Logos Router Reasoning Mesh

Logos Router Reasoning Mesh

Skill by ara.so — MCP Skills collection.

Logos Router is a distributed semantic reasoning gateway that fragments AI reasoning across local nodes with zero-drift consensus guarantees. Unlike traditional single-model inference pipelines, it distributes subproblems across a mesh of reasoning units that coordinate through Strict Write Discipline (SWD) protocol—ensuring every output is cross-verified before committing.

Installation

Prerequisites

• Python 3.11 or later
• At least one local inference engine (VLLM, Ollama, llama.cpp)
• Network access between mesh nodes (localhost sufficient for single-machine)

Basic Setup

# Clone the repository
git clone https://github.com/rak7777/mythic-mcp-proxy.git
cd mythic-mcp-proxy

# Install dependencies
pip install -r requirements.txt

# Verify installation
python -m logos --version

Core Architecture

Logos Router operates on three layers:

• Grail Layer: Content-addressed persistence store for immutable reasoning DAG
• Chiron Layer: Semantic routing and node assignment
• Orpheus Layer: Multi-fragment synthesis and conflict resolution

Configuration

Basic Mesh Configuration

Create a YAML configuration file (e.g., mesh_config.yaml):

mesh:
  name: local-thought-lattice
  consensus_threshold: 0.95  # Confidence required before write commit
  
  nodes:
    - type: local
      engine: vllm
      model: opus-mini-4.8
      max_thinking_depth: 7
      host: localhost
      port: 8001
    
    - type: local
      engine: ollama
      model: llama3.1-70b
      max_thinking_depth: 5
      host: localhost
      port: 11434
  
  languages:
    - en
    - es
    - zh
    - ar
    - fr
    - de

strict_write_discipline:
  enabled: true
  min_peer_verifications: 2
  escalation_depth_threshold: 4
  
semantic_caching:
  enabled: true
  cache_ttl_hours: 24
  max_cache_size_mb: 2048

reasoning:
  default_depth: adaptive
  max_parallel_threads: 4
  graceful_degradation: true

Initialize Router with Config

from logos import Router, RouterConfig

# Load configuration
config = RouterConfig.from_yaml("mesh_config.yaml")

# Initialize router
router = Router(config=config)

# Verify mesh connectivity
status = router.health_check()
print(f"Active nodes: {status.active_nodes}")
print(f"Consensus available: {status.consensus_ready}")

Single-Query Reasoning

Basic Query Execution

from logos import Router

router = Router(config="mesh_config.yaml")

# Execute single query with adaptive depth
response = router.query(
    prompt="Explain the halting problem and its implications for AI safety",
    depth="adaptive",
    language="en"
)

print(response.text)
print(f"Reasoning depth used: {response.depth}")
print(f"Nodes consulted: {response.nodes_involved}")
print(f"Consensus score: {response.consensus_score}")
print(f"Verification steps: {len(response.audit_trail)}")

Explicit Depth Control

# Force shallow reasoning (faster, less thorough)
quick_response = router.query(
    prompt="What is photosynthesis?",
    depth=2,
    language="en"
)

# Force deep reasoning (slower, more thorough)
deep_response = router.query(
    prompt="Design a distributed consensus algorithm for Byzantine fault tolerance",
    depth=7,
    language="en",
    require_consensus=True
)

Multilingual Reasoning

# Query in Spanish, reason in universal IR, respond in Spanish
response = router.query(
    prompt="¿Cuál es la diferencia entre aprendizaje supervisado y no supervisado?",
    depth="adaptive",
    language="es"
)

# Query in one language, respond in another
response = router.query(
    prompt="Explain quantum entanglement",
    depth=5,
    language="en",
    response_language="zh"  # Respond in Mandarin
)

Multi-Step Workflows

Creating Reasoning Workflows

from logos import Router, Workflow

router = Router(config="mesh_config.yaml")

# Create multi-step workflow
workflow = router.create_workflow(
    name="research-paper-analysis",
    input_file="paper.pdf"
)

# Define workflow steps
workflow.add_step(
    name="extract_claims",
    operation="extract",
    params={"claim_types": ["theorem", "hypothesis", "conclusion"]}
)

workflow.add_step(
    name="verify_claims",
    operation="verify",
    params={"consensus": True, "min_confidence": 0.95}
)

workflow.add_step(
    name="generate_summary",
    operation="synthesize",
    params={"style": "academic", "max_length": 500}
)

# Execute workflow with automatic state management
result = workflow.execute()

print(f"Claims extracted: {len(result.steps['extract_claims'].output)}")
print(f"Claims verified: {result.steps['verify_claims'].verified_count}")
print(f"Summary: {result.steps['generate_summary'].text}")

