rdmsm4x-changelog-20260830-0035-fleet-local-llm-worker-routing

Summary

Configured universal local LLM worker routing across the fleet message bus (agent_msg.zsh), establishing dedicated asynchronous worker endpoints: mlx@<host> (Apple Silicon Metal), ollama@<host> (GGUF daemon), and tosh@<host> (Intel Mac AMD dGPU).

Details of Changes

1. Fleet LLM Routing Architecture

Configured three distinct local inference worker daemons:

• mlx@<host> (Apple Silicon): Powered by mlx-agent-worker and Apple MLX (mlx-lm). Executes on Metal GPU with zero-copy unified memory.
• ollama@<host> (General Purpose): Powered by ollama-agent-worker connecting to local Ollama daemon on port 11434 (glm-4.7-flash, qwen3.6, gpt-oss).
• tosh@<host> (Intel Workstations): Powered by tosh-agent-worker for Intel Macs with discrete AMD GPUs (rdmpw3275m, rdmpw3265m).

2. Binaries & LaunchAgents Created

• ~/dev/apps/apple-mlx/bin/mlx-agent-worker -> ~/bin/mlx-agent-worker
• ~/dev/fleet/tooling/bin/ollama-agent-worker -> ~/bin/ollama-agent-worker
• ~/dev/fleet/tooling/bin/tosh-agent-worker -> ~/bin/tosh-agent-worker
• ~/dev/fleet/tooling/launchd/com.eastcoastscience.ollama-agent-worker.plist
• ~/dev/fleet/tooling/launchd/com.eastcoastscience.tosh-agent-worker.plist
• ~/dev/apps/apple-mlx/launchd/com.eastcoastscience.mlx-agent-worker.plist

3. Primary Agent Posture

Claude Code, Codex, and Antigravity retain their normal high-reasoning primary models. They can asynchronously delegate or prompt local engines via:

• Fleet bus: agent_msg.zsh send --to <worker>@<host>
• MCP tools: mlx_generate, mlx_chat in ~/.claude/settings.json and ~/.gemini/antigravity-cli/mcp/apple-mlx/

4. Verification Evidence

• Live round-trip bus tests passed:
  • mlx@rdmsm4x: message 20260830-003234-570A6B03 -> reply 20260830-003258-E4A0E260.
  • ollama@rdmsm4x: message 20260830-003327-DB5C84BF -> reply 20260830-003413-239223AC.
• Broadcast 20260830-003432-366A1815 dispatched to all fleet nodes.