feat: Add Ollama environment variable support and fix compatibility

[CodeNKasch]

Overview

This PR enhances Ollama integration with environment variable support and fixes compatibility issues, enabling seamless use of local LLMs without requiring Gemini API keys.

Key Features

🔧 Environment Variables

• GIA_DEFAULT_MODEL: Set your preferred model (Ollama or Gemini)

  • Example: export GIA_DEFAULT_MODEL="ollama::llama3.2"
  • When set to an Ollama model, no Gemini API key is needed
  • Priority: -m flag > GIA_DEFAULT_MODEL > gemini-2.5-flash-lite

• OLLAMA_BASE_URL: Connect to remote Ollama servers

  • Example: export OLLAMA_BASE_URL="http://192.168.1.100:11434"
  • Default: http://localhost:11434
  • Automatic URL normalization to /v1/ endpoint

🐛 Ollama Compatibility Fix

Fixed critical issue where Ollama's OpenAI compatibility layer couldn't handle multiple text parts in a single message:

• Problem: Messages with multiple parts (role + file + prompt) would fail
• Solution: Automatically merge multiple text-only parts into single text for Ollama
• Benefit: Preserves Gemini's prompt caching while fixing Ollama compatibility

📚 Documentation

Updated README.md and CLAUDE.md with:

• Quick setup examples for Ollama without API keys
• Environment variable configuration
• Model selection priority
• Custom server configuration

🧪 Testing

• Added comprehensive unit tests for URL normalization
• Added tests for multi-part message merging
• Added tests for message content preservation (Gemini caching)
• Manual testing confirms all features working correctly

Usage Examples

# Set default model to Ollama (no Gemini API key needed)
export GIA_DEFAULT_MODEL="ollama::llama3.2"
gia "What is the capital of France?"

# Connect to remote Ollama server
export OLLAMA_BASE_URL="http://192.168.1.100:11434"
gia -m "ollama::llama3.2" "your prompt"

# Use with file inputs (multi-part merging works automatically)
gia -m "ollama::llama3.2" -f data.txt "Summarize this"

Technical Details

• Renamed OLLAMA_MODEL to GIA_DEFAULT_MODEL for consistency
• Removed unused created_at field from ConversationSummary
• Removed unused notify_rust import
• URL normalization ensures proper /v1/ endpoint format
• Multi-part message merging preserves multimodal content (images, audio)