phpkaiharness

phpkaiharness is a pluggable AI Agent Harness for web app (php web app). It gives you an LLM execution loop enriched with safety, Cognitive Graph Memory ..., all without leaving the PHP ecosystem.

Comment

🚀 PHPKAIHARNESS 3.0

The Autonomous AI Agent Operating Layer for Web APP

Production-grade AI orchestration, cognitive memory, semantic intelligence, and enterprise safety, all inside the Web APP ecosystem.

🌟 Overview

phpkaiharness is a next-generation AI Agent Harness built for Web applications. It transforms LLMs into autonomous, context-aware agents capable of:

  • Reasoning
  • Remembering
  • Verifying
  • Safely interacting with application data
  • Learning across sessions

Combining Cognitive Graph Memory, Quantum Memory, Semantic Cache, RAG Context Injection, and Enterprise Guardrails, the platform delivers faster, smarter, and more reliable AI experiences while reducing token costs.

⚡ Core Features

🤖 Autonomous Agent Engine

Drive the complete:

Think → Act → Observe

execution cycle with:

  • Autonomous tool usage
  • Configurable max iterations
  • SSE token streaming
  • Runaway-loop protection

🔀 Intelligent LLM Failover

Automatic provider failover chain:

Qwen Cloud
   ↓
Ollama
   ↓
LM Studio
   ↓
OpenRouter
   ↓
Laravel AI

Capabilities

  • Semantic Cache
  • Ontological Context Injection
  • Cognitive Graph Memory
  • Provider circuit breaking
  • Provider prioritization
  • High availability architecture

💾 Semantic Cache

SQLite-backed semantic response caching.

Benefits

✅ Up to 97% faster repeated requests

✅ Reduced API costs

✅ Zero-token cache hits

Example

Hot tier ratio for 500GB/day

≈

Hot tier sizing for 500GB per day

Both can return the same cached result.

🕵️ PII Protection Layer

Automatic redaction before requests leave your application.

Protected Data

  • Email addresses
  • IP addresses
  • Credit card numbers
  • API keys

Masking is applied to:

  • Incoming prompts
  • Outgoing responses

⏱️ Rate Limiting

Token-bucket rate limiter backed by SQLite.

Protects against:

  • HTTP 429 errors
  • API cost spikes
  • Excessive request bursts

🛡️ Enterprise Guardrails

Policy-driven execution control.

Features

  • Tool allowlists
  • Tool denylists
  • Argument validation
  • Terminal protection
  • Scope enforcement

Risky actions are blocked before execution.

🧠 Model Prompt Optimizer

Automatically rewrites prompts for optimal performance.

Supported Profiles

  • Qwen 3.5
  • Gemma 4

Benefits:

  • Better tool calling
  • Improved instruction following
  • Higher reasoning accuracy

🔗 Ontological Context Injection

RAG-powered live data enrichment.

Workflow

User Prompt
      ↓
Embedding Search
      ↓
Relevant Eloquent Records
      ↓
Context Injection
      ↓
LLM Response

Grounds responses using real application data.

🕸️ Cognitive Graph Memory

Persistent cross-session knowledge graph.

Capabilities

  • Entity extraction
  • Relationship mapping
  • Long-term memory
  • Multi-turn reasoning
  • Knowledge accumulation

⚛️ Quantum Memory Harness

Quantum-inspired intelligent memory retrieval.

Scoring Formula

S_fused = α · S_cos + β · S_interfere

Where:

  • S_cos = cosine similarity
  • S_interfere = phase interference score

Advantages

  • Multi-hop traversal
  • Entangled memory discovery
  • Contextual recall
  • Enhanced relevance scoring

✅ Draft Verification

Every response goes through a secondary validation step:

Draft Generated
      ↓
Verification Pass
      ↓
Fact Validation
      ↓
Final Response

Reduces hallucinations and factual inaccuracies.

💡 Thinking Budget

Structured reasoning injection:

Think
  ↓
Act
  ↓
Observe

Improves planning quality and complex task execution.

