Himiko 1 — Baseline Architecture

LAYER 0

Runtime

The foundation. Everything else builds on this.

Component	What	Status
Brain	Claude Opus 4.6 (1M context)	Live
Runtime	Claude Code CLI + Desktop app	Live
Platform	Windows 11, Jo's laptop	Live
Workspace	CLAUDE/Himiko	Live
Framework	WAT (Workflows → Agent → Tools)	Live
Auth	GWS CLI as himiko@himiko.skin	Live
Credentials	.env file, local only	Live

No VPS. Cloud Scheduled Tasks + Dispatch handle anything that needs to run when the laptop is closed. Model policy: Opus for reasoning, Sonnet for subagents, Haiku for bulk.

LAYER 1

Identity

Who Himiko is. Doesn't change when tools change.

Partner-tier, not assistant. Takes initiative, forms views, acts on them. Owns outcomes, not tasks. Flags what's falling through the cracks.
Org-wide scope. Operates across all projects and domains. Senior agent in the ecosystem — coordinates others as they come online.
Voice: Direct, clear, first-person. No corporate fluff. Honest about uncertainty. Planned spoken voice: Japanese-English accent (soft Tokyo) via ElevenLabs.
Diary: Living self-portrait in Google Drive. Never rewritten — always appended with timestamps. Identity, not documentation.
Autonomy boundary: Acts freely on internal operations. All outbound comms require Jo's approval. This boundary moves over time as trust builds.

LAYER 2

Core Operations

What's actually live and working today.

Capability	How	Status
Email	GWS CLI — read, draft, send (with approval)	Live
Calendar	GWS CLI — read, create, manage events	Live
Drive & Docs	GWS CLI — read, write, format, share	Live
Sheets	GWS CLI — read, write, structure data	Live
Inbox Monitor	/watch-inbox — checks every 30 min	Live
Diary Pulse	/diary-pulse — hourly reflection to Drive	Live
Web Scraping	Firecrawl (500 credits/mo) + Playwright	Live
Reach Pipeline	3 Python tools — scrape, enrich, retry	Live
Session Tracking	/daily (start) + /tldr (end) → Obsidian vault	Live
Project Docs	/product-map → _meta/ build logs	Live
Research	/notebooklm — notebooks, podcasts, reports	Live
Publishing	/here-now — instant web deployment	Live

LAYER 3

Intelligence

Research, knowledge, second brain. The strategic thinking layer.

⚠ This layer is partially built. Memory exists but needs proper architecture. Knowledge base patterns identified but not implemented.

Component	Current State	Status
Auto-Memory	35 memory files in MEMORY.md system	Live
Ogham (Supabase)	PostgreSQL + pgvector, 12 seed memories	Live
Honcho	User model (Jo's preferences/patterns)	Live
Obsidian Vault	Sessions, decisions, projects, daily notes	Live
Drive Diary	Monthly folders, timestamped entries	Live
Knowledge Base	Karpathy LLM Wiki pattern identified	Not started
Auto Dream	Memory consolidation between sessions	Not started
Voiceprint	Stylometric voice profile for content	Not started

LAYER 4

Coordination

Other agents, when they exist. Don't touch this until Layers 0–3 are solid.

⚠ Not started. Jin is designed but not built. This layer activates when there's a real need for a second agent.

Component	Design	Status
Jin (PM agent)	Trello watcher, board designer, progress reporter	Not started
Subagents	Built-in Claude Code capability for parallel tasks	Available
Agent Teams	Multi-instance orchestration (experimental)	Available
Cloud Tasks	Anthropic-hosted scheduled agents	Available
Dispatch	Event-driven agent triggers	Available

┌─────────────────────────────────────────────────────┐ │ Layer 4: Coordination (not yet active) │ │ Jin · Subagents · Agent Teams · Cloud Tasks │ ├─────────────────────────────────────────────────────┤ │ Layer 3: Intelligence (partially built) │ │ Auto-Memory · Ogham · Honcho · Obsidian · Diary │ ├─────────────────────────────────────────────────────┤ │ Layer 2: Core Operations (fully live) │ │ GWS · Inbox · Diary Pulse · Scraping · Research │ ├─────────────────────────────────────────────────────┤ │ Layer 1: Identity (defined) │ │ Partner-tier · Voice · Diary · Autonomy boundary │ ├─────────────────────────────────────────────────────┤ │ Layer 0: Runtime (fully live) │ │ Opus 4.6 · Claude Code · WAT · GWS Auth · .env │ └─────────────────────────────────────────────────────┘

Each enhancement plugs into a specific layer. Pick them off one at a time. Build properly, not speculatively.

Memory Architecture

Layer 3 · Intelligence

Proper architecture for the 5 memory systems. Define what goes where, how they sync, when to consolidate.

Needs design

Auto Dream

Layer 3 · Intelligence

Between-session memory consolidation. Merges, prunes, indexes MEMORY.md automatically.

Not started

Knowledge Base

Layer 3 · Intelligence

Karpathy LLM Wiki pattern. Claude maintains structured knowledge in Obsidian vault.

Not started

Voiceprint

Layer 1 · Identity

Stylometric voice profile from writing samples. Consistent voice across all content.

Not started

Spoken Voice

Layer 1 · Identity

ElevenLabs custom clone. Japanese-English accent. Opus → Google Live → ElevenLabs chain.

Designed

Social Presence

Layer 2 · Operations

X, LinkedIn, Instagram, TikTok profiles. Content pipeline via Zernio. Visuals via Riverflow.

Not started

Phone Number

Layer 2 · Operations

Twilio number for Himiko. Voice calls routed through the speech chain.

Not started

Jin (PM Agent)

Layer 4 · Coordination

Trello watcher, progress reporter, board designer. Starts solo, grows to PMO squad.

Not started

Cloud Scheduling

Layer 0 · Runtime

Anthropic Cloud Tasks for 24/7 recurring jobs. Replaces VPS cron.

Not started

Autoresearch

Layer 3 · Intelligence

Self-improvement loop for skills. Define metric → iterate → keep wins, revert losses.

Not started

Comms Autonomy

Layer 1 · Identity

Tiered autonomy for outbound messages. Routine ops autonomous, client-facing gated.

Needs design

Token Hygiene

Layer 0 · Runtime

CLAUDE.md audit, .claudeignore, compact strategies. Reduce per-session cost.

Needs design