feat: Local LLM support via Ollama for plan execution

[mmorris35]

Summary

Add support for using local LLMs (via Ollama) as an alternative to Claude Haiku for executing development plans. This enables:

• Zero API costs after hardware investment
• Privacy — code never leaves the machine
• Offline execution — no internet required

Motivation

Models like Qwen3-Coder-Next-80B now rival Claude on coding benchmarks and can run locally on Apple Silicon Macs with 64GB+ unified memory. For teams with suitable hardware, this eliminates per-token costs entirely.

Proposed Implementation

1. Ollama-compatible executor agent

• Use Ollama's OpenAI-compatible API (localhost:11434/v1/chat/completions)
• New executor template that works with local models
• Configurable model selection (qwen3-coder-next, codellama, deepseek-coder, etc.)

2. Configuration options

{
  "executor": {
    "provider": "ollama",
    "model": "qwen3-coder-next",
    "baseUrl": "http://localhost:11434",
    "contextWindow": 128000
  }
}

3. Prompt tuning

• May need DevPlan-specific system prompts optimized for open models
• Test and document which models work best with DevPlan format

Hardware Requirements

Model	Min RAM	Speed (M4 Pro)
Qwen3-Coder-Next-80B (Q4)	64GB	~10-15 tok/s
DeepSeek-Coder-33B (Q4)	24GB	~25-30 tok/s
CodeLlama-34B (Q4)	24GB	~25-30 tok/s

Success Criteria

• Executor agent can use Ollama as backend
• At least one model (Qwen3-Coder-Next) tested end-to-end with a sample plan
• Documentation for local LLM setup
• Performance comparison vs Haiku (speed, quality)

Related

• Ollama OpenAI compatibility: https://ollama.com/blog/openai-compatibility
• Qwen3-Coder-Next: https://ollama.com/library/qwen3-coder-next