Project

Master OS

A production AI operating system spanning 9 git repositories and 7 business domains. One command drives the full session loop: scouts, interview, parallel-by-brand ticket dispatch, FSM-gated merge, and daily wiki distillation.

Role Architect & Operator
Stack Python · Anthropic SDK · Claude Code · MCP · Bash hooks
Status Production — running daily
Phase Phase 6 (shipped 2026-05-07)

What it is

The Master OS is a custom AI operating system I designed, built, and operate solo across seven business domains: Sandbox (content and curriculum engine), UMANO (music events and record label), VioletStudios (AI production tools), Suite52 (artist persona), Blockwise Intelligence (algorithmic trading — hard-siloed), MaxProduct (AI product incubator), and Personal (career and professional assets).

Nine independent git repositories are unified under a single orchestration layer. One command — /start — drives the full session loop. The system runs daily, compounds institutional knowledge into a growing wiki, and enforces trust boundaries at the infrastructure level rather than relying on model instruction-following.

The premise: a solo operator running multiple businesses simultaneously can achieve 10x throughput if the orchestration layer is correct. Correctness requires infrastructure enforcement, not convention. The Master OS is that infrastructure.

Architecture highlights

Orchestration model Hybrid: main session owns the loop and retains full toolkit (Task, AskUserQuestion); subagents execute per-brand with restricted palettes
Trust boundaries PreToolUse hooks (deterministic) — not CLAUDE.md instructions (advisory). Hooks are invariants; CLAUDE.md is advice.
Merge gate FSM-validated: verifier verdict must be pass, pytest green, ruff clean, system_doctor A 19/19 — before any branch merges to main
Knowledge system 336+ page wiki with daily distillation pipeline — distillers classify, librarian writes; distillers never write directly (authority delegation preserved)
Repo topology 9 git repos unified by machine-readable registry; worktrees for isolated parallel ticket execution
Agent contracts 8 declared agent contracts validated by system_doctor; any missing contract surfaces as a housekeeping ticket

Key engineering decisions

Hooks-deterministic / advisory split

Trust boundaries that must hold 100% of the time live in PreToolUse hooks. The ticket-state-guard.json hook blocks direct writes to tickets.json — every state transition must go through scripts/tickets.py, which validates against a finite state machine. The pre-merge-gate.json hook intercepts every merge on a claude/T-* branch and fails it unless all checks pass. Correctness doesn't depend on the model reading its instructions.

4-layer Blockwise Intelligence silo defense

The live trading system is isolated by four independent layers: settings.json deny-list (Claude Code config), advisory rules file (CLAUDE.md-layer), PreToolUse boundary-guard hook (runtime interception), and safety_guard.py (Python runtime check). Any one layer can fail without breaching the silo — all four must fail simultaneously. Defense-in-depth for a system where a misread could trigger financial consequences.

Distillation with authority delegation

Each session produces logs, merged tickets, and research notes. The first /start of each calendar day runs distillation: parallel distiller subagents classify candidates as high/medium/low confidence, then route writes to designated authority agents (librarian for brand wiki pages, operator-global for cross-brand docs). Distillers never write directly — authority chains are preserved across the distillation boundary. High-confidence findings auto-promote; medium surfaces in the next interview for approval; low stays in the daily ledger only.

Intelligent cadence

/start replaced 10 discrete Phase 5 routines. Each phase is skippable based on freshness signals from session_state.py:should_run(phase). Run it 10 times in a day — it skips re-scouting if no new commits in the 9 tracked repos, skips distillation if already run today, but always runs dispatch (the ticket queue must drain). Same command whether checking in after 20 minutes or returning after a week. The system decides what's stale.

Scale (as of May 2026)

336+ wiki pages
9 git repositories
8 declared agent contracts
19/19 system_doctor checks passing
7 business domains
241+ tests passing

Evidence of shipping

Phase 6 shipped 2026-05-07 — a complete replacement of the 10-routine Phase 5 system with a single /start session loop. Parallel-by-brand dispatch is operational: concurrent ticket executors across brand worktrees with zero collision, enforced by worktree isolation and the FSM merge gate. The system has processed hundreds of tickets across all seven domains since Phase 6 launch.

The best evidence the system works: this portfolio site was planned, scaffolded, designed, and deployed via Master OS ticket execution — a ticket executor in an isolated worktree wrote the Astro pages while a parallel executor handled wiki updates.