Arquitectura general
Section titled “Arquitectura general”CONTROL PLANEubuntu-ialab192.168.1.30
INFERENCE BACKEND ACTIVORX9070 · 192.168.1.50llama.cpp v2.14.0 (Vulkan/ROCm)
INVENTARIO OFFLINERX7900XT · 192.168.1.60Nodo apagado — expected_offline, no routable, no activoModelos activos
Section titled “Modelos activos”| Modelo | Estado | Host | Uso |
|---|---|---|---|
| llama-3.1-8b-instruct | Activo | 192.168.1.50:1234 | Minimal, observe, greetings, light prompts |
| qwen2.5-coder-14b-instruct | Activo | 192.168.1.50:1234 | Coding, report, architecture, reasoning |
| nomic-embed-text-v1.5 | Activo | 192.168.1.50:1234 | Embeddings, semantic recall |
| qwen3.6-27b | Desactivado | 192.168.1.50:1234 | No borrado del disco, disponible para tests manuales |
| qwen2.5-coder-32b | DOWN | 192.168.1.60:1234 | Nodo RX7900XT apagado |
Nota: active != loaded != discoverable != disabled. qwen3.6-27b está DISABLED. El modelo previsto para RX7900XT es
gpt-oss-20b-derestricted Q4_K_M.
Servicios principales
Section titled “Servicios principales”| Servicio | Puerto | Proceso | Tráfico |
|---|---|---|---|
| ailab-gateway | 8008 | openai_gateway.py | Único entrypoint de chat |
| ailab-router | 8083 | router_api.py (FastAPI) | API interna (/status, /profiles, /replay) |
| ailab-live-api | 8084 | live_api.py | API de estado, embeddings |
| ailab-hermes-status | 8095 | hermes/endpoint.py | Estado Hermes Enterprise |
| ailab-docs | 4322 | Astro preview | Documentación |
| ailab-metrics | 3010 | Next.js SSR | Dashboard público |
| ailab-heartbeat | — | Heartbeat persistente | Latido de cluster |
| ailab-live-state | — | State snapshot | Snapshot periódico |
| ailab-runner | — | GitHub Actions Runner | CI/CD |
Endpoints always-on (10/10)
Section titled “Endpoints always-on (10/10)”| Endpoint | Propósito |
|---|---|
| GET /health | Salud del gateway |
| GET /slo/health | Estado SLO completo |
| GET /runtime/sensors | Snapshot completo de sensor fusion (FASE 30I) |
| GET /runtime/maturity | Descriptores de madurez del runtime |
| GET /runtime/topology | Topología y dominio de fallo |
| GET /runtime/governance | Visibilidad de decisiones governance |
| GET /runtime/routes/semantics | Semántica operacional por route-family |
| GET /runtime/reports/discipline | Disciplina de reportes operacionales |
| GET /runtime/reports/evidence | Catálogo de evidencia y thresholds |
| GET /hermes/status | Estado completo Hermes Enterprise (SOUL, capabilities, operators, hooks, MCP, governance) |
Hermes Enterprise — componentes disponibles para consulta
Section titled “Hermes Enterprise — componentes disponibles para consulta”Hermes Enterprise Core está desplegado en runtime/hermes/ y expone los siguientes subsistemas vía GET /hermes/status en el puerto :8095:
- SOUL — semilla ontológica: truth_model, boundaries, protocols, domains, identity, schema
- Capability Registry — registro de capacidades del runtime con validación y políticas
- Operator Registry — operadores registrados con validación profunda (12 reglas)
- Hook Registry — ciclo de vida de hooks (todos declarativos,
enabled: false) - MCP Registry — servidores MCP declarados (prometheus, marketplace-mcp como planned)
- Dynamic Governance — 4 modos (NORMAL/ELEVATED/DEGRADED/LOCKDOWN), 6 señales trigger, anti-flapping 30s
Todos los componentes están en modo declarativo-validación. No hay enforcement activo ni hooks ejecutándose.
Observabilidad
Section titled “Observabilidad”| Componente | Host | Puerto |
|---|---|---|
| Prometheus | 192.168.1.40 | 9090 |
| Grafana | 192.168.1.40 | 3000 |
| Prometheus config | /home/albert/docker/monitorizacion/prometheus/ | prometheus.yml |
| Grafana provisioning | /home/albert/docker/monitorizacion/grafana/provisioning/ | Dashboards JSON auto-load |
100+ métricas ailab_*, 15 dashboards (carpeta AI-LAB, TIER 1 y TIER 2), 19 alertas activas.
Sensor fusion: 13 dominios observados, 4 métricas nuevas (ailab_sensor_fusion_*, ailab_observed_runtime_context_size_bytes).