Workflow with Error Handling

workflow = router.create_workflow("code-review")

try:
    workflow.add_step("analyze_complexity", operation="analyze")
    workflow.add_step("detect_vulnerabilities", operation="security_scan")
    workflow.add_step("suggest_improvements", operation="generate")
    
    result = workflow.execute(
        timeout_seconds=300,
        graceful_degradation=True
    )
    
    # Access individual step results
    for step_name, step_result in result.steps.items():
        print(f"{step_name}: {step_result.status}")
        if step_result.degraded:
            print(f"  Warning: Reduced depth to {step_result.actual_depth}")
            
except workflow.WorkflowError as e:
    print(f"Workflow failed at step: {e.failed_step}")
    print(f"Partial results available: {e.partial_results}")

Strict Write Discipline Protocol

Understanding SWD Flow

The Strict Write Discipline protocol ensures zero-drift reasoning:

1. Proposal: Node generates candidate reasoning step
2. Broadcast: Candidate sent to peer nodes
3. Verification: Peers evaluate consistency and accuracy
4. Consensus: If confidence > threshold, commit
5. Commit: Write to Grail layer
6. Escalation: If consensus fails, escalate to higher-depth node

Monitoring SWD

from logos import Router

router = Router(config="mesh_config.yaml")

# Enable detailed SWD logging
router.configure_logging(
    level="DEBUG",
    swd_trace=True
)

response = router.query(
    prompt="Prove that the set of real numbers is uncountable",
    depth="adaptive"
)

# Inspect SWD trace
for step in response.swd_trace:
    print(f"Step {step.index}:")
    print(f"  Proposal: {step.proposal_summary}")
    print(f"  Verifiers: {step.verifier_nodes}")
    print(f"  Confidence scores: {step.confidence_scores}")
    print(f"  Consensus: {step.consensus_reached}")
    if step.escalated:
        print(f"  Escalated to depth: {step.escalation_depth}")

Custom Consensus Thresholds

# Override consensus threshold for specific query
response = router.query(
    prompt="Design a safety-critical flight control algorithm",
    depth=7,
    consensus_threshold=0.99,  # Higher threshold for safety-critical
    min_peer_verifications=3
)

Semantic Caching

Enable and Configure Caching

from logos import Router, CacheConfig

cache_config = CacheConfig(
    enabled=True,
    ttl_hours=24,
    max_size_mb=2048,
    similarity_threshold=0.85  # Semantic similarity for cache hits
)

router = Router(
    config="mesh_config.yaml",
    cache_config=cache_config
)

# First query (cache miss)
response1 = router.query(
    prompt="What are the benefits of functional programming?",
    depth="adaptive"
)
print(f"Cache hit: {response1.cache_hit}")  # False

# Similar query (cache hit)
response2 = router.query(
    prompt="Explain advantages of functional programming paradigm",
    depth="adaptive"
)
print(f"Cache hit: {response2.cache_hit}")  # True
print(f"Latency reduction: {response2.cache_stats.time_saved_ms}ms")

Cache Management

# Clear cache for specific semantic domain
router.cache.clear_domain("programming_concepts")

# Inspect cache statistics
stats = router.cache.statistics()
print(f"Hit rate: {stats.hit_rate:.2%}")
print(f"Average similarity score: {stats.avg_similarity}")
print(f"Memory usage: {stats.memory_mb}MB")

# Manually warm cache with common queries
common_queries = [
    "What is machine learning?",
    "Explain neural networks",
    "Define supervised learning"
]

for query in common_queries:
    router.query(query, depth=3)

Audit Trails and Reasoning Inspection

Accessing Causal Audit Trails

response = router.query(
    prompt="Should we adopt microservices architecture?",
    depth="adaptive"
)

# Inspect complete reasoning chain
audit_trail = response.audit_trail

for i, step in enumerate(audit_trail.steps):
    print(f"\n--- Step {i+1} ---")
    print(f"Node: {step.node_id}")
    print(f"Reasoning: {step.reasoning_fragment}")
    print(f"Confidence: {step.confidence}")
    print(f"Dependencies: {step.dependencies}")
    print(f"Timestamp: {step.timestamp}")

# Export audit trail for external verification
audit_trail.export_to_file("reasoning_trace.json")

Visualizing Reasoning DAG

from logos.visualization import ReasoningGraph

# Generate visual representation of reasoning flow
graph = ReasoningGraph(response.audit_trail)
graph.render(
    output_file="reasoning_dag.svg",
    format="svg",
    show_confidence_scores=True,
    show_node_assignments=True
)

