Skip to content
thndrs

Web Search

Search Backends

Section titled “Search Backends”

thndrs owns web search and exposes three backend settings through --websearch or the websearch config key:

  • duckduckgo (the default): query DuckDuckGo’s HTML endpoint.
  • searxng: query a configured SearXNG instance using its JSON API.
  • none: disable the web_search tool.

The selected backend is part of runtime and session metadata. Search is never inferred from prompt wording and no provider-specific search header is sent.

SearXNG

Section titled “SearXNG”

SearXNG requires an HTTP(S) base URL. Configure it with --websearch-url, websearch_url, or THNDRS_WEBSEARCH_URL:

Terminal window
thndrs --websearch searxng --websearch-url http://127.0.0.1:8080

The application requests /search?q=...&format=json and normalizes the returned title, URL, and snippet fields. The configured SearXNG host may be loopback or private because it is an explicit local service. SearXNG result URLs remain untrusted and are fetched only through the public URL policy.

Result handling

Section titled “Result handling”

Results are capped at ten per call (five by default), then each selected public URL is fetched through the bounded Lectito-backed extraction path. Results keep their original order; an extraction failure is reported next to that result so later results can still be used. Search and extraction output includes truncation and transport diagnostics when relevant.

Search results should usually be paired with read_url before making claims about page content.

Local Extraction

Section titled “Local Extraction”

Extraction follows the same read-only posture as the repository tools.

HTML and XHTML pages are extracted with Lectito into readable Markdown and plain text. Other allowed text-like content types, such as JSON, XML, plain text, feeds, YAML, CSV, and JavaScript, are returned as raw text. Binary content is rejected.

Public URL Safety

Section titled “Public URL Safety”

Local URL reads only accept public http and https URLs. Private-network targets are rejected before fetching and again after redirects. Redirect count, content type, response size, and total request time are bounded.

This policy also applies to result URLs returned by SearXNG.

Truncation Metadata

Section titled “Truncation Metadata”

Search and extraction results carry truncation metadata so the transcript can show when output was capped.

Search Transcript Entries

Section titled “Search Transcript Entries”

Search activity renders through normal tool transcript entries and remains provider-neutral.