Research a question across three independent sites

README.md

@@ -22,6 +22,7 @@ Browser tab
 ├── Answer check ────► every reply, read back against the tool results
 └── Tool loop ───────► search provider  (DuckDuckGo, Wikipedia, or LangSearch/Jina with your key)
                     ├► r.jina.ai        (page reader)
+                    ├► MediaWiki        (Wikipedia pages, no reader budget)
                     └► MCP servers over HTTP
 ```
 
@@ -205,6 +206,7 @@ Three details make the app work from a repository sub-path rather than a domain
 | -------------- | ------------------------------------------------------------------- |
 | `web_search`   | Full web search with no key; Wikipedia, LangSearch or Jina instead. |
 | `read_page`    | Fetches a URL and returns its readable text.                        |
+| `research`     | Search, read three independent sites, return quoted passages.       |
 | `calculator`   | Exact arithmetic via a hand-written parser.                         |
 | `current_time` | Local date, time, and timezone.                                     |
 | `weather`      | Current conditions and a three-day outlook, from several forecasts. |
@@ -233,7 +235,7 @@ Search and `read_page` share the reader's budget of 20 requests a minute per IP,
 
 **The search itself now carries the facts a 0.8B model would otherwise spend a round guessing at.** Every `web_search` result is stamped with today's local date, so "current" and "today's news" have a date without calling `current_time`. A German question searches German Wikipedia and, on DuckDuckGo, prefers German results (`kl=de-de`); English _who was Ada Lovelace_ is not mistaken for German because bare `was` is also English. German Wikipedia is smaller, so an empty result there falls through to English rather than telling the model the subject does not exist.
 
-The `research-question` skill used to teach only "search, then answer from the snippet". That is how Wikipedia's lead about an office never naming the incumbent became a wrong answer. It now also shows opening the page when the snippet is not enough, and answering a German office-holder question in German from a German source.
+The `research-question` skill does not offer `web_search`. It offers `research`, which searches, picks three **different sites** (`investor.nvidia.com` and `nvidianews.nvidia.com` count as one), reads them in parallel, and returns the passages that bear on the question. Wikipedia pages go through MediaWiki, so a typical call spends one reader request on the search and two on the other sites rather than six. A page that will not open becomes its search snippet; if nothing readable comes back the tool throws rather than telling the model the subject does not exist. `lookup-term` still searches and optionally reads one page — a name does not need three sources.
 
 **LangSearch is the way off that shared budget without paying for one.** `api.langsearch.com` is a search API rather than a results page, its free tier allows 1,000 searches a day and one a second, and a key needs no card — so a search stops competing with `read_page` for the same 20 requests a minute. Two things about it are worth knowing before choosing it. Its snippets are index text rather than prose, lower-cased and with spaces around the punctuation, which a 0.8B model reads less confidently than a sentence. And it answers in an envelope: a refusal it decides to report with a 200 arrives as a `msg` and no result set, so `searchLangSearch` raises that rather than passing an empty list to a model that would relay it as "this does not exist". Long summaries are available per result and are switched off — each is the whole page behind the result, which would leