IntentGuard × OpenClaw — Brain / Body Architecture

The Differentiation

IntentGuard thinks. OpenClaw acts. wrapper.ts connects them.

🧠 IntentGuard Brain

The cognitive engine. Measures trust, authenticates identity, and decides what's permissible based on mathematical proof.

Core Functions

8-step Trust Debt pipeline (agents 0-7)
FIM geometric authentication (20-dimensional identity vector)
Sovereignty scoring (0.0–1.0 trust metric)
Federation: bot-to-bot tensor overlap, drift detection, quarantine
Grid state sync to tesseract.nu (12-cell Reality Bridge)
CEO Loop: 24/7 autonomous executor

Patent-Pending Formula

TrustDebt = Σ((Intent_i - Reality_i)² × Time_i × SpecAge_i × Weight_i)

Key Files

src/wrapper.ts — Brain↔Body connector
src/runtime.ts — Sovereign Engine orchestrator
src/ceo-loop.ts — Always-on executor
src/pipeline/runner.ts — 8-step pipeline
src/auth/geometric.ts — FIM identity + permissions
src/federation/ — Bot-to-bot trust network

🦞 OpenClaw Body

The execution layer. Receives decisions, routes to terminals, captures output, and reports back through every communication channel.

Core Functions

Discord bot with 9 cognitive room channels
Terminal dispatch via native IPC (no focus stealing)
ThetaSteer categorization (12×12 tesseract grid)
3-tier LLM routing: Ollama → Sonnet → Opus
Bidirectional I/O: dispatch prompt → capture output → post to Discord
43 skills (voice, screen, browser, email, messaging, IoT)

Multi-Channel Gateway

ws://127.0.0.1:18789 — OpenAI-compatible API

Key Files

openclaw/src/runtime.ts — Discord bot + orchestrator
openclaw/skills/claude-flow-bridge.ts — Terminal router
openclaw/skills/thetasteer-categorize.ts — 12×12 grid
openclaw/src/task-store.ts — Task lifecycle journal
openclaw/src/output-poller.ts — Terminal output capture
openclaw/src/scheduler.ts — Proactive cron jobs

⚡ System Architecture Bridge

┌──────────────────────────────────────────────────────────────────────┐ │ 🧠 INTENTGUARD (Brain) │ │ │ │ ┌─────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────┐ │ │ │ 8-Step │ │ FIM Geometric│ │ Sovereignty │ │ CEO Loop │ │ │ │ Pipeline │ │ Auth Engine │ │ Score Engine │ │ (24/7) │ │ │ │ (agents 0-7) │ │ (20-dim vec) │ │ (0.0→1.0) │ │ │ │ │ └──────┬──────┘ └──────┬───────┘ └──────┬───────┘ └────┬─────┘ │ │ │ │ │ │ │ │ └────────────────┼─────────────────┼────────────────┘ │ │ │ │ │ ├──────────────────────────┼─────────────────┼────────────────────────┤ │ wrapper.ts FIM interceptor │ │ spawns OpenClaw hooks tool calls │ │ as child process for permission checks │ ├──────────────────────────┼─────────────────┼────────────────────────┤ │ │ │ │ │ 🦞 OPENCLAW (Body) │ │ │ │ ▼ ▼ │ │ ┌──────────────────────────────────────────────────────────────┐ │ │ │ Gateway ws://127.0.0.1:18789 (OpenAI-compatible API) │ │ │ │ ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐ ┌────────────┐│ │ │ │ │Discord │ │WhatsApp│ │Telegram│ │ Slack │ │ iMessage ││ │ │ │ └───┬────┘ └────────┘ └────────┘ └────────┘ └────────────┘│ │ │ └──────┼───────────────────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────┐ ┌─────────────┐ ┌──────────────────────┐ │ │ │ ThetaSteer │──▶│ Model Route │──▶│ Claude Flow Bridge │ │ │ │ Categorize │ │ H1→Ollama │ │ 9 Terminal Rooms │ │ │ │ 12×12 Grid │ │ H3→Sonnet │ │ Parallel IPC │ │ │ │ │ │ H5→Opus │ │ │ │ │ └──────────────┘ └─────────────┘ └──────────┬───────────┘ │ │ │ │ │ ┌──────┬──────┬──────┬──────┬──────┬──────┬──────┬──────┬──────┐│ │ │🔨 │📐 │🎩 │🔒 │🎤 │🧪 │🎭 │🧭 │🌐 ││ │ │build │archi │oper │vault │voice │lab │perf │nav │net ││ │ │iTerm │Code │kitty │Wez │Term │Cursor│Term │rio │Msgs ││ │ └──────┴──────┴──────┴──────┴──────┴──────┴──────┴──────┴──────┘│ │ │ output capture │ │ ▼ │ │ ┌──────────────────────┐ │ │ │ data/spec.md (living)│ ← learnings appended │ │ │ data/tasks.jsonl │ ← task journal │ │ │ tesseract.nu/api │ ← grid state sync │ │ └──────────────────────┘ │ └──────────────────────────────────────────────────────────────────────┘

