████████╗██╗  ██╗███████╗    ██╗  ██╗██╗██╗   ██╗███████╗
╚══██╔══╝██║  ██║██╔════╝    ██║  ██║██║██║   ██║██╔════╝
   ██║   ███████║█████╗      ███████║██║██║   ██║█████╗  
   ██║   ██╔══██║██╔══╝      ██╔══██║██║╚██╗ ██╔╝██╔══╝  
   ██║   ██║  ██║███████╗    ██║  ██║██║ ╚████╔╝ ███████╗
   ╚═╝   ╚═╝  ╚═╝╚══════╝    ╚═╝  ╚═╝╚═╝  ╚═══╝  ╚══════╝

─────────────────────────────────────────────
  trigger diagnostics  ·  behavioral telemetry
  hermes agent  ·  openclaw  ·  pre-v1
─────────────────────────────────────────────

STATUS

Pre-v1. Active development starts after two upstream commitments clear. What lives here today is the fixture corpus — production-shape multi-skill agent workflows the v1 tooling will run against.

WHAT THIS BECOMES

The Hive will be a skill trigger diagnostic and behavioral telemetry layer for Hermes Agent and OpenClaw. Two questions, one tool:

• Predictive: given a skill corpus and a set of realistic prompts, which skill should fire, which actually would, and which collide?
• Empirical: once skills are running, which one did fire, on what prompt, at what cost, with what outcome?

The gap between prediction and reality is where the useful findings live. Skill authors don't currently have a way to test whether their description actually triggers on the phrasings users send. Operators don't have a clean view of which installed skills earn their context budget. The Hive fills both.

FIXTURE CORPUS

Production-shape workflows with real integrations (Google Calendar, Stripe, multi-language flows). These exist as the dogfood corpus for the v1 trigger diagnostic — realistic targets, not toy examples.

   THE HIVE
   ├── skills/
   │   ├── restaurant-agent ────── Bookings, menu, upselling
   │   ├── real-estate-agent ───── Lead qualification, multilingual
   │   └── restaurant-agent-openclaw ── Cross-platform proof
   └── tools/
       ├── calendar_tool.py ────── Google Calendar integration
       └── paywall/app.py ─────── Stripe subscription billing

---

AGENT ROSTER

SKILL:restaurant-agent

Framework: Hermes Agent

Front-of-house AI that handles the entire reservation flow. Asks for party size, date, time, name, phone — then books it directly on Google Calendar. No human touches the booking.

• Reservation management with real calendar events
• Menu and dietary inquiries (halal, vegan, gluten-free)
• Natural upselling (sharing platters for groups, weekend specials)
• Complaint escalation to human manager
• 24/7 on WhatsApp + Telegram

CUSTOMER: Hi, table for 4 Saturday at 7pm
AGENT:    Perfect. Name and phone number?
CUSTOMER: Johnson, 050-123-4567
AGENT:    Booked. Saturday 7pm, party of 4.
          Our sharing platters are great for groups — ask your server.
          [✓ Calendar event created]

SKILL:real-estate-agent

Framework: Hermes Agent

Lead qualification engine that speaks English, Arabic, and French. Captures budget, timeline, area preference, and pre-approval status. Scores every lead as hot, warm, or cold.

• Buyer and seller qualification
• Multi-language (EN / AR / FR)
• Viewing bookings with calendar integration
• Automatic lead scoring
• Property matching against active listings

CUSTOMER: مرحبا، أبحث عن شقة غرفتين نوم بميزانية 250 ألف
AGENT:    مرحبا! ما المنطقة المفضلة؟ وما الجدول الزمني للانتقال؟
          هل لديك موافقة مبدئية على التمويل؟

SKILL:restaurant-agent-openclaw

Framework: OpenClaw

Same skill, different framework. Proves cross-platform deployment — one skill definition works on both Hermes and OpenClaw without rewriting.

---

TOOLS

calendar_tool.py

Direct Google Calendar API integration. No MCP middleware, no timeouts. Creates, lists, and deletes calendar events from the command line or agent skill.

python3 tools/calendar_tool.py create "Reservation - Smith, party of 4" "2026-06-01T19:00:00" 120

paywall/app.py

Stripe subscription paywall for freemium AI chatbots. One free question, then $19.95/month unlimited access. Full Checkout + webhook flow.

STRIPE_SECRET_KEY=sk_test_xxx python3 tools/paywall/app.py
# → http://localhost:8080
# Test card: 4242 4242 4242 4242

FREE QUESTION ──→ AI RESPONDS ──→ PAYWALL ──→ STRIPE CHECKOUT ──→ UNLIMITED

---

DEPLOYMENT

# Install Hermes Agent
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

# Copy skills into Hermes
cp -r skills/restaurant-agent ~/.hermes/skills/
cp -r skills/real-estate-agent ~/.hermes/skills/

# Install OpenClaw
npm install -g openclaw@latest

# Copy OpenClaw skill
openclaw skills install skills/restaurant-agent-openclaw/

# Set up Google Calendar
pip3 install google-api-python-client google-auth-oauthlib
python3 tools/calendar_tool.py list-calendars

# Launch
hermes chat

---

STACK

FRAMEWORKS    Hermes Agent · OpenClaw
INTELLIGENCE  LLM-judge backend — model-agnostic
              default       Claude Sonnet 4 (best accuracy-per-dollar in current testing)
              routing       OpenRouter
              under test    Claude · Llama · DeepSeek · GPT (multi-model comparison v1.x)
RUNTIME       Python 3.11+ · CLI
TELEMETRY     SQLite (v2)
OUTPUT        Markdown · JSON · CI exit codes

---

ROADMAP

v1 — Skill trigger diagnostic

• SKILL.md parser (frontmatter + body, Hermes and OpenClaw)
• Prompt-set schema with expected-fire annotations
• LLM-judge: ranked skill selection per prompt with confidence
• Per-skill context cost + collision report
• Markdown + JSON output, CI exit codes

v2 — Behavioral telemetry

• Runtime tracer for Hermes and OpenClaw
• SQLite trace store (compatible interface planned)
• Prediction vs observation reconciliation reports

Things this will never become: voice agents, CRM connectors, skill generation, security scanning. Scope discipline is the whole point.

Tetsuo Labs · 0xTetsuo