Tool reference

Every tool and parameter, so you can audit the surface area.

web_search

Parameter	Type	Notes
query	string	required
max_results	number	1–20, default 8
recency	string	d day · w week · m month · y year
engine	string	ddg (default) · brave · bing · multi (parallel, RRF-fused)
deep	boolean | number	read top results, attach query-relevant excerpts. true = 4, or 1–8
refresh	boolean	skip the 10-minute cache

Rows include the URL, publication date when the engine provides one, and a via line showing which engines served the results. Deep rows show an excerpt: line per result.

Fallback chain: DDG html → DDG lite → DDG POST → r.jina.ai proxy → Bing RSS. Any primary-engine failure auto-retries through multi.

fetch_page

Parameter	Type	Notes
url	string	http/https only (SSRF-guarded)
query	string	returns intro + most query-relevant passages (BM25) instead of the page head
max_chars	number	default 8000, max 50000
raw	boolean	return raw HTML instead of extracted markdown
timeout	number	1000–60000 ms
headers	object	custom request headers
no_cache / no_wayback / allow_http_errors	boolean	escape hatches

Handles HTML (markdown extraction), JSON (pretty-printed), plain text and PDFs (poppler if installed, internal parser otherwise). Bot-walls retry via the Wayback Machine; thin/SPA pages re-render through a reader proxy. The output header reports the source: direct, wayback, jina, github-issue-api, stackexchange-api, wikipedia-rest, …

Site adapters

Known URL shapes skip scraping entirely:

• github.com/{o}/{r} — API metadata + raw README
• github.com/{o}/{r}/issues|pull/{n} — issue/PR + comments as markdown
• github.com/{o}/{r}/blob/{ref}/{path} — raw file
• stackoverflow.com/questions/{id} — question + top answers
• news.ycombinator.com/item?id= — thread via Algolia
• reddit.com/.../comments/... — post + top comments via .json
• wikipedia.org/wiki/{title} — REST html → markdown

search_stackoverflow / search_wikipedia / search_npm / search_github / search_hn

Single query (+ max) each. All free APIs, no keys; search_github honours an optional GITHUB_TOKEN env var for rate limits.

/research <topic>

Runs web + HN + GitHub + Wikipedia searches in parallel with live footer progress, fetches the most promising pages, then asks the model for a grouped briefing: documents worth reading, recurring conclusions, and disagreements between sources.