cap_web_search

Web Search

Use this skill when the user needs current public web information that should be fetched through the built-in search capability instead of answered from memory.

When to use

• The user asks for current news, recent developments, live information, or up-to-date facts.
• The user asks to search the web for a topic, company, product, documentation page, or article.
• The task needs a few likely sources or result links before summarizing.
• The answer would be risky if guessed from stale model knowledge.

Available capability

• web_search: search the web with the configured provider and return concise formatted results.

Provider behavior

• The runtime prefers Tavily when a Tavily API key is configured.
• If Tavily is not configured but Brave Search is configured, it falls back to Brave.
• If neither provider key is configured, the capability returns an error instead of search results.

Calling rules

• Call web_search directly. Do not route web search through CLI wrappers unless the user explicitly asks for console commands.
• Input must be a JSON object with one required field:

{
  "query": "your search query"
}

• Keep queries short and concrete. Prefer one clear search intent per call.
• If the user asks a compound question, split it into separate searches when needed instead of cramming everything into one long query.
• Use search when freshness matters. Do not use it for stable facts that can be answered locally.

Output shape

• The capability returns plain text, not structured JSON.
• Results are formatted as a short numbered list.
• Each item typically contains:
  • title
  • URL
  • short snippet/content
• If no result is found, the output is No web results found.
• Common error strings include:
  • Error: no search provider credentials configured
  • Error: invalid input JSON
  • Error: missing query
  • Error: search request failed (...)
  • Error: failed to parse search results

Recommended workflow

1. Decide whether the user needs fresh web information or a normal answer.
2. Write one focused query.
3. Call web_search.
4. Read the returned snippets and URLs.
5. Summarize the findings for the user and mention uncertainty when the results are weak or mixed.