verified clean (tests only). All mirrors byte-identical to CG Pro where they exist.
• 🏷 Toolkit labels (PR #6874). list_mcp_tools derives toolkits for previously-unlabelled Pro tools via the new WP_MCP_AI_Tool_Registry::get_tool_toolkit() resolver + wp_mcp_ai_tool_toolkit filter; the Newsletter 9.4.6+ bootstrap fatal is fixed (loader requires it on plugins_loaded), and create_company moves to the canonical write/state-changing flags.
• 📡 SSE keepalives (PR #6879). Inter-step comment frames at four chat-stream boundaries keep long agentic runs alive through proxies with idle-read timeouts; the Phase 4b deadline-triggered offload ships as the new proposal doc.
• 📦 Versioning — bumped to 1.1.95 across all version-bearing files. Pro addon: 1.1.95. Media Worker: 3.3.0 → 3.4.0. Media Studio: 0.6.1 (unchanged). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). ChatGPT Plugin addon: 0.1.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — in-window byte-identical mirrors keep the sub-project track). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.10.03 (unchanged). Tool count: ~350 base + ~1,312 Pro (~1,662 total — unchanged) — no registrations in-window. Provider count: 18 chat providers (unchanged). Addon count: 29 (unchanged). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 62 (+1 — mcp-ai-wpoos-toolkit-audit). Stale build ZIPs removed: the 1.1.93 oOS set (9 root + 2 optional-components + 19 toolkit-addons = 30 files).

v1.1.94 — October 3, 2026

ChatGPT Plugin Addon, Decision-Scope Guard & MCP Grant Enforcement

• 🤖 ChatGPT plugin addon + OAuth 2.1 resource-server (PR #6871). addons/chatgpt-plugin/ (0.1.0) packages the site's native MCP bridge for OpenAI agent surfaces — plugin.json + mcp.json with a .codex-plugin overlay, three runtime skills, bin/ stamp/package/validate scripts, subtree-split sync to the nvoos-chatgpt-plugin mirror. Base ships the RFC 9728 contract: WP_MCP_AI_OAuth_Resource_Server, /.well-known/oauth-protected-resource + /.well-known/openai-apps-challenge, 401 WWW-Authenticate on the MCP route, per-tool securitySchemes in tools/list, the nvoos_get_profile tool (+1 base), and RFC-9728-resource-identifier token acceptance in the REST authenticator.
• 🛡 Decision-scope guard (PR #6866, Proposal 052). Every Jev dispatch declares a domain (advisory/content/operations/verification) and an authority ceiling (inform → suggest → act); banned domains fail closed to the deterministic fallback; act authority requires the wp_mcp_ai_decision_act_domains filter. The WPMCPAI.Decisions.ScopeDeclared sniff (severity 5) makes undeclared dispatches a CI build failure — six existing sites declared with zero behavior change.
• 🔐 Toolkit MCP grants enforced + exposed (PR #6872). The v1.1.88 deny-by-default grant gate now actually runs outside wp-admin (the grant-lookup class definition loads in every context) — non-granted assistant-scoped tools/call returns -32601; granted servers' tools flow onto the wp_mcp_ai_chat_effective_tools seam (chat payload, tools/list, direct execution) and the estimator counts them via wp_mcp_ai_prompt_window_toolkit_tool_slugs.
• ⚖ OOS engine chat parity (PR #6872). The OOS chat path runs the same wp_mcp_ai_attention_tool_slugs → wp_mcp_ai_chat_effective_tools pipeline as the legacy path, and the WordPress EventDispatcher translates domain events into the documented legacy wp_mcp_ai_* argument tuples — fixing the ArgumentCountError → HTTP 500 on every OOS chat.
• 🔌 FlowHub MCP OAuth login proxy (PR #6867). get_oauth_client_options() resolves the connection proxy (via connection_ref or gateway-URL match) and persists it in the OAuth flow state — discovery, DCR, token exchange, refresh, and revoke all ride the proxy.
• 🔍 Safe-mode REST error masking fixed (PR #6869, closes #6860). Admins get ?verbose_errors=1 per-request detail; every masked response carries a 10-char correlation ref backed by a server-side Recent Errors entry; the original HTTP status is preserved. Ported byte-identical to the Content Graph AI RequestGuard.
• 🏥 Site Health fatal + false positives fixed (PR #6870). get_site_health pre-loads wp-admin/includes/misc.php before the wp_check_php_version() polyfill (no more redeclare fatal on WP 6.9); the database-schema check verifies wp_mcp_ai_activated_version; cold analytics caches no longer warn at 0% hit rate. Docker dev environment hardening (in-container port 8000, DISABLE_WP_CRON + cron loop, PHP 8.3 images, WP_DEBUG_LOG outside the webroot).
• 📦 Versioning — bumped to 1.1.94 across all version-bearing files. Pro addon: 1.1.94. ChatGPT Plugin addon: 0.1.0 (new — inventory #30). Media Worker: 3.3.0 (unchanged). Media Studio: 0.6.1 (unchanged). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged — in-window RequestGuard/GetSiteHealthTool/OosShadowRunner ports keep the sub-project track). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — no port waves in-window). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.10.03 (unchanged — #6865 is test-drift-only). Tool count: ~350 base + ~1,312 Pro (~1,662 total — +1 base) — nvoos_get_profile (live registry authoritative). Provider count: 18 chat providers (unchanged). Addon count: 29 (+1 — the ChatGPT Plugin addon). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 61 (unchanged). Stale build ZIPs removed: the 1.1.92 oOS set (9 root + 2 optional-components + 19 toolkit-addons = 30 files).

