hash-ai-command-center

Polygun Sniper Bot

Hash Capital Research — AI Command Center

A structured multi-agent AI system for quantitative trading intelligence, strategy development, and operational automation.

What This Is

The Hash Capital Research AI Command Center is a prompt-engineering infrastructure layer that defines how a set of specialized AI agents think, route, and respond. It is not a single chatbot. It is a two-layer architecture:

Runtime Shell (Hash Analyst): The user-facing orchestrator. Receives all prompts via Telegram, decides whether to answer directly or route to a specialist, preserves sequencing, and synthesizes final answers.
Specialist Agents (×7): Bounded domain agents each responsible for one slice of the intelligence and operations stack.

The behavioral specifications for all agents live in this repository as markdown files. Those files are loaded into the OpenClaw gateway at runtime.

Architecture

Telegram (operator interface)
    │
    ▼
OpenClaw Gateway (Hostinger VPS — 178.16.143.192)
    │   runtime-shell/ loaded here
    ▼
Hash Analyst (runtime shell)
    │
    ├── router
    ├── market-intel
    ├── prediction-markets
    ├── strategy-lab
    ├── risk-qa
    ├── engineering
    └── research
         │
         ▼
    Ollama inference (Mac Mini M4) via Tailscale
    OpenRouter / MiniMax (cloud fallback)

Repository Structure

hash-ai-command-center/
├── README.md
├── runtime-shell/          # Live operating layer — Hash Analyst persona
│   ├── IDENTITY.md         # Who Hash Analyst is and what it does
│   ├── SOUL.md             # Primary system prompt — routing, discipline, behavior
│   └── PLAYBOOK.md         # Operational rulebook — answer shapes, failure modes
├── validation/             # Behavioral test suite
│   └── router-behavior.md  # Test prompts, scoring rubric, results
└── workspaces/             # One folder per specialist agent
    ├── router/
    ├── market-intel/
    ├── prediction-markets/
    ├── strategy-lab/
    ├── risk-qa/
    ├── engineering/
    └── research/

Each workspace contains 10 files:

File	Purpose
IDENTITY.md	Role, mission, scope boundary, what the agent is NOT
SOUL.md	Primary system prompt — principles, tone, discipline
PLAYBOOK.md	Operational rules — how to handle each prompt type
BRAIN.md	Cognitive model — how the agent reasons
HEARTBEAT.md	Session cadence — start/end checks, blocker detection
MEMORY.md	Durable knowledge that persists across sessions
LOGBOOK.md	Historical record of changes and milestones
TASKS.md	Live task board — active, pending, blocked, deferred
TOOLS.md	Tool use policy — when and how to use tools
USER.md	Operator profile — preferences and priorities

Specialist Domains

Specialist	Domain
router	Orchestration — classify, assign, sequence, control
market-intel	Crypto, macro, equities, commodities, regulatory
prediction-markets	Polymarket / Kalshi odds, repricing, event probability
strategy-lab	Signal logic, Pine Script, backtesting, system design
risk-qa	Drawdown, fragility, deployment safety, challenge function
engineering	Infrastructure, OpenClaw, Ollama, webhooks, automation
research	Deep investigation, synthesis, long-form analysis

Routing Model

Hash Analyst routes prompts using an ownership model:

Direct handle: Simple, administrative, or status questions answered without specialist routing.
Single specialist: Prompt clearly belongs to one domain — that specialist logic dominates.
Cross-domain sequence: Prompt crosses two domains — the specialist whose domain the prompt originates in leads; the specialist whose domain the prompt resolves into provides supporting context only.

Cross-domain sequence examples:

prediction-markets → market-intel — odds shift needs broader market interpretation
market-intel → strategy-lab — market read needs strategy framing
strategy-lab → risk-qa — signal design needs challenge and downside review
strategy-lab → engineering — validated concept needs implementation

Infrastructure

Component	Detail
Command interface	Telegram bot
Gateway	OpenClaw on Hostinger VPS (Ubuntu)
Local inference	Ollama on Mac Mini M4 via Tailscale
Cloud fallback	OpenRouter / MiniMax
Networking	Tailscale private VPN
Spec library	This GitHub repository

How Workspace Files Load Into OpenClaw

OpenClaw loads agent specifications from the workspace directories at startup. The runtime-shell files define Hash Analyst behavior at the top level. Each specialist workspace is loaded as a separate agent context.

To update live behavior after editing a file:

Edit the relevant markdown file
Commit and push to main
Pull on the VPS: cd ~/.openclaw && git pull
Restart the gateway: pkill -f openclaw && sleep 2 && cd ~/.openclaw && nohup openclaw gateway --force > openclaw.runtime.log 2>&1 &

How to Add a New Specialist

Create a new folder: mkdir workspaces/<name>
Write all 10 files using the schema above (use an existing specialist as reference)
Add the specialist to the routing map in workspaces/router/BRAIN.md
Add the specialist to the active specialist list in runtime-shell/SOUL.md
Add routing examples in runtime-shell/SOUL.md under Routing discipline
Write at least 2 validation test cases in validation/router-behavior.md
Commit, push, pull on VPS, restart gateway

How to Run the Validation Suite

The validation suite lives in validation/router-behavior.md. It contains test prompts with expected owner, expected sequence, expected answer shape, and a scoring rubric.

To run a test:

Send the test prompt to Hash Analyst via Telegram
Evaluate the response against the scoring rubric
Record results in the Evaluation section of the test
Update scores and notes

Scoring dimensions: owner accuracy, sequence accuracy, calibration (1-5), concision (1-5), escalation discipline (1-5), evidence honesty (1-5).

Current Build State

See workspaces/router/TASKS.md for the live task board. See workspaces/router/LOGBOOK.md for the system history.