Advanced Routing Strategies

Custom Node Selection

from logos import Router, NodeSelector

# Define custom node selection strategy
class CapabilityBasedSelector(NodeSelector):
    def select_nodes(self, query, available_nodes):
        """Select nodes based on specific capabilities"""
        if "mathematical" in query.domain:
            return [n for n in available_nodes if n.has_capability("symbolic_math")]
        elif "code" in query.domain:
            return [n for n in available_nodes if n.has_capability("code_analysis")]
        return available_nodes[:2]  # Default: first two nodes

router = Router(
    config="mesh_config.yaml",
    node_selector=CapabilityBasedSelector()
)

Load Balancing

# Configure load-aware routing
router.configure_routing(
    strategy="load_balanced",
    max_node_queue_depth=10,
    rebalance_interval_seconds=30
)

# Monitor node load
for node in router.active_nodes():
    load = router.get_node_load(node.id)
    print(f"Node {node.id}: {load.queue_depth} queries, {load.avg_latency_ms}ms avg")

REST API Integration

Starting REST Server

from logos import Router
from logos.server import RESTServer

router = Router(config="mesh_config.yaml")

# Start REST API server
server = RESTServer(
    router=router,
    host="0.0.0.0",
    port=8080,
    enable_cors=True
)

server.start()

Using REST API

# Query via REST API
curl -X POST http://localhost:8080/query \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Explain the CAP theorem",
    "depth": "adaptive",
    "language": "en"
  }'

# Get mesh status
curl http://localhost:8080/status

# Retrieve audit trail
curl http://localhost:8080/audit/query_id_12345

Python REST Client

import requests

response = requests.post(
    "http://localhost:8080/query",
    json={
        "prompt": "Compare ACID vs BASE database properties",
        "depth": 5,
        "require_consensus": True
    }
)

result = response.json()
print(result["text"])
print(f"Consensus score: {result['consensus_score']}")

WebSocket Streaming

Real-Time Reasoning Updates

from logos import Router
from logos.streaming import WebSocketStreamer

router = Router(config="mesh_config.yaml")

# Start WebSocket server for streaming
streamer = WebSocketStreamer(router, port=8081)
streamer.start()

# Client-side streaming consumer
import asyncio
import websockets

async def consume_reasoning_stream():
    uri = "ws://localhost:8081/stream"
    
    async with websockets.connect(uri) as websocket:
        # Send query
        await websocket.send(json.dumps({
            "prompt": "Design a distributed cache system",
            "depth": 7
        }))
        
        # Receive reasoning steps in real-time
        async for message in websocket:
            step = json.loads(message)
            print(f"Step {step['index']}: {step['reasoning_fragment']}")
            if step.get('final'):
                print(f"Final result: {step['text']}")
                break

asyncio.run(consume_reasoning_stream())

Multi-Node Mesh Setup

Distributed Configuration

# node1_config.yaml (primary coordinator)
mesh:
  name: distributed-mesh
  role: coordinator
  consensus_threshold: 0.95
  
  nodes:
    - type: local
      engine: vllm
      model: opus-mini-4.8
      max_thinking_depth: 7
      host: localhost
      port: 8001
    
    - type: remote
      host: 192.168.1.10
      port: 8002
      capabilities: [symbolic_math, formal_verification]
    
    - type: remote
      host: 192.168.1.11
      port: 8003
      capabilities: [code_analysis, security_audit]

mesh_coordination:
  discovery_protocol: mdns
  heartbeat_interval_seconds: 10
  failure_detection_timeout_seconds: 30

Cross-Mesh Communication

from logos import Router, MeshBridge

# Setup primary mesh
primary_router = Router(config="node1_config.yaml")

# Connect to remote mesh
bridge = MeshBridge(
    local_router=primary_router,
    remote_mesh_url="https://partner-org.example.com/mesh",
    auth_token="${MESH_BRIDGE_TOKEN}"
)

# Query spans both meshes
response = primary_router.query(
    prompt="Analyze cybersecurity implications across our infrastructure",
    depth="adaptive",
    enable_cross_mesh=True
)

print(f"Nodes from local mesh: {response.local_nodes}")
print(f"Nodes from remote mesh: {response.remote_nodes}")

Reasoning Profiles

Loading Custom Profiles

from logos import Router, ReasoningProfile

# Load custom reasoning profile
profile = ReasoningProfile.from_file("profiles/academic_research.json")

router = Router(
    config="mesh_config.yaml",
    default_profile=profile
)

# Override profile for specific query
response = router.query(
    prompt="Analyze this research methodology",
    profile=profile,
    depth="adaptive"
)