📟 Cyber HUD Dashboard

A futuristic cyber-teal monitoring center.

Includes

  • Real-time workflow tracing
  • Session explorer
  • Agent playground
  • Feature configuration panel
  • Telemetry analytics
  • Live status indicators

Every feature displays:

ACTIVE
or
DEACTIVATED

in real time.

☁️ Native Qwen Cloud Integration

Built-in DashScope support.

Features

  • Hybrid credential resolution
  • Structured JSON output
  • Streaming support
  • Qwen reasoning control
  • Shared application configuration

Resolution chain:

global_settings
      ↓
Harness Config
      ↓
Laravel AI SDK
      ↓
Environment Variables

🏗️ Technology Stack

PHP 8.5
Laravel 13
SQLite
Qwen Cloud
Laravel AI SDK
Guzzle HTTP
PSR-14 Events
Pest Testing

📊 Benchmark Results

Benchmarked on a production Laravel 13 CTI platform.

🚀 Cache Performance

7 / 17 Requests
= 41%

served directly from cache.

Average Response Time

53.87 ms

versus

2710 ms

for raw API calls.

💰 Token Savings

41%

of benchmark requests consumed:

0 Tokens

Projected production cache hit rate:

70–90%

🧠 Response Quality

Harness Output:

2812 Characters

Raw API:

1990 Characters

Result

+41% richer responses

through memory-enhanced context.

📚 Knowledge Growth

After only 17 sessions:

Cognitive Memory

84 Facts

stored in the graph.

Quantum Memory

182 Nodes

created and linked.

✅ Accuracy Improvements

  • Zero hallucinations on cached database queries
  • Verified tool results stored permanently
  • Memory compounds over time
  • Institutional knowledge evolves continuously

🛠 Major Challenges Solved

OpenAI Tool Call Compatibility

Converted internal format:

{
  "id": "tool1",
  "name": "search",
  "arguments": {}
}

to OpenAI-compatible format:

{
  "id": "tool1",
  "type": "function",
  "function": {
    "name": "search",
    "arguments": "{}"
  }
}

Qwen Thinking Model Hangs

Automatically applies:

enable_thinking = false

to:

  • qwen3
  • qwq

preventing infinite reasoning loops.

Unified Credential Resolution

Implemented a 5-level fallback chain:

Host Database
      ↓
Laravel AI SDK
      ↓
Harness Config
      ↓
.env
      ↓
Default Values

Windows Compatibility

Replaced:

getenv('HOME')

with:

storage_path()

for universal support.

🏆 Achievements

✅ Production Ready

✅ 93 Passing Tests

✅ Cognitive Knowledge Graph

✅ Quantum Memory Retrieval

✅ Real-Time Telemetry

✅ Multi-Provider AI Support

✅ Enterprise Safety Controls

✅ Zero Dependencies Beyond LLM APIs

🔮 What's Next

Vector-Powered Semantic Cache

Replacing Levenshtein matching with:

  • pgvector
  • sqlite-vec
  • Qwen Embeddings
  • Native vector stores

Multi-Agent Orchestration

Router Agent
      ↓
Specialized Agents
      ↓
Collaborative Execution

Powered by:

  • qwen-turbo
  • qwen-plus
  • qwen-max

Real-Time WebSocket Telemetry

Future support for:

  • Laravel Reverb
  • WebSockets
  • Live token streams
  • Workflow visualization

Distributed Memory Graph

Scaling memory using:

  • Tenant sharding
  • WAL optimization
  • Faster retrieval
  • Concurrent writes

Packagist Release

Installation becomes:

composer require kai/phpkaiharness

Memory Quality Decay

Implementing:

  • Automatic relevance decay
  • Fact quality scoring
  • Knowledge graph cleanup
  • Noise reduction

🚀 Vision

phpkaiharness is evolving from an AI framework into a self-improving intelligence platform that remembers, verifies, learns, and compounds knowledge over time.

The first request gets an answer. The thousandth request inherits institutional memory. 🧠⚡

Built With

Share this project:

Updates

Submission history