---
name: phaseo-web-search-tools
description: "Configures and debugs web grounding on Phaseo: the `gateway:web_search` and `gateway:web_fetch` tools, native provider web search, citations and request-log checks for search-heavy flows. Use when responses need current web information or citations, or search tool calls fail or return nothing. Not for calling your own function tools (use an SDK skill) or building agent loops (phaseo-agent-sdk-typescript)."
license: MIT
---

# Phaseo Web Search Tools

Use this skill when the work is specifically about search, source grounding, or fetched-page context rather than generic gateway wiring.

## Workflow
1. Identify the search mode:
- native provider web search
- `gateway:web_search`
- `gateway:web_fetch`
- mixed search plus follow-up fetch grounding

2. Choose the lowest-complexity path:
- prefer native provider web search when the chosen provider already supports it and the request surface can pass it through
- use `gateway:web_search` when you need provider-agnostic search behavior
- add `gateway:web_fetch` only when one or more concrete URLs must be grounded into page text

3. Keep the request contract explicit:
- native provider search should still route through `web_search_options` capability support
- `gateway:web_search` should declare bounded result count and only include full text when needed
- `gateway:web_fetch` should keep `max_chars` tight to control latency and token volume

4. Validate through logs before finishing:
- confirm request detail views show web-search observability
- confirm native search calls, managed search results, citations, or fetch metadata appear where expected
- verify usage counters such as `server_tool_use.web_search_requests` and `server_tool_use.web_fetch_requests`

## Practical rules
- Treat search and fetch as separate cost/latency decisions.
- Prefer highlights/snippets over full page text until the workflow proves it needs more.
- When debugging provider-native search, inspect both the request surface schema and capability filtering before changing model/provider choices.
- When grounding fetched content, check truncation flags before assuming the model saw the full page.

## Use this skill for
- coding-agent implementations that need search-backed answers
- request-log debugging for missing citations or search results
- comparing native provider search against gateway-managed search
- grounding workflows that chain search into fetched-page extraction