summarize-source

name	summarize-source
description	Condense a single source — a URL, article, document, or pasted text — into key points the user can read at a glance. Use for "summarize this", "TL;DR", "give me the key points", or as a summarizer sub-agent merging one of several sources. Triggers: summarize, tldr, key points, gist, condense, brief me on this, what's the takeaway.
license	MIT
compatibility	Requires JarvisVR agent-backend. Works on provided text offline; URL fetch uses the configured provider where available.
metadata	{"agent":"research-agent","category":"research","version":"1.0","author":"jarvisvr"}
allowed-tools	web_search show_panel show_document

Summarize Source

Produce a faithful, compact summary of one source and present it as a sectioned panel (or open the original in a document_viewer). This is also the unit of work for a summarizer sub-agent in a fan-out (protocol §9.2 handoff).

Steps

1. Acquire the content. Provided text → use directly. A URL → web_search (or fetch via provider) to obtain the page text.
2. Extract structure: a one-line thesis, 3–5 key points, and any numbers / dates / names worth keeping.
3. Compose the summary. Lead with the thesis; bullet the points; preserve source attribution. Stay faithful — never add claims the source doesn't make.
4. Render a panel with sections, or show_document to open the original alongside.
5. Speak a 1-sentence TL;DR; offer the full panel for detail.

Output

panel (show_panel, props per registry.json):

{ "widget_type": "panel",
  "props": { "title": "Summary — Quest 4 review",
             "body": "Strong display gains, same comfort tradeoffs.",
             "sections": [
               { "heading": "Key points", "text": "1) Brighter pancake lenses 2) Better passthrough 3) Heavier" },
               { "heading": "Source", "text": "TechNews, 2026" } ],
             "scrollable": true } }

Spoken TL;DR:

{ "text": "TL;DR: a meaningful display upgrade, but comfort is unchanged.", "final": true }

As a sub-agent (fan-out)

When merging several sources, each summarizer sub-agent (a1.1, a1.2, …) summarizes one source and returns its {thesis, points, source}; the parent research-agent reconciles overlaps and dedupes before handing to result-synthesis / present-data.

Edge cases

• Paywalled / unreachable URL → summarize the snippet you have and flag that it's partial; never invent the body.
• Very long doc → summarize hierarchically (per-section, then overall) and open it in document_viewer with page_count set.
• Conflicting claims within the source → present them as the source's open question, don't resolve silently.
• Code/tabular content → use code_viewer / data_table instead of prose.