Perplexity provider and runtime options
Adds Perplexity as a new provider for web search, answers, and research. Runtime options are now forwarded to Claude, Codex, and Gemini, giving callers fine-grained per-request control, and Gemini gains a web_contents tool that extracts page content via URL Context.
🚀 Features
Broaden runtime options to Claude, Codex, and Gemini
The options object that tools accept at call time is now forwarded to the Claude, Codex, and Gemini providers, giving callers fine-grained control over each request without changing the shared tool interface.
Claude supports model, effort, maxTurns, maxThinkingTokens, maxBudgetUsd, and a thinking object. Codex supports model, modelReasoningEffort, and webSearchMode. Gemini supports provider-native options: model plus generation_config for search, model plus config for answer and contents, and caller-supplied research options passed through to the underlying API.
Runtime options override provider defaults from the config, but managed tool inputs and tool wiring stay fixed.
Gemini web contents via URL Context
The Gemini provider now supports the web_contents tool, which extracts the main textual content from one or more public URLs. The implementation uses Gemini's built-in URL Context tool (urlContext), letting the model read and return cleaned page content without an external extraction service.
Retrieval metadata is included in the output: successful fetches are counted in the summary, and failures (paywall, error, unsafe) are surfaced as retrieval issues. A new contentsModel setting controls which model is used for content extraction, defaulting to gemini-2.5-flash.
Perplexity provider for grounded web workflows
pi-web-providers now supports Perplexity as a provider for web_search, web_answer, and web_research. You can enable it from /web-providers, authenticate it with PERPLEXITY_API_KEY, and use Perplexity-specific defaults such as Sonar for grounded answers and sonar-deep-research for longer investigations.
This matters if you want a provider that combines direct search results with grounded model responses without switching to a separate extension.
Provider-specific search options for web_search
web_search now accepts a provider-specific options object, so providers can expose richer search controls without forcing every feature into the generic top-level tool shape. For example, Perplexity-backed searches can now pass values such as country, search_mode, search_domain_filter, and search_recency_filter through the same managed tool.
This keeps pi-web-providers flexible as more providers are added, while preserving a stable shared interface for search queries and result limits.