v1.1.93 — October 3, 2026

Fleet Status Monitoring, WP-CLI Parity, FlowHub MCP Apps & October Model Catalog

• 📡 Media worker fleet status monitoring (PR #6849 — worker 3.2.0 → 3.3.0). Opt-in (STATUS_ENABLED=1) status module: heartbeat validation, at_risk → major_outage state machine, Redis/in-memory store, dead man's switch sweeper, synthetic HTTP/TLS checks through the SSRF guard, HMAC-signed webhook + email alerts, OpenMetrics, public status page. Plugin half: WP_MCP_AI_Media_Worker_Config, heartbeat emitter, alert poller, remote_monitor source, 2 new base tools (get_fleet_status, get_site_uptime), GET /mcp-ai/v1/status/sites, /status fleet slash sub-command; Pro incident auto-create/resolve + Status Dashboard Fleet section.
• 🖥 WP-CLI parity & hardening (PR #6852, Proposal 050). includes/cli/ extraction with manage_options gating, wp mcp-ai tool call, new security/model Base commands + nine Pro command trees (crm, incident, maintenance, schedule, workflow, vault, remote-site, communication, media-studio), --fields/--format=count, canonical mcp-ai namespaces with legacy aliases, docs/operations/wp-cli.md, WP-CLI 2.9+.
• 🔌 FlowHub MCP Apps (PRs #6854, #6861). FlowHub MCP connections join the assistant MCP Apps "Add from Remote Sites" dropdown with the OAuth login flow and chat-time connection_ref resolution; the connection's proxy now carries into every MCP Apps gateway request (proxy_url/proxy_auth via http_api_curl, OAuth call sites included) — no direct egress needed on geo-blocked deployments.
• 💹 Financial Planner OpenStock parity (PR #6857, Proposal 051, B1–B3). WP_MCP_AI_Finnhub_Provider (BYO-key, rotation on 401/403/429, SWR cache, cached/realtime modes) at the yfinance filter seam; +3 Pro tools (watchlist_sync, market_overview_widget, the previously orphaned import_financial_planning_blueprint registered); price_alerts data-mode parity; openstock-market-analyst blueprint; masked key inputs.
• 🧠 October 2026 model catalog (PR #6863). Catalog v2026.10.03 — 238 models, 18 providers (Z.AI joins): gpt-6-sol/gpt-6-luna, claude-sonnet-5-5/claude-opus-5-5/claude-mythos-5-1, gemini-3.8-live, gemini-omni-1.1-flash, Cloudflare Workers AI refresh, Z.AI glm-5.3 family; codex-line/Kimi-k2 removals with migration entries; deprecations with sunset dates; defaults refreshed; Cloudflare pricing fixes.
• 🐞 Schedules, Media Studio, and validators fixed (PRs #6853, #6855, #6856, #6858, #6859). Fashion transforms pass the resolved Gemini image model (no more /ai/generate 500); tool-rule validation failures return HTTP 400; required_settings falls back to the credential resolver; Media Studio 0.6.0 → 0.6.1. Schedules no longer trigger on create/save (timing-change-only reschedules; recurring starts one interval out); the edit modal stops sending workflow_steps on non-workflow schedules. aspect_ratio: 'auto' passes the validated Gemini tools.
• 🔒 Advisory + preset/manifest repairs (PRs #6850, #6851). undici 7.30.0 across 8 addon lockfiles (GHSA-w293-vg96-wgc3); the fleet pair joins site_management preset + coverage manifests.
• 📦 Versioning — bumped to 1.1.93 across all version-bearing files. Pro addon: 1.1.93. Media Worker: v3.2.0 → v3.3.0. Media Studio: 0.6.0 → 0.6.1. SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — no port waves in-window). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.10.03 (238 models; Z.AI joins). Tool count: ~349 base + ~1,312 Pro (~1,661 total — +2 base +3 Pro) — the fleet pair + the financial trio (live registry authoritative). Provider count: 18 chat providers (Z.AI joins). Addon count: 28 (unchanged). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 61 (unchanged). Stale build ZIPs removed: the 1.1.91 oOS set (6 files) + the superseded media-studio (0.1.0/0.2.0/0.6.0) and saas-controller (0.1.0/0.2.0) addon ZIPs (5 files — 11 total).

v1.1.92 — October 2, 2026

Media Studio AI Fashion Production, Inline Vision Data URLs & Provider-Content Fixes

• 👗 Media Studio fashion production suite (PRs #6839, #6844 — addon 0.1.0 → 0.6.0). The fashion-studio AI mode lands in Media Studio: NV_oOS_Media_Studio_AI_Service drives eight industry transforms (on-model, model-swap, face-swap, background, recolor, packshot, detail-repair, try-on) through Gemini via edit_gemini_image; REST /ai/* routes carry the D-1 consent/acknowledgment gate + D-3 cost tripwires ($0.25 per-image / $10 per-job review, $100 hard cap, unknown pricing → review); a lazy-loaded SPA mode (media-library picker, presets, review-confirm dialogs, compliance chips); forced disclosure watermark on face outputs, metadata disclosure floor. Phase 2 (Pro) — mcp_ai_fashion_model identity CPT + consent gating, mcp_ai_fashion_job batch/review queue (approve/reject/reroll, WooCommerce + media collection export), the fashion_studio Pro module, fashion presets. Phase 3 (base) — marketplace output pipeline (amazon/woocommerce/social/web profiles, GD upscaling ≤2×, format conversion, alt text, white-bg validation). Phase 4 (base) — IPTC 2025.1 XMP provenance into JPEG/PNG/WebP + best-effort C2PA. Phase 5 — sidecar-backed fashion video transform + 8 new Pro tools (fashion_onmodel_generate, fashion_model_swap, fashion_background_generate, fashion_recolor, fashion_packshot, fashion_virtual_tryon, fashion_batch_job, fashion_identity_manage — registered in the main $pro_tools map, self-gate via is_available()) + the Workflow Builder's 10th fashion preset category.
• 🖼 Inline vision data URLs + payload optimization (PR #6846). submit_document_prompt and every OpenAI/DeepSeek vision path no longer hand remote image URLs to provider servers — that fetch failed on hotlink-protected, CDN-fronted, staged, or freshly uploaded media (Failed to download image from …). The new WP_MCP_AI_Image_Data_Url helper converts image segments to base64 data URLs with the WordPress server as fetcher (local attachments read off disk, remote URLs downloaded server-side); both OpenAI payload paths + DeepSeek prefer inline bytes with URL fallback; logs truncate data URLs. Oversized JPEG/PNG downscaled to a 2048px vision-tile cap, opaque PNGs re-encoded JPEG q82 (alpha/GIF/WebP preserved), header-only dimension probe, originals never modified, optimized bytes only when smaller. New Chat Client → Features → Inline Image Optimization toggle (default on, three settings layers) + wp_mcp_ai_image_inline_* filters.
• 🐞 Provider content + credential edge cases fixed (PR #6845). The six research tools fataled on array AI content (Gemini/OpenAI-compatible gateways return message.content as parts → preg_match(): Argument #2 must be of type string, array given) — the new WP_MCP_AI_Tool_Research_Content_Normalization trait flattens parts arrays at every choke point and repairs research_project's uncompilable JSON-extraction regexes; WP_MCP_AI_DeepSeek_Client::normalize_response() flattens array content blocks (chat, transcripts, tools); create_post taxonomy auto-apply matches term names on word boundaries and never auto-applies single-word tags (still suggested); the vision tools see the configured OpenAI key via the merged settings view + WP_MCP_AI_Credential_Resolver (the dashboard stores keys in the separate wp_mcp_ai_credentials option they never read).
• 🧩 Pro tool coverage manifest (PR #6847). Regenerated with the 9 fashion-stack tool classes — the test_tool_class_manifest_is_up_to_date CI failure is closed.
• 📦 Versioning — bumped to 1.1.92 across all version-bearing files. Pro addon: 1.1.92. Media Studio: 0.1.0 → 0.6.0 (own track — the fashion production suite; the in-window inventory row is already 0.6.0). Media Worker: v3.2.0 (unchanged). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged). nvoos-content-graph-pro: 1.0.0 (unchanged — no port waves in-window). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Model catalog: v2026.09.22 (unchanged — zero includes/data/ diff in-window). Tool count: ~347 base + ~1,309 Pro (~1,656 total — +8 Pro) — the fashion set (live registry authoritative). Provider count: 15 chat providers (unchanged). Addon count: 28 (unchanged). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 61 (unchanged). Stale build ZIPs removed: the 1.1.90 build set (30 files: 9 in build/ incl. 3 .sha256, 2 in build/optional-components/, 19 in build/toolkit-addons/).

