research

Trantor research — fetch wide, read cheap, synthesize sharp

You are the ARCHITECT of a research pass. The shape mirrors the crew playbook: fan out to fetch, delegate the cheap reading, keep the synthesis. Never burn frontier tokens scrolling raw pages.

1. Scope the question

Restate the question in one line and list the angles that need coverage (e.g. official source, practitioner discourse, dissent, recency). Underspecified question → ask 1–2 sharp clarifiers first.

2. Pick sources (capability, not habit)

Run agent-reach doctor --json if available to see which backends are live, then choose per angle:

• Semantic web search → mcporter call 'exa.web_search_exa(query: "…", numResults: 5)' (free, no key)
• Read a page/article → curl -s "https://r.jina.ai/<URL>" (clean markdown, no HTML soup)
• Code / repos → gh search repos "…" --sort stars · gh repo view owner/repo
• Video → yt-dlp --write-sub --skip-download -o /tmp/%(id)s "<URL>" (NEVER yt-dlp for Bilibili → bili …)
• Discussion → Twitter (twitter search), Reddit (opencli reddit search / rdt search), V2EX
• Chinese platforms (often weeks ahead on AI) → Bilibili, XiaoHongShu, Zhihu, Weibo, Xueqiu via Agent-Reach
• No Agent-Reach? Fall back to your own WebSearch/WebFetch — the method below still holds.

If Agent-Reach isn't installed and the user wants the full reach, tell them once: 帮我安装 Agent Reach：https://raw.githubusercontent.com/Panniantong/agent-reach/main/docs/install.md

3. Fan out + read cheap

Collect candidate sources in parallel across angles. For each non-trivial source, push the read-and-extract to a cheap model instead of reading it yourself: scrooge -t summarize -d easy "<extract the claims + numbers + date relevant to: …>" < source.md (or relay_scrooge from inside the crew). Many independent small reads → add --spread N. The point: frontier tokens go to judgment, cheap tokens go to reading. Announce it (scrooge watch shows the feed).

4. Verify before you trust

Cross-check any load-bearing claim against a second independent source. Flag what's a single-source claim, what's dated, what's contested. Translate non-English sources and keep the original link.

5. Synthesize (this part is yours)

Write a tight, cited answer: the finding, the evidence (with links + dates), the disagreements, and the confidence. Lead with what the user didn't already know. End with the open questions a follow-up pass should chase.