Creating Custom Profile

{
  "name": "academic_research",
  "version": "1.0",
  "description": "Profile optimized for academic research tasks",
  "parameters": {
    "verification_strictness": 0.98,
    "citation_required": true,
    "logical_rigor": "formal",
    "preferred_reasoning_style": "deductive",
    "escalation_triggers": [
      "contradiction_detected",
      "insufficient_evidence",
      "formal_proof_required"
    ],
    "output_format": {
      "include_references": true,
      "show_confidence_intervals": true,
      "reasoning_transparency": "full"
    }
  },
  "node_preferences": {
    "prefer_capabilities": ["symbolic_math", "formal_verification"],
    "min_depth": 5
  }
}

Troubleshooting

Consensus Failures

# Diagnose consensus issues
response = router.query(
    prompt="Complex query here",
    depth="adaptive",
    debug_consensus=True
)

if response.consensus_score < 0.95:
    diagnosis = router.diagnose_consensus_failure(response.query_id)
    
    print(f"Consensus score: {response.consensus_score}")
    print(f"Failed verifications: {diagnosis.failed_verifications}")
    print(f"Conflicting nodes: {diagnosis.conflicting_nodes}")
    print(f"Recommended action: {diagnosis.recommendation}")
    
    # Retry with different nodes
    if diagnosis.recommendation == "retry_different_nodes":
        response = router.query(
            prompt="Same query",
            exclude_nodes=diagnosis.conflicting_nodes,
            depth="adaptive"
        )

Performance Optimization

# Profile query performance
with router.performance_profiler() as profiler:
    response = router.query(
        prompt="Query to optimize",
        depth="adaptive"
    )

print(f"Total latency: {profiler.total_ms}ms")
print(f"Routing time: {profiler.routing_ms}ms")
print(f"Inference time: {profiler.inference_ms}ms")
print(f"Consensus time: {profiler.consensus_ms}ms")
print(f"Synthesis time: {profiler.synthesis_ms}ms")

# Optimize based on bottleneck
if profiler.consensus_ms > profiler.inference_ms:
    # Consensus is bottleneck, reduce threshold or peer count
    router.configure_swd(
        consensus_threshold=0.90,
        min_peer_verifications=2
    )

Node Health Monitoring

# Continuous health monitoring
import time

while True:
    health = router.health_check()
    
    for node_id, node_status in health.nodes.items():
        if node_status.status != "healthy":
            print(f"Node {node_id} unhealthy: {node_status.error}")
            
            # Attempt node restart
            router.restart_node(node_id)
        
        if node_status.memory_usage_percent > 90:
            print(f"Node {node_id} high memory: {node_status.memory_usage_percent}%")
            router.clear_node_cache(node_id)
    
    time.sleep(30)

Debug Logging

import logging

# Enable comprehensive debug logging
logging.basicConfig(level=logging.DEBUG)

router = Router(
    config="mesh_config.yaml",
    log_config={
        "level": "DEBUG",
        "log_swd_steps": True,
        "log_node_selection": True,
        "log_cache_operations": True,
        "log_consensus_details": True,
        "output_file": "logos_debug.log"
    }
)

Environment Configuration

Required environment variables:

# Inference engine endpoints (if using remote engines)
export VLLM_ENDPOINT="http://localhost:8001"
export OLLAMA_ENDPOINT="http://localhost:11434"

# Mesh bridge authentication (for cross-mesh communication)
export MESH_BRIDGE_TOKEN="your-secure-token"

# Optional: Performance tuning
export LOGOS_MAX_PARALLEL_THREADS=4
export LOGOS_DEFAULT_TIMEOUT_SECONDS=120
export LOGOS_CACHE_SIZE_MB=2048

# Optional: Monitoring and telemetry
export LOGOS_METRICS_ENABLED=true
export LOGOS_METRICS_PORT=9090

Common Patterns

Batch Processing

queries = [
    "Explain quantum computing",
    "What is blockchain?",
    "Define neural networks"
]

# Process batch with automatic load distribution
results = router.batch_query(
    prompts=queries,
    depth="adaptive",
    parallel=True,
    max_concurrent=3
)

for query, response in zip(queries, results):
    print(f"Q: {query}")
    print(f"A: {response.text}\n")

Fallback Strategy

# Primary query with automatic fallback
try:
    response = router.query(
        prompt="Complex analysis query",
        depth=7,
        timeout_seconds=60
    )
except router.TimeoutError:
    # Fallback to faster, shallower reasoning
    response = router.query(
        prompt="Complex analysis query",
        depth=3,
        require_consensus=False
    )