Hackathon Project

LangClaw-Fix

LangGraph-powered AI agents for OpenClaw — bringing stateful, multi-model workflows to Telegram with code-enforced safety and autonomous tool use.

ZhipuAI GLM LangGraph OpenClaw Telegram Python 3.11+

View on GitHub Explore ↓

The Problem

What Traditional Skills Can't Do

LangGraph agents as OpenClaw skill backends unlock a class of scenarios that prompt-only and script-based skills structurally cannot handle.

Medical Triage

Multi-turn state tracking + safety gates — must confirm allergy history before proceeding, track which checks are complete across a long conversation.

State Tracking Human-in-the-Loop

Vehicle Diagnostics

Autonomous decisions — agent judges whether this is a DIY fix or requires a mechanic, based on severity, user skill level, and part availability.

Autonomous Decisions Multi-Model

IT Helpdesk

Multi-turn state + dynamic decisions — remember which troubleshooting steps were already tried, adapt the next suggestion based on what failed.

State Tracking Autonomous Decisions

Home Inspection

Multi-model orchestration — vision model assesses damage from photos, reasoning model estimates repair cost, search model finds local contractors.

Multi-Model Tool Orchestration

LangGraph enables: Multi-Turn State → Human-in-the-Loop Gates → Autonomous Decisions → Multi-Model Orchestration

Showcase

Universal Appliance Repair

A single conversation, multiple AI capabilities — from photo identification to guided repair to automatic escalation.

Identify appliance make, model, and error code from photo

Vision

Safety gate — "Have you checked for gas smell?"

Human-in-the-loop

Validate safety response — is it genuinely safe?

LLM Judgment

Diagnose root cause from symptoms and error codes

Reasoning

Search for model-specific repair manuals (only when needed)

Autonomous Tool

Walk user through fix step by step, adapting to feedback

Agent Loop

Escalate with formal report if repair fails after 3 attempts

Enforced Policy

The Solution

LangGraph Agents for OpenClaw

Stateful

SQLite checkpoint survives restarts, tracks every attempt precisely

Safe

LLM judge with ambiguity handling + code-level injection detection

Observable

Every model call logged with real-time tracing — no black boxes

Autonomous

Agent decides when to instruct, ask, search, or escalate based on context

Multi-Model

4 ZhipuAI GLM models, each chosen for its specific strength

Deployed

Runs on Telegram via OpenClaw with async polling (no timeouts)

How It Works

Architecture

→

OpenClaw Gateway

→

HTTP Daemon :19292

→

LangGraph StateGraph

vision_triage

GLM-4.6V

safety_gate

interrupt

safety_classify

GLM-4-Plus

diagnosis

GLM-4-Plus

agent_decide

GLM-4-Plus

manual_search

GLM-5

escalation

GLM-4.5-Air

complete

—

Async polling: daemon returns {status:"processing"} immediately. OpenClaw polls /poll_result/{SID}. Avoids timeouts on slow vision calls (30–60s). Request deduplication via description + photo hash.

AI Models

ZhipuAI Model Allocation

4 models, each chosen for its specific strength

Task	Model	Why
Photo analysis	GLM-4.6V	Multimodal vision — identifies appliance from photo
Safety classification	GLM-4-Plus	Reliable LLM judge with ambiguity handling
Root cause diagnosis	GLM-4-Plus	Reasoning over symptoms and error codes
Repair guidance	GLM-4-Plus	Autonomous decision-making (instruct/ask/search/escalate)
Manual search	GLM-5	Built-in web search tool support
Escalation report	GLM-4.5-Air	Structured, professional writing

Proof

Benchmark Comparison

Same repair scenario, three approaches

Metric	LangGraph Agent	Script Wrapper	Prompt-Only
Total time	55.5 s	81.0 s	N/A
Backend tokens	1,664	1,510	0
Total tokens	4,064	3,610	11,100
Safety enforcement	LLM judge + ambiguity	LLM YES/NO	Prompt (bypassable)
State persistence	SQLite checkpoint	JSON file	None
Agent autonomy	Full ✓	None ✗	Prompt ✗
Process startup	0 s (daemon)	~3 s / call	N/A

vs Script Wrapper: 31% faster with comparable token cost. vs Prompt-Only: 63% fewer tokens. The only approach with real safety enforcement and agent autonomy.

Stack

Tech Stack

AI Models

ZhipuAI GLM-4.6V · GLM-4-Plus · GLM-5 · GLM-4.5-Air

Agent Framework

LangGraph

Platform

OpenClaw → Telegram via SKILL.md

Runtime

Python 3.11+ · asyncio · httpx

Deployment

Async HTTP daemon

Safety

LLM judge · ambiguity handling · injection detection

Get Started

Quick Start

# Clone and setup

git clone https://github.com/nyang01/LangClaw_FIX.git

cd LangClaw_FIX && chmod +x setup.sh && ./setup.sh

# Configure API key (.env)

ZAI_API_KEY=your-zhipuai-api-key

# Start daemon

cd agent && python daemon.py