Use nemotron-3-super-120b: confirmed tool calling works without Venice

Exclude Venice provider via OpenRouter extra_body routing

Switch to qwen3-coder:free - verified tool support + 1M context window

Switch to hermes-3-405b for tool calling; cap arxiv to 5 results with 1 retry

Fix context overflow: truncate content to 3k chars, reduce results; switch to nemotron-120b

Fix verbose log capture: strip ANSI instead of dropping, add logging handler

Use moonshotai/kimi-k2.6:free (verified available on OpenRouter)

Add litellm dep; use openrouter/ prefix with qwen3-8b:free

Switch to qwen/qwen3-14b:free to avoid CrewAI google/ provider detection

Use base_url approach with gemma-3-27b-it:free; no LiteLLM prefix stripping

Add OpenRouter model fallback chain (5 free models) to survive rate limits

Switch to deepseek-chat-v3 free model (avoid Venice provider spend limit)

Remove google-generativeai packages to stop CrewAI auto-selecting Gemini default

Fix OpenRouter: remove base_url, let LiteLLM route via openrouter/ prefix

Fix OpenRouter model slug to free llama-3.3-70b

Switch LLM to OpenRouter (gemini-2.0-flash); drop direct Gemini dependency

Fix: switch to Docker SDK, lazy embedder loading, bug fixes

Initial commit