Ollama Integration + Model Setup

[Omoshirokunai]

What

Connect the side panel to a local Ollama/llamafile server. Implement model selection, streaming responses, and graceful offline fallback.

Model Selection & Download

Auto-Download (Current Sprint 2)

• Add setup modal: "Download qwen2.5 (~400MB)?"
• Download + launch llamafile binary on first-use
• Seamless for users, zero manual steps
• Setup wizard on first-use: "Connect to local AI model"
• Model selector populated from GET /api/tags endpoint

Acceptance Criteria

• Prompt input field at bottom of side panel: placeholder "Ask or reply..."
• Model dropdown get models from GET localhost:11434/api/tags
  • Shows model names (e.g., "qwen2.5:0.5b", "neural-chat:7b")
  • Default selection: first model in list
• Send only on explicit action: Enter key or Send button click
  • No auto-send; no character limit auto-trigger
• POST request to localhost:11434/api/generate
• Response streams into timeline in real-time as ai_suggestion event
• "Use" button: records ai_interaction with acceptance: 'fully_accepted'
• "Dismiss" button: records ai_interaction with acceptance: 'rejected'
• Error handling:
  • Ollama not reachable: show "Local model unavailable. Check Ollama is running."
  • Network error: retry (max 3 attempts)
  • Response timeout (>5s): cancel and show "Response took too long"
• llamafile tested and documented:
  • Same endpoint works (localhost:11434/api/generate)
  • Note any differences (e.g., model loading time)