Summary

• Added Tavily as a configurable, highest-priority search backend alongside existing Exa, Brave, and DuckDuckGo backends
• The priority chain is now: TAVILY_API_KEY → EXA_API_KEY → BRAVE_API_KEY → DuckDuckGo
• Existing backends are fully preserved and unchanged

Files changed

• internal/search/tavily.go (new) — TavilySearcher implementation using net/http POST to https://api.tavily.com/search, mapping results[].{title, url, content} to SearchResult
• internal/search/search.go — Added TAVILY_API_KEY check as first priority in resolveSearcher() with empty-key guard
• internal/search/search_test.go — Added tests for Tavily searcher (success, HTTP error), resolver priority (Tavily over Exa), empty key error; updated existing tests to unset TAVILY_API_KEY

Dependency changes

• None — uses net/http from the Go stdlib

Environment variable changes

• Added TAVILY_API_KEY — set to a Tavily API key to enable the Tavily backend (takes priority over all other backends when set)

Notes for reviewers

• No new Go module dependencies required
• Follows the exact same patterns as ExaSearcher and BraveSearcher (struct with apiKey/baseURL/client, test server override via unexported fields)
• Go toolchain was not available in the build environment so tests could not be run locally; CI should validate

Automated Review

• Passed after 1 attempt(s)
• Final review: The Tavily migration is well-implemented and correct. tavily.go follows the existing exa.go pattern closely: same struct layout, HTTP client setup, error wrapping style, and //nolint:errcheck usage. The Tavily REST API is called correctly — POST to /search with api_key in the JSON body, which matches Tavily's documented API contract. The resolveSearcher() priority insertion is clean and consistent. Tests are comprehensive, cover the happy path, HTTP errors, and resolver priority; existing tests are correctly guarded with unsetEnv(t, "TAVILY_API_KEY") to prevent interference. One minor test weakness noted but not blocking.