v1.1.91 — October 1, 2026

FlowHub MCP Mode, OAuth Discovery & JSON Envelope Protection

• 🔌 FlowHub Connection MCP mode (PR #6836). FlowHub Remote Sites connections gain a flowhub_mode selector (api default | mcp) — MCP mode designates the connection as the backend for the FlowHub toolkit MCP server. WP_MCP_AI_FlowHub_Connection_Helper::get_mcp_connection_id() resolves the first enabled MCP-mode connection; the toolkit MCP REST controller injects its connection_id into tool arguments only when the caller passes none — MCP-triggered refresh/sync calls route through the explicit-connection path so decrypted credentials and the connection's encrypted proxy apply via http_api_curl. Explicit connection_id always wins; without an MCP-designated connection, behavior is byte-for-byte unchanged; the seam is generic (WP_MCP_AI_Toolkit_Server_Base::get_mcp_connection_id()).
• 🔐 MCP App OAuth discovery per MCP spec (PR #6835). WP_MCP_AI_MCP_App_OAuth_Client::discover_metadata() walks the full chain — RFC 8414 metadata (+ §3.2 path insertion for path-scoped servers), RFC 9728 protected-resource metadata (trying every authorization_servers entry), the 401 WWW-Authenticate: Bearer resource_metadata probe, OIDC fallback, and WordPress REST fallback. RFC 8414 documents are accepted only with both endpoints; every attempt is recorded (URL → HTTP status / transport error) and the metabox failure alert surfaces the real transport error; server-advertised default_scope is honored; parse_www_authenticate() splits multiple challenges per RFC 7235 §2.1; 10 s per-probe cap.
• 🔗 OAuth redirect allowlists (PRs #6831, #6832). LinkedIn, QuickBooks, Mailjet, and Yahoo Sports connect buttons no longer silently bounce to wp-admin — each flow allowlists its consent-page host (www.linkedin.com, appcenter.intuit.com, app.mailjet.com, api.login.yahoo.com) via the allowed_redirect_hosts filter, deriving the host from the provider's authorize-endpoint filter; matching ports land in the Content Graph AI platform.
• 🛡 JSON envelope protection (PR #6827). Orchestration CCT reads gate on the physical table — table_exists() probe → is_storage_ready() (table present and every required column; schema drift trips the gate → transients fallback, logged; per-request cache with a reset_storage_cache() seam). The new WP_MCP_AI_Db_Output_Guard::run() suppresses $wpdb error output around callbacks (logging failures with the last query) and wraps the central tool dispatch — execute_tool() + both REST tool handlers — so no surface can leak database error HTML into a JSON response again; the four agent-command-center read AJAX handlers gain with_suppressed_db_errors().
• 🐞 Security events display + double-render fixed (PRs #6829, #6830). The Security tab's "Recent Security Events (last 10)" table and the compliance CSV exporter now read the canonical event_type/ip_address keys (with legacy fallbacks), label known event slugs, and show user display names (Guest for unauthenticated) — stored entries were always correct, only the display keys were wrong. The orchestration dashboard no longer renders twice: the self-instantiating class gained a duplicate new in the bootstrap loader, registering two distinct callbacks on the same page hook (duplicate DOM IDs breaking the auto-refresh JS) — the redundant new is removed.
