Adds Tavily Search as a native Router path. Returns web search results wrapped in an OpenAI-compatible ChatCompletion shim — Thompson Sampling can compete Tavily against LLMs on web_scraping and research goals.

Changes

• Added tavily/ vendor prefix to _VENDOR_PREFIX_MAP
• Added _make_chat_completion() generic shim helper for non-LLM providers
• Added _call_tavily() — calls Tavily API, wraps result as ChatCompletion
• Env var: TAVILY_API_KEY
• Tests: 4 new test cases (dispatch, response shape, missing API key)

Usage

from kalibr import Router

router = Router(
    goal="web_scraping",
    paths=["tavily/advanced", "deepseek-chat"],
    success_when=lambda out: len(out) > 50
)
response = router.completion(messages=[{"role": "user", "content": "latest AI benchmarks"}])

Supports tavily/basic and tavily/advanced search depths.

Tavily is a hackathon co-sponsor.