tang-agents

Fault-tolerant multi-agent system for collaborative task execution with policy drafting, review, dispatch, parallel ministry actions, and comprehensive fault...

Security Scan

The code and instructions align with a local multi-agent framework: it doesn't request secrets or external install URLs and appears to implement the described departments and fault‑tolerance features, with only minor packaging quirks to review.

Purpose & Capability

The package name/description match the included Python modules: agents for drafting, review, dispatch, and six ministries are implemented. No unrelated credentials, binaries, or config paths are requested. Minor note: requirements.txt lists 'asyncio' (part of the stdlib) and 'aiohttp' — aiohttp is plausible for async networking but the presence of asyncio on PyPI is unnecessary.

Instruction Scope

SKILL.md only instructs installing requirements and running the TangSystem example. It does not instruct reading unrelated files, environment variables, or sending data to unknown endpoints. The runtime instructions are scoped to running the local multi-agent flow.

Install Mechanism

There is no custom install spec; SKILL.md advises 'pip install -r requirements.txt' which pulls packages from PyPI. This is common but has moderate risk compared to a purely instruction-only skill because dependencies (aiohttp) will be installed from the public registry. The 'asyncio' entry is unnecessary and may be a packaging oversight.

Credentials

The skill declares no required environment variables, no primary credential, and no config paths. The code shown does not access environment secrets or external auth tokens.

Persistence & Privilege

The skill is not forcibly persistent (always:false) and uses normal agent invocation. It does not request system-wide changes or modify other skills' configurations in the provided code.

Assessment

This skill appears internally consistent and implements the described local multi-agent framework. Before installing: (1) inspect the rest of tang_agents.py to confirm there are no hidden network calls, hardcoded endpoints, or code that sends data externally (the file provided in the package should be reviewed end-to-end); (2) verify whether aiohttp is actually used — if not, the requirements file can be tightened; (3) run the package in a sandboxed environment first (or in a virtualenv) to observe network activity and avoid side effects; (4) prefer pinned dependency versions from trusted sources, and be cautious about installing packages from PyPI in sensitive environments. If you want, I can scan the remaining portion of tang_agents.py for any network requests or suspicious patterns (HTTP requests, sockets, subprocess usage) and report back.

Tang Agents - Multi-Agent Collaboration System

Overview

A fault-tolerant multi-agent collaboration framework inspired by the Tang Dynasty's Three Departments and Six Ministries system. This system implements separation of powers, parallel execution, and comprehensive fault tolerance.

Architecture

User Request
    ↓
[Zhongshu] Policy Drafting
    ↓
[Menxia] Review & Veto (Checks & Balances)
    ↓
[Shangshu] Task Dispatch
    ↓
[Six Ministries] Parallel Execution
    - Libu: Permissions
    - Hubu: Resources
    - Libu-Content: Content Generation
    - Bingbu: Security
    - Xingbu: Compliance
    - Gongbu: Deployment
    ↓
Output / Human Approval

Installation

pip install -r requirements.txt

Quick Start

import asyncio
from tang_agents import TangSystem

async def main():
    system = TangSystem()
    
    # Process a task
    edict_id = await system.process("Your task description")
    
    # Check status
    status = system.dashboard.get_status(edict_id)
    print(status)
    
    # Display dashboard
    system.dashboard.display()

asyncio.run(main())

Fault Tolerance Features

1. Timeout Handling: Each agent has independent timeout (3-15s)
2. Circuit Breaker: Auto-pause after 3 consecutive failures
3. Graceful Degradation: Return template data on failure
4. Failure Isolation: One ministry failure doesn't crash others
5. Retry Mechanism: Menxia review allows up to 3 retries

License

MIT