Knowledge Base (AnythingLLM Enterprise)
Section titled “Knowledge Base (AnythingLLM Enterprise)”| Recurso | Valor |
|---|---|
| Host AnythingLLM | 192.168.1.50:3001 |
| Workspaces activos | 7 |
| Total vectores | 1304 |
| Embedder | text-embedding-multilingual-e5-small (Q8_0) |
| Documentos importados | 84 (canónico) + 53 (evidence) + 19 (MCP/A2A) + 8 (runbooks) + etc. |
| Baseline | Congelada en CP-ANYTHINGLLM-ENTERPRISE-04-COMPLETE |
| Estado | ✅ RAG E2E validation 100% |
Checkpoints principales
Section titled “Checkpoints principales”Fases tempranas (Phase 1–12)
Section titled “Fases tempranas (Phase 1–12)”phase-1-stablephase-2-stablephase-2-gpu-telemetryphase-3-grounded-opencode-runtimephase-4-opencode-router-livephase-5-cognitive-agent-routerphase-6-weighted-intent-routingphase-6-distributed-cognition-v1phase8-cognitive-observability-stablephase12-supervised-self-optimizationBloques 21–27 — Runtime profiling, memoria, SLO temprano
Section titled “Bloques 21–27 — Runtime profiling, memoria, SLO temprano”CP-21B-STABLECP-22B-STABLECP-23A-FOUNDATIONCP-23A-MEMORY-SAFECP-23A-MODEL-ALIAS-FIXCP-23B-QUALITY-GATECP-23B-RECALL-STABILITYCP-24-ANALYTICSCP-25-OPENCODE-PRODUCTIONCP-26-OPENWEBUI-PRODUCTIONCP-26.1-OBSERVABILITY-v2CP-26.1.1-COMPLETION-FINALIZATION-FIXCP-26.1.2-REPORT-ROUTING-FIXCP-26.2-UX-COGNITIVE-QUALITYCP-27-RUNTIME-STABILIZATIONBloques 28–30 — Agentes, streaming, SLO enforcement, sensor fusion
Section titled “Bloques 28–30 — Agentes, streaming, SLO enforcement, sensor fusion”CP-28.1-PLANNER-RUNTIME-SKELETON-STABLECP-28.2-READONLY-EXECUTOR-STABLECP-28.2-B-READONLY-BURNIN-STABLECP-28.3-SANDBOX-WRITE-STABLECP-28.3-B-SANDBOX-WRITE-BURNIN-STABLECP-29.2-B-STREAMING-BURNIN-STABLECP-29.3-THREE-MODEL-RUNTIME-STABLECP-29.4-SLO-ENFORCEMENT-STABLECP-29.4.4-ERROR-TAXONOMY-STABLECP-29.4.4-C-SLO-HEALTH-ENDPOINT-STABLECP-29.4.4-D-PARALLEL-TOOLCALL-HARDENING-STABLECP-30A-RUNTIME-STATE-FOUNDATION-STABLECP-30B-MODEL-STATE-AWARE-STABLECP-30E-GOVERNANCE-VISIBILITY-STABLECP-30F-ROUTE-SEMANTICS-STABLECP-30G-OPERATIONAL-REPORTING-STABLECP-30H-RUNTIME-EVIDENCE-ENFORCEMENT-STABLECP-30I-RUNTIME-SENSOR-FUSION-STABLECP-30I-B-SENSOR-FUSION-HARDENED-STABLECP-30I-C-SENSOR-SUMMARY-EXPOSURE-STABLECP-30I-D-SENSOR-SEMANTICS-NORMALIZED-STABLECP-30I-E-OPERATIONAL-RESPONSE-FORMATTING-STABLECP-30I-F-RUNTIME-COGNITIVE-COMPRESSION-STABLECP-30I-F0-RUNTIME-MODEL-ROUTING-CLEANUP-STABLECP-30I-G-RUNTIME-GROUNDING-STABLEBloques 31–36 — Madurez semántica, cognición, precisión
Section titled “Bloques 31–36 — Madurez semántica, cognición, precisión”CP-OBS-31A-OBSERVABILITY-SOURCE-OF-TRUTH-STABLECP-OBS-31A.1-PROMETHEUS-AUTHORITY-AUDIT-STABLECP-OBS-31A.2-GRAFANA-DRIFT-AUDIT-STABLECP-OBS-31A.3-RUNTIME-OBSERVABILITY-ALIGNMENT-STABLECP-OBS-31A.4-OBSERVABILITY-REMEDIATION-PLAN-STABLECP-OBS-31A.5-EXECUTOR-STABLECP-31B-RUNTIME-SEMANTIC-MATURITY-STABLECP-31B-HF1-OPENCODE-CONTEXT-ALIGNMENT-STABLECP-31C-OPERATIONAL-REPORTING-DISCIPLINE-STABLECP-35C-LIVE-AUTHORITY-BACKED-COGNITION-STABLECP-35D-OPERATIONAL-FAST-PATH-STABLECP-35D-HF1-FASTPATH-ROUTING-PRIORITY-STABLECP-36A-OPERATIONAL-INCIDENT-INTELLIGENCE-STABLECP-DEV-36X-CODEBASE-MEMORY-INTEGRATION-STABLECP-DOC-36X-GITNEXUS-STRUCTURAL-COGNITION-STABLECP-DOC-36X-SPANISH-LOCALIZATION-STABLECP-36B-RUNTIME-PRECISION-MODE-STABLEBloques 37–40 — Estabilidad, release hardening, SLO drift watch
