octen-research

Octen Research

Multi-source research on Octen. The shape is always the same: fan out, filter, ground, then synthesize — never a single search dressed up as a report.

When this skill applies

Use it when the answer needs coverage: several entities, several angles, or a claim that must hold across sources.

Do not use it for a single fact, a single entity, or a page you already have the URL for — those are one search or one extract. Routing for those lives in the octen-web-search skill.

Before searching: resolve time

A relative window ("recent", "last quarter", "past 6 months") is not a filter until it becomes two timestamps. Take today's date from your environment context, work out the bounds, and state them before the first call — an approximate window silently changes what the fan-out covers.

Pass them as start_time / end_time, or use time_range when one of the coarse windows (day / week / month / year) is exactly what was asked for.

Step 1 — fan out

One broad_search with the whole question. Do not pre-split it into sub-queries; decomposition is what the tool does.

Set max_queries to the scope:

Scope	max_queries
Focused comparison, 2–3 entities	3–5
Multi-facet research	5–10
Landscape scan	10–20
Exhaustive survey	20–30

For a question about recent events, reach for news_search instead. Narrow with include_domains when the credible sources are known (arxiv.org, a vendor's docs domain), and exclude_domains to drop content farms.

Results come back grouped per sub-query and are not deduplicated — the same URL can appear under several sub-queries. Dedupe by URL before counting anything.

Step 2 — filter before you read

Judge candidates from the highlights, not by fetching everything. Drop duplicates, listicles, and pages that only restate the query. extract is the expensive step; spend it on the handful that carry real evidence.

Step 3 — ground

extract the survivors, up to 20 URLs in one call. Leave query unset — this step exists to read the page, and setting query returns ranked excerpts instead of the body, which is the opposite of grounding. Each result carries category and page_structure — use them to throw out login walls and nav pages before spending context on what you fetched. Retry failed or empty extracts with mode: "advanced".