9 Cognitive Rooms

Each room maps to a terminal app with its own IPC method

🔨 Builder iTerm2 (AppleScript) C1 Operations.Grid

📐 Architect VS Code (System Events) A2 Strategy.Goal

🎩 Operator Kitty (remote socket) C3 Operations.Flow

🔒 Vault WezTerm (CLI) A1 Strategy.Law

🎤 Voice Terminal (AppleScript) B3 Tactics.Signal

🧪 Laboratory Cursor (System Events) C2 Operations.Loop

🎭 Performer Terminal (AppleScript) A3 Strategy.Fund

🧭 Navigator Rio (System Events) B1 Tactics.Speed

🌐 Network Messages (System Events) B2 Tactics.Deal

8-Step Trust Debt Pipeline

IntentGuard's brain: agents 0-7 measure the gap between intent and reality

Outcome Requirements Parser

Report HTML, docs

0-outcome-requirements.json

Database Indexer

Requirements + codebase

1-indexed-keywords.json

Category Generator

Keywords

2-categories-balanced.json

Matrix Builder

Message Execution Flow

From Discord message to terminal output to spec update

Discord message received

Voice memo, text, or reaction in any room channel

ThetaSteer categorization

Maps to 12×12 grid → full notation + hardness (1-5) + target model

FIM auth interceptor (Brain check)

IntentGuard validates: Identity ∩ Requirement ≥ Threshold

Model routing

H1-2 → Ollama (local) | H3-4 → Sonnet (API) | H5 → Opus (API)

Terminal dispatch

Claude Flow Bridge sends to room's terminal via native IPC (no focus steal)

Output capture + stabilization

Poll every 2s → detect shell prompt → 5s stability → complete

Post results to Discord

Status emoji + summary + output (split >1900 chars to file attachment)

Spec update + grid sync

Learnings appended to data/spec.md → sovereignty updated → tesseract.nu synced

Skills Dashboard

43 eligible, 12 remaining (env vars / config needed)

Eligible Skills (43)

Skill	Category	Status
🔐 1password	Security	eligible
📝 apple-notes	Productivity	eligible
⏰ apple-reminders	Productivity	eligible
🐻 bear-notes	Productivity	eligible
📰 blogwatcher	Content	eligible
🫐 blucli	Audio	eligible
📸 camsnap	Media	eligible
📦 clawhub	Skills	eligible
🧩 coding-agent	Dev	eligible
🎮 discord	Comms	eligible
♊️ gemini	AI	eligible
🧲 gifgrep	Media	eligible
🐙 github	Dev	eligible
🎮 gog	Workspace	eligible
📦 healthcheck	System	eligible
📧 himalaya	Email	eligible
📨 imsg	Messaging	eligible
📦 mcporter	MCP	eligible
📊 model-usage	Analytics	eligible
📄 nano-pdf	Docs	eligible
🖼️ openai-image-gen	AI	eligible
🎙️ openai-whisper	Speech	eligible
☁️ openai-whisper-api	Speech	eligible
🧿 oracle	AI	eligible
🛵 ordercli	Commerce	eligible
👀 peekaboo	Screen	eligible
📜 session-logs	Debug	eligible
📦 skill-creator	Skills	eligible
🌊 songsee	Audio	eligible
🔊 sonoscli	Audio	eligible
🎵 spotify-player	Audio	eligible
🧾 summarize	Content	eligible
✅ things-mac	Productivity	eligible
🧵 tmux	Terminal	eligible
🎞️ video-frames	Media	eligible
📱 wacli	Messaging	eligible
🌤️ weather	Utility	eligible
🌊 claude-flow-bridge	Workspace	eligible
📡 output-capture	Workspace	eligible
🎮 system-control	Workspace	eligible
🧭 terminal-dispatch	Workspace	eligible
🧊 thetasteer-categorize	Workspace	eligible
🎤 voice-memo-reactor	Workspace	eligible

Blocked Skills (12)

These need API keys or service configuration — not installable binaries.

Skill	Missing	Type
🫧 bluebubbles	`channels.bluebubbles`	config
🎛️ eightctl	`eightctl`	binary
📍 goplaces	`GOOGLE_PLACES_API_KEY`	env
🍌 nano-banana-pro	`GEMINI_API_KEY`	env
📝 notion	`NOTION_API_KEY`	env
💎 obsidian	`obsidian-cli`	binary
💡 openhue	`openhue`	binary
🗣️ sag	`ELEVENLABS_API_KEY`	env
🗣️ sherpa-onnx-tts	`SHERPA_ONNX_*`	env
💬 slack	`channels.slack`	config
📋 trello	`TRELLO_API_KEY`	env
📞 voice-call	`plugins.voice-call`	config

What Fires Next

Immediate: Set API keys (Notion, Trello, ElevenLabs, Gemini) to unlock 4 more
Short-term: Configure Slack + BlueBubbles channels for messaging skills
Medium-term: eightctl, openhue, obsidian-cli need upstream releases or custom builds
Target: 50+/55 eligible (90%+ coverage)