Section titled “Bloques 37–40 — Estabilidad, release hardening, SLO drift watch”CP-38A-RUNTIME-DEEP-AUDIT-01-STABLECP-38B-GATEWAY-SHUTDOWN-GRACEFUL-01-STABLECP-38C-GITNEXUS-NAPI-ERROR-TRIAGE-01-STABLECP-38D-RUNTIME-STABILITY-SNAPSHOT-01-STABLECP-39A-OPENCODE-GATEWAY-CONTRACT-HARDENING-01-STABLECP-39B-RUNTIME-OBSERVABILITY-ALERTS-01-STABLECP-39C-COGNITIVE-HEALTH-FOLLOWUP-01-STABLECP-39E-RUNTIME-STABILIZATION-RELEASE-CLOSE-01-STABLECP-40A-POST-RELEASE-SLO-DRIFT-WATCH-01-STABLEGobernanza, routing, multi-GPU readiness
Section titled “Gobernanza, routing, multi-GPU readiness”CP-GITNEXUS-FIRST-ACTIVATION-01CP-SLO-ENFORCEMENT-01CP-VALIDATION-AUTHORITY-01CP-AUTONOMOUS-OBSERVABILITY-TRIAGE-01CP-OPERATOR-INTENT-REASONING-01CP-MULTIGPU-READINESS-01CP-DYNAMIC-NODE-REGISTRY-01CP-AI-LAB-MULTI-NODE-ROUTING-01CP-INTELLIGENT-FALLBACK-ENGINE-01CP-CAPABILITY-SCHEDULER-01CP-49A-POOL-ADMIN-API-READONLY-01Hermes Enterprise
Section titled “Hermes Enterprise”CP-E01A-SOUL-ENTERPRISE-SKELETON-STABLECP-E02A-CAPABILITY-REGISTRY-SKELETON-STABLECP-E02B-CAPABILITY-REGISTRY-VALIDATOR-STABLECP-E02C-OPERATOR-REGISTRY-VALIDATOR-STABLECP-E03A-OPERATOR-REGISTRY-SKELETON-STABLECP-E04A-HOOK-REGISTRY-SKELETON-STABLECP-E05-MCP-REGISTRY-SKELETON-STABLECP-E06-DYNAMIC-GOVERNANCE-STABLECP-E07-ENTERPRISE-RUNTIME-STATUS-ENDPOINT-STABLECP-HERMES-ENTERPRISE-CORE-01CP-HERMES-ENTERPRISE-FOUNDATION-01CP-HERMES-OPERABILITY-TUNING-01CP-HERMES-DOCS-ASTRO-ENTERPRISE-01Knowledge Base (AnythingLLM)
Section titled “Knowledge Base (AnythingLLM)”CP-ANYTHINGLLM-ENTERPRISE-04-COMPLETEFase actual
Section titled “Fase actual”El runtime se encuentra en una fase consolidada tras la finalización de:
- Hermes Enterprise Core — SOUL, Capability Registry, Operator Registry, Hook Registry, MCP Registry, Dynamic Governance y Status Endpoint (
:8095) completados y documentados. 185 tests PASS. - AnythingLLM Enterprise — Knowledge Base completa con 1304 vectores, 7 workspaces activos, RAG E2E validation 100%. Baseline congelada en CP-ANYTHINGLLM-ENTERPRISE-04-COMPLETE.
- Documentación Astro — 10 páginas Hermes Enterprise publicadas. Build exitoso (275 págs, 0 errores).
El baseline observacional se apoya en 30I-D (sensor semantics). Sobre ese baseline, el runtime añadió bloques de estabilidad (37–40), gobernanza federada y capacidades enterprise.
Multi-GPU sigue pospuesto: RX7900XT permanece como inventario expected_offline. No hay scheduler Multi-GPU implementado. Las precondiciones (30H–31B) están cerradas desde CP-31B, pero el nodo RX7900XT sigue apagado y no hay plan activo de reactivación.