• 🧩 RF-DETR presets + coverage manifest (PR #6828). rfdetr_catalog_search joins the ecommerce preset ("Product operations"); rfdetr_detect stays deliberately out of presets (gated behind enable_vision_analysis_toolkit, matching its siblings); the Pro tool coverage manifest is regenerated with the two missing class basenames.
• 🔒 Dependency advisories (PRs #6833, #6834). nodemailer ^9.1.1 → ^10.0.9 in addons/pro + addons/media-worker (GHSA-g57g-f23g-4646, alerts #977/#978 — quoted local-part envelope injection) with the vendor copy spec switched to dist/ (nodemailer 10 ships no top-level lib/) and the vendor bundle refreshed 8.0.5 → 10.0.9; fast-uri override floor → >=4.1.5 across the four alert-bearing trees (GHSA-jvvf-x445-j334, alerts #979–#982 — mailto header injection), verified additive-only via npm-pack diff.
• 🌐 Ecosystem port (PR #6825). The RF-DETR vision cluster ports byte-identical into nvoos-content-graph-pro — the Roboflow inference service (three trust tiers, fail-closed credentials, SSRF/HTTPS discipline), WP_MCP_AI_Tool_Rfdetr_Detect, the vision count normalizer, the e-commerce catalog search (44-tool e-commerce registration), the D8-compat image-dHash copy, and the slim standalone init with the toolkit_vision_analysis module gate — dual-matrix green with the documented deviations; tracker row F3 appended (sub-cluster 1 complete).
• 📦 Versioning — bumped to 1.1.91 across all version-bearing files. Pro addon: 1.1.91. Media Worker: v3.2.0 (unchanged — the #6833 nodemailer bump carries no version change). SaaS Controller: 0.3.0 (unchanged). Design System addon: 0.3.0 (unchanged). nvoos-content-graph: 1.0.8 (unchanged). nvoos-content-graph-ai: 1.0.4 (unchanged). nvoos-content-graph-ai-platform: 2.0.0 (unchanged — OAuth ports + build refresh). nvoos-content-graph-pro: 1.0.0 (unchanged — two in-feature ports landed: the RF-DETR cluster (#6825) and the FlowHub MCP-mode mirrors (#6836)). Checkout API: 0.1.2 (unchanged). Docs Hub addon: 0.5.1 (unchanged). Comic Reader addon: 0.5.0 (unchanged). Chat SPA addon: 0.7.0 + Canvas Toolkit addon: 0.2.0 (unchanged). Model catalog: v2026.09.22 (unchanged — no model PRs in-window). Tool count: ~347 base + ~1,301 Pro (~1,648 total — unchanged) — no tool registrations in-window (FlowHub MCP mode is a connection-binding seam; #6828 only wires the existing RF-DETR pair into presets/manifest; live registry authoritative). Provider count: 15 chat providers (unchanged). Addon count: 28 (unchanged). Bundled skills: 75 base + 41 Pro (unchanged). Coding-time agent skills: 61 (unchanged). Stale build ZIPs removed: the 1.1.89 build set (30 files: 9 in build/ incl. 3 .sha256, 2 in build/optional-components/, 19 in build/toolkit-addons/).

v1.1.90 — September 30, 2026

RF-DETR Vision Cognition, Strict MCP Scope & Memory Identity Closure

• 🔍 RF-DETR vision cognition (PR #6824, Proposal 049). New Pro WP_MCP_AI_Roboflow_Inference_Service — one HTTP client across three trust tiers (key-less self-hosted Docker Inference server on loopback/private hosts; dedicated deployments; Serverless Cloud API) with fail-closed credentials (va_roboflow_api_key as the raw Authorization header for serverless/dedicated, HTTPS