Current State → Next State

The differentiation gap and what needs bridging

🧠 Brain: Where We Are

Pipeline: 8-step Trust Debt engine functional (agents 0-7)
Auth: FIM geometric 20-dim identity vector implemented
Sovereignty: Self-scoring engine (currently Grade C: 4,423 units)
Federation: Tensor overlap + drift detection built but untested at scale
CEO Loop: 24/7 autonomous executor operational
Grid: tesseract.nu reader/writer connected

Gaps

Pipeline runs manually (npm run queen) — not triggered by body events
FIM interceptor installed but permission thresholds may be too permissive
Federation untested between real instances
Sovereignty score doesn't yet gate body actions dynamically

🦞 Body: Where We Are

Discord: Bot live with 9 room channels + commands
Skills: 43/55 eligible (78% coverage)
Terminals: 9 rooms with native IPC (parallel safe)
Categorization: ThetaSteer 12×12 grid + hardness routing
Proactive: Queue drain (30s), retry (5m), health check (5m), doc scan (6h)
Overnight: 5-room parallel pipeline runner

Gaps

Gateway runs as launchd service but body runtime still manual (npm start)
No WebSocket reconnection — Discord disconnect = dead bot
Output poller can miss fast tasks (2s poll interval)
System Events rooms (Cursor, VS Code, Rio) still fragile (focus stealing)

⚡ Bridge: Where We Need To Go

Phase 1: Tighten the Brain↔Body Loop

Dynamic sovereignty gating: Trust Debt score should throttle body actions in real-time (low trust = only Ollama, high trust = Opus)
Event-driven pipeline: Body task completions should trigger Brain re-measurement (not just overnight)
FIM threshold tuning: Calibrate 20-dim thresholds based on actual usage patterns from task-store.jsonl

Phase 2: Body Resilience

Launchd the runtime: Create ai.openclaw.runtime.plist for auto-restart Discord bot
WebSocket heartbeat: Auto-reconnect on Discord disconnect
Process output hooks: Replace polling with child_process stdout streaming (already partially done)
Remaining 12 skills: Set API keys to reach 50+/55

Phase 3: Federation

Multi-instance: Run 2+ IntentGuard brains with different sovereignty vectors
Tensor handshake: Test bot-to-bot alignment scoring
Drift quarantine: Validate that divergent bots get isolated
Grid consensus: Multiple brains writing to shared tesseract.nu grid

Keep-Alive Architecture

Ollama heartbeat monitoring and token tier routing with failure isolation

🔋 Heartbeat & Routing System System

Ollama Health Classification

Every 60 seconds, OpenClaw checks http://localhost:11434/api/tags to determine Ollama health:

Healthy: Ollama responds → H1-H2 tasks routed to local models (FREE)
Degraded: Ollama slow/partial → fallback to Sonnet API ($0)
Down: Ollama unreachable → all tasks escalate to paid tiers

Every 5 minutes, health status is posted to Discord #cron channel.

Token Tier Routing Table

┌─────────┬──────────────┬───────────────┬─────────────────────────────────┐ │ Tier │ Provider │ Cost │ When to Use │ ├─────────┼──────────────┼───────────────┼─────────────────────────────────┤ │ Tier 0 │ Ollama │ FREE │ H1-H2 tasks, prototyping │ │ Tier 1 │ Sonnet 4.5 │ $0 (moderate) │ H3-H4 tasks, production │ │ Tier 2 │ Opus 4.6 │ $$ (high) │ H5 tasks, critical decisions │ │ Tier 3 │ Human │ $$$ (manual) │ Escalation, approvals, review │ └─────────┴──────────────┴───────────────┴─────────────────────────────────┘

Failure Isolation Per Subdivision

Each task subdivision has independent failure handling:

flows: Terminal failures isolated to room → other rooms unaffected
docs: Parse failures logged but don't block spec updates
discord: WebSocket reconnection on disconnect → queue preserved
cron: Missed schedule = immediate retry + Discord alert

4 Task Subdivisions

Parallel work streams with independent failure domains

🌊 flows Bridge

Claude Flow terminal tasks

→ Terminal dispatch to 9 rooms
→ Output capture & stabilization
→ IPC method per room (AppleScript, CLI, socket)
→ Result posting to Discord

📚 docs Brain

Document scanning, spec parsing

→ data/spec.md learning updates
→ Codebase indexing (agent 1)
→ Outcome requirements parsing (agent 0)
→ 6-hour overnight doc scan cron

🎮 discord Body

Messages, voice, reactions

→ Message ingestion from 9 room channels
→ Voice memo → Whisper transcription
→ Reaction-based task queue (👍/📌/❌)
→ WebSocket connection management

⏰ cron Grid

Scheduled jobs, health checks

→ Queue drain (every 30s)
→ Task retry (every 5m)
→ Ollama health check (every 60s → Discord every 5m)
→ Overnight 5-room parallel runner (3am)