NEXUS AI Hub

Technical summary

NEXUS AI Hub is the AI agent orchestrator running on Linux servers and multi-cloud environments (successor of oci-manager): central registry of satellite nodes with heartbeats, hardware metrics and latency, smart routing of tasks, comments and ideas to 8 specialists, LGPD/GDPR sanitization of secrets, smart truncate of huge logs and remote node management.

Executed scope

• plugin.json: v1.0.0, slug nexus-ai-hub, legacy_slugs ["oci-manager","ai-hub"], "AI" category, is_core: false, public author_url and website (perciocastelo.com.br).
• Registration: 2 tabs (ai-hub and nexus-ai-hub legacy → tab.php), "AI Hub" doc (bot icon), oci-v2-assets JS, 5 modals (oci-explorer, oci-preview, oci-logs, ia-chat-test, ai-hub-settings).
• Central schema (legacy, consumed via DB::fetchAll, no DDL in the plugin) — oci_nodes (id PK, name, ip, hostname, node_key st- + 24 bytes, provider, status, last_pulse, persona_name, persona_avatar) and oci_ai_endpoints (id PK, node_id FK, ip, url on port /v1/chat/opencode, model, status, last_seen_at).
• Personas — 18 agents with avatars (assets/ia/people1|robo): automatic assignment without repetition (filters used ones, with fallback), re-assignment with force, master detected by SERVER_ADDR, SITE_URL hostname, "Independent" provider or id 7.
• AI panel routes — ai_explain_task (explain task), ai_explain_comment (explain comment with context), ai_ask_ideas (ideas/refactors via Torvalds), ai_upload_report (PDF via pdftotext or plain text: LGPD sanitization + segmentation into 8,000-char chunks analyzed sequentially, "Part X of Y" header).
• Multi-node routing — active /v1/chat/opencode endpoints deduplicated per node, nsx01 priority, 90s global budget, node_lock_<id> locks (180s; 300s on failure), X-Proxy-Secret/X-Nexus-Proxy-Secret headers + Host, rewrite to http://ip, 55s timeout (1.5s connect); local fallback run_local_opencode (35s, opencode-web setuid wrapper for www-data, build-line filter).
• 8 specialist subagents — JSON manifests with keywords (+1) and regex (+3), minimum 1: Torvalds (ideas), SecurityAI, HASHAI and CVEAI (security category preserves hashes — they are the forensic analysis target), NginxAI, ApacheAI (web servers), HestiaAI (sysadmin), PipelineAI (CI/CD devops).
• Orchestrated delegation — sleep(3) between typing presence and comments (Linus delegates → the specialist introduces itself) via presence.php (working/idle, released in finally).
• LGPD/GDPR — lgpd_sanitize_content masks Google/OpenAI/AMZN keys, tokens, passwords, emails, .pem/.key/.cert/.crt/.pub files, private key blocks and hashes (except security forensics); privacy alert at the top of the reply when sanitized.
• External references — Semgrep rule (semgrep.dev/r), CVE via MITRE API (cveawg.mitre.org) and Wikipedia OpenSearch (3 terms with generic-word blacklist).
• Smart truncate — description >7,000 chars → first 4,500 + last 2,000; last 15 comments capped at 1,500 (1,000+400); final 12,000 safety.
• Node management — registration preserves the existing node_key; remote commands via m.php (purge, rebuild_llm, update_agent, logs lines 1–500, reboot, shutdown; 8s/20s timeout; X-Nexus-Proxy-Secret; NEXUS-Dashboard/7.6 User-Agent; "sudo requires password" live-log detection); synchronous latency ping via /health (3s).
• Security — IP whitelist (allowed_ips) enforced on all 4 AI routes; sync token rotation (st- + 24 bytes with created_at) and optional node_key rotation (kgn<ip>_ + 16 bytes).
• API: 17 actions — oci_rotate_tokens, oci_token_status, oci_nodes_list, oci_nodes_status (maps persona_name/is_master), oci_endpoints_list, oci_node_delete, oci_node_register, oci_node_assign_avatar, oci_endpoint_register, oci_node_command, oci_node_ping, ai_explain_task, ai_explain_comment, ai_ask_ideas (session_write_close for parallel processing), ai_upload_report ($_POST merge), ai_hub_settings_get/set (allowed_ips + persona_category humans|robots).
• UI (tab.php, 295 lines, slate + violet/blue theme) — Active AI Agents table (persona avatar, violet badge, last_pulse, LLAMA: UP/DOWN, IP/provider with ms latency, 12-point CPU/RAM sparklines, purge/rebuild_llm/update_agent/restart/logs/SSH/copy key/delete actions, "Master Shield Active" on the master) and IA Endpoints Active (model, provider, URL, chat test); top bar with settings (whitelist), masked sync token with countdown + rotation, Refresh (refreshOciAll() — renamed from refreshAll() to avoid the Alpine.js namespace clash with the core) and "Recruit Node".
• Fastr — /oci-nodes → oci_nodes_status (no argument). No cron (on demand).
• Documented Safe Mode — persistent opencode serve --port 8092 daemon keeps the model in memory and a Python wrapper on port 8091 intercepts /v1/chat/opencode via lightweight sockets (40s+ → ~2.8s on 1 Core ARM/1GB instances); telemetry/title bypass (agent.title.disable) with a whitelisted light model; MemoryMax=300M removed from systemd using native swap (2.5GB).

Stack and tools

• PHP 8 (no framework)
• MySQL 8 (oci_nodes, oci_ai_endpoints, node_lock_*/oci_sync_token/ai_hub_settings configs)
• Alpine.js + Tailwind CSS (premium light, slate/violet/blue)
• cURL + REST (/v1/chat/opencode endpoints, m.php, health)
• OpenCode (Safe Mode daemon v3.2 + local fallback)
• NEXUS plugin system (PluginManager: tabs, docs, API actions, modals, JS, Fastr)

Operational tags

• IA
• Agentes
• Orquestração
• Multi-cloud
• OCI
• Satélites
• OpenCode
• Subagentes
• LGPD
• DevOps
• Plugin NEXUS

Operational result

• Multi-cloud fleet: Oracle Cloud, OVH and independent nodes with heartbeat, CPU/RAM sparklines and ms latency — all in the panel.
• Automatic routing: tasks, comments and idea requests are forwarded to the active, least-loaded node (nsx01 priority), with anti-race locks and local fallback.
• Visible specialists: 8 subagents delegate with typing messages and their own personas — the user sees who is analyzing what.
• Privacy by default: secrets and hashes are masked before leaving for the cloud; security forensics preserve the hashes (the analysis target).
• Huge logs become analyses: PDF/text upload segmented into sequential parts with per-part summaries.
• Operational security: IP whitelist, rotating sync token and one-click node_key rotation.
• Manageable fleet: purge, LLM rebuild, agent update, live logs, SSH, restart and removal straight from the satellites table.
• /oci-nodes chat command.