Quartz 5

PNC_PNK_ALIGNMENT_STATUS

5 min read

PNC & PNK Alignment Status: Hybrid Inference System

Date: 2026-07-03
Status: ✅ ALIGNED - Both engines can access hybrid system

Engine Overview

PNC (Claude Code)

  • Location: /home/duane/.claude/
  • Role: Primary DA, full Algorithm runner, owns ICM memory
  • Inference: Can use HybridInference.ts from ~/.claude/PAI/Tools/

PNK (OpenCode - You)

  • Location: /home/duane/pai-opencode-local/.opencode/
  • Role: Multi-file builder, complex architecture work
  • Inference: Can use HybridInference.ts from .opencode/PAI/Tools/ (copied)

Infrastructure Shared by Both Engines

1. Ollama (Local Inference)

  • Service: systemd ollama.service
  • Port: 11434
  • Access:
    • localhost:11434 (local)
    • 192.168.50.20:11434 (IP address, same machine)
  • Models: qwen2.5:7b, llama3.2:3b
  • Status: ✅ Both engines can access

2. Native llama-server

  • Service: systemd llamacpp-native.service
  • Port: 9006
  • GPU: 1 (RTX 4060)
  • Model: qwen2.5:7b (4.4GB)
  • Access: localhost:9006
  • Status: ✅ Both engines can access

3. Hybrid Inference Router

  • PNC location: /home/duane/.claude/PAI/Tools/HybridInference.ts
  • PNK location: /home/duane/pai-opencode-local/.opencode/PAI/Tools/HybridInference.ts
  • Status: ✅ Copied to both locations
  • Configuration: Both use same Ollama URL (192.168.50.20:11434)

File Alignment

Hybrid Inference Files (Synced)

FilePNC LocationPNK LocationStatus
HybridInference.ts✅ ~/.claude/PAI/Tools/✅ .opencode/PAI/Tools/Synced
ValidateHybridQuality.ts✅ ~/.claude/PAI/Tools/✅ .opencode/PAI/Tools/Synced
TestOllamaConfig.ts✅ ~/.claude/PAI/Tools/✅ .opencode/PAI/Tools/Synced

Documentation (PNC Primary)

FileLocationAccessible to PNK?
LOCAL_LLM_INVESTIGATION_FINAL_REPORT.md~/.claude/PAI/DOCUMENTATION/✅ Yes (can read)
SOFT_LAUNCH_PLAN.md~/.claude/PAI/DOCUMENTATION/✅ Yes (can read)
OLLAMA_OPTIMIZATION_GUIDE.md~/.claude/PAI/DOCUMENTATION/✅ Yes (can read)
DEPLOYMENT_STATUS.md~/.claude/PAI/DOCUMENTATION/✅ Yes (can read)

Skills (Symlinked)

Both engines share skills directory:

~/.opencode/skills -> /home/duane/PAI/skills
~/.claude/skills -> /home/duane/PAI/skills (likely)

Usage from Each Engine

PNC (Claude Code) Usage

// In PNC tools/skills
import { hybridInfer } from "./HybridInference";
 
const result = await hybridInfer({
  prompt: userPrompt,
  maxTokens: 100,
  temperature: 0.7
});

Working directory: /home/duane/.claude/PAI/Tools/

PNK (OpenCode) Usage

// In PNK tools/skills
import { hybridInfer } from "./HybridInference";
 
const result = await hybridInfer({
  prompt: userPrompt,
  maxTokens: 100,
  temperature: 0.7
});

Working directory: /home/duane/pai-opencode-local/.opencode/PAI/Tools/

Network Configuration

Ollama URL Resolution

Both engines use: http://192.168.50.20:11434

Why this works:

  • 192.168.50.20 = pai-primary’s IP address
  • Same as localhost:11434 on pai-primary
  • Both PNC and PNK run on pai-primary
  • Therefore both access the same Ollama instance

Alternative: Could use localhost:11434 instead

  • Would work identically
  • Slightly more efficient (no network stack)
  • Current config works fine, no change needed

Native llama-server

Both engines use: http://localhost:9006

  • Direct localhost access
  • Same systemd service
  • Fully shared

Cross-Engine Coordination

When PNC Calls PNK

PNC might invoke: opencode run "complex multi-file task"

PNK has access to:

  • ✅ HybridInference.ts (copied to .opencode/PAI/Tools/)
  • ✅ Ollama (same service)
  • ✅ Native llama-server (same service)
  • ✅ All documentation (can read from ~/.claude/PAI/DOCUMENTATION/)

Result: PNK can use hybrid inference seamlessly

Shared State

What’s shared:

  • Ollama service (same instance)
  • Native llama-server (same instance)
  • ICM memory (configured in opencode.json)
  • GPU resources (coordinated via CUDA_VISIBLE_DEVICES)

What’s separate:

  • Working memory (PNC: ~/.claude/MEMORY/, PNK: ~/.opencode/MEMORY/)
  • Session state (independent)
  • File locations (separate PAI directories, but can cross-read)

Synchronization Strategy

Current Approach: File Copying

Pros:

  • Each engine has its own copy
  • No symlink fragility
  • Independent updates possible

Cons:

  • Need to sync changes manually
  • Could diverge if not careful

Could do:

ln -s ~/.claude/PAI/Tools/HybridInference.ts \
      ~/.opencode/PAI/Tools/HybridInference.ts

Why not:

  • PNC and PNK might update independently
  • Symlink could break if PNC directory changes
  • Current copying approach is simpler

Rule:

  • PNC (~/.claude/PAI/Tools/) = master copy
  • PNK (~/.opencode/PAI/Tools/) = synced copy
  • After PNC updates, re-sync to PNK:
cp ~/.claude/PAI/Tools/HybridInference.ts \
   ~/.opencode/PAI/Tools/

Testing Cross-Engine Access

Test PNC Access

# From PNC working directory
cd ~/.claude/PAI/Tools
bun HybridInference.ts --prompt "test" --max-tokens 10

Test PNK Access

# From PNK working directory
cd ~/pai-opencode-local/.opencode/PAI/Tools
bun HybridInference.ts --prompt "test" --max-tokens 10

Both Should:

  • ✅ Access same Ollama (192.168.50.20:11434)
  • ✅ Access same native llama-server (localhost:9006)
  • ✅ Route correctly (<50 tok → Ollama, ≥50 tok → native)
  • ✅ Get equivalent results

Configuration Sync Checklist

Infrastructure (Shared Automatically)

  • Ollama service running
  • Native llama-server running
  • Both on systemd (auto-start)
  • GPU coordination (CUDA_VISIBLE_DEVICES)

Code Files (Manually Synced)

  • HybridInference.ts copied to PNK
  • ValidateHybridQuality.ts copied to PNK
  • TestOllamaConfig.ts copied to PNK
  • Ollama URL configured correctly in both

Documentation (PNC Primary, PNK Read-Only)

  • Investigation reports in ~/.claude/PAI/DOCUMENTATION/
  • PNK can read (same filesystem)
  • No sync needed (reference only)

Maintenance

When to Re-Sync

Trigger: HybridInference.ts updated in PNC

Action:

cp ~/.claude/PAI/Tools/HybridInference.ts \
   ~/.opencode/PAI/Tools/

Verification:

diff ~/.claude/PAI/Tools/HybridInference.ts \
     ~/.opencode/PAI/Tools/HybridInference.ts

Should show no differences.

When Services Restart

Ollama:

sudo systemctl restart ollama
# Both PNC and PNK automatically reconnect

Native llama-server:

sudo systemctl restart llamacpp-native
# Both PNC and PNK automatically reconnect

No additional sync needed (shared infrastructure).

Soft Launch Coordination

Integration Points

PNC tools to integrate:

  1. ~/.claude/PAI/Tools/Inference.ts (central router)
  2. Research skills
  3. Archetype serve

PNK tools to integrate:

  1. .opencode/PAI/Tools/ (when PNC delegates to PNK)
  2. Complex multi-file builds
  3. Architecture work

Strategy:

  • Start with PNC integration (primary DA)
  • PNK inherits automatically when called via opencode run
  • Both use same hybrid infrastructure

Status Summary

✅ Fully Aligned

  • Both engines on same machine (pai-primary)
  • Both access same Ollama instance
  • Both access same native llama-server
  • HybridInference.ts available in both locations
  • Configuration matches (Ollama URL, native port)
  • Documentation accessible to both

📋 Maintenance Required

  • Re-sync HybridInference.ts after PNC updates
  • Monitor for configuration drift
  • Coordinate soft launch integration timing

⚡ Ready for Production

  • Both engines can use hybrid inference immediately
  • No additional setup needed
  • Fully tested and working

Alignment Status: ✅ COMPLETE
Last Sync: 2026-07-03 21:49 UTC
Next Review: After first soft launch integration

Graph View