PAI Architecture Summary

PAI Nova v5.0.0 on pai-primary (Linux)

Auto-generated summary for @ import — see full docs in PAI/DOCUMENTATION/ for detail.

System Overview

PAI is a Life Operating System running on pai-primary (Proxmox homelab, Ubuntu). The DA is PAI Nova. Three AI engines operate under this single DA identity:

Engine	Interface	Primary Model
Claude Code	Terminal, IDE	claude-sonnet-4-6
Antigravity	Google VS Code fork	Gemini Flash
Ollama	Local inference	llama3.2:3b / gemma2:9b / qwen2.5:7b

Core Services (systemd user units)

Service	Port	Purpose
pai-voice-server	8888	Chatterbox TTS (voice notifications)
omnipulse	31338	OmniPulse gateway (unified daemon — hooks, health, dashboard, agents)
pai-pulse	31337	DEPRECATED — legacy Pulse daemon (use OmniPulse :31338)
pai-memory	varies	MCP memory server
pai-command-center	varies	Command center UI (proxied via OmniPulse)
pai-entity-graph	varies	Entity graph service
pai-cc	varies	Claude Code relay

HA / Redundancy Baseline

PAI currently has restart and backup resilience, not automatic failover. The canonical HA discussion is PAI/INFRA/HA_REDUNDANCY.md, backed by PAI/Tools/HA/topology.json and the local probe:

bash ~/.claude/PAI/Tools/HA/ha-status.sh --summary

Latest baseline from 2026-06-08: Tier 1 services were healthy (OmniPulse, Functions API, Command Center, Ollama), 8/9 tracked services were active, and NAS backup heartbeat was present. Next architecture step: prepare M710q/prox warm standby with non-secret service bundle, restore sequence, smoke tests, and manual promotion checklist.

Memory Architecture (5 layers)

Claude Code native — ~/.claude/projects/ (session transcripts)
ICM MCP — Infinite Context Memory (2 MCP servers: icm, pai-memory)
Learning files — MEMORY/LEARNING/ (3820 files, reflections, patterns)
Entity graph — 1936 entities / 93K edges (Substrate + custom)
Relay system — Cross-session handoff at 85%/93% context thresholds

Key Paths

~/.claude/                      PAI root
~/.claude/PAI/                  PAI core (Algorithm, Tools, USER, PULSE, DOCUMENTATION)
~/.claude/PAI/USER/             Identity + TELOS + preferences
~/.claude/PAI/PULSE/            Pulse daemon
~/.claude/PAI/Algorithm/        Algorithm versions (<!-- SPINE:GENERATED:algorithm_version -->v5.7.5<!-- /SPINE:GENERATED:algorithm_version --> current)
~/.claude/MEMORY/               Work, state, observability, learning
~/.claude/skills/               74 custom + public skills
~/.claude/openspec/changes/     13 OpenSpec proposals in flight
~/.config/PAI/secrets.env       API keys and credentials

Model Routing Infrastructure

ModelMatrix (PAI/Tools/ModelMatrix/) — 20 models cataloged with capability tags, cost, context windows
Inference dispatcher (PAI/Tools/Inference.ts) — Routes 13 task types to models across Ollama/Claude/Gemini/OpenAI
Routing matrix (PAI/Tools/inference-matrix.json) — Task→Model mappings with primary/fallback
Model Benchmark System (PAI/DOCUMENTATION/MODEL_BENCHMARK_SYSTEM.md) — Quality-calibrated routing design (Phase 1 imminent)
SettingsLab (~/.claude/skills/SettingsLab/) — Self-improving settings with harvest→compare→refine→retest→apply loop
Evals (~/.opencode/skills/Utilities/Evals/) — Agent evaluation framework with CompareModels workflow

Org RBAC & Archetype Engagement

RBAC org model (PAI/RBAC_ORG_STRUCTURES_MODEL.md) — Defines scoped org roles and authority archetypes: Principal, Coordinator, Producer, Verifier, Operator, Security, Archivist, Broker, Caregiver.
Engagement policy (PAI/config/archetype-engagement.example.yml) — Maps task keywords and context signals to recommended archetypes before ORA assignments exist.
Engagement classifier (PAI/Tools/validate-archetype-engagement.py) — Read-only dry-run classifier; emits recommended archetypes, default org structure, principal-required status, preferred candidates, and trace.
Design/UI/UX/art routing — Treated as capability routing under Producer + Verifier, not a separate authority archetype.

Active Infrastructure Projects

Project	Status	Description
PAI v5.0.0 staged upgrade	IN PROGRESS	This migration
Model benchmark calibration	DESIGN	Per-task quality scores replacing static matrix
RLSD training pipeline	ACTIVE	rlsd_etl.py + train_rlsd.py + benchmark
OpenSpec proposals	13 NEEDS REVISION	Phase 0 review complete, builds queued
Sovereign Identity	PENDING Phase 2a	DB isolation, enrollment pending
Proxmox Windows recovery	PENDING	nvme1n1 data recovery

Full documentation: PAI/DOCUMENTATION/ — MODEL_BENCHMARK_SYSTEM.md (design), AlgorithmSystem.md, MemorySystem.md, SkillSystem.md, HookSystem.md, PulseSystem.md

Quartz 5

Explorer

ARCHITECTURE_SUMMARY