feat: add Tavily as configurable search and crawl provider

[tavily-integrations]

Summary

Adds Tavily as an optional, parallel search and content extraction provider alongside the existing Serper.dev (search) and Jina (crawl) tools. No defaults are changed — operators can switch by setting search_tool: tavily and/or crawl_tool: tavily in their YAML config.

What changed

• tools/search.py: Added TavilySearchTool (subclass of SearchToolBase, name=tavily) using AsyncTavilyClient.search(), and TavilyWebCrawlTool (subclass of WebCrawlToolBase, name=tavily) using AsyncTavilyClient.extract(). Both map results to existing OrganicResults/SerpEntry/WebContentV2 dataclasses with proper provenance metadata.
• tools/types.py: Added TAVILY = "tavily" to both SerpProvider and WebProvider enums.
• src/agents/base.py: Imported and registered TavilySearchTool and TavilyWebCrawlTool in TOOL_CLASS_SEARCH / TOOL_CLASS_CRAWL lists.
• config/redsearcher.yaml and config/deepseek.yaml: Added inline comments documenting tavily as an alternative value for search_tool and crawl_tool.
• .env: Added TAVILY_API_KEY="" placeholder.
• requirements.txt: Added tavily-python>=0.5.0.

Files changed

• tools/search.py
• tools/types.py
• src/agents/base.py
• config/redsearcher.yaml
• config/deepseek.yaml
• .env
• requirements.txt

Dependency changes

• Added tavily-python>=0.5.0 to requirements.txt

Environment variable changes

• Added TAVILY_API_KEY (read by TavilySearchTool and TavilyWebCrawlTool via AsyncTavilyClient)

Notes for reviewers

• This is an additive migration — all existing Serper.dev and Jina code is untouched
• Tavily client is lazily imported inside methods to avoid import errors when the package isn't installed
• Both tools use the existing retry, caching, and registry infrastructure
• To activate: set search_tool: tavily and/or crawl_tool: tavily in your config YAML and provide TAVILY_API_KEY

🤖 Generated with Claude Code

Automated Review

• Passed after 1 attempt(s)
• Final review: The additive Tavily migration is correct and well-scoped. Both TavilySearchTool and TavilyWebCrawlTool conform to the SearchToolBase/WebCrawlToolBase interfaces, use the correct Tavily SDK methods (AsyncTavilyClient.search() and .extract()), properly map response fields to the existing SerpEntry/WebContentV2 types, and follow existing patterns for caching and retry logic. Enum values, registrations, dependency, env placeholder, and config comments are all present. Existing Serper/Jina tooling is fully intact. Three minor issues noted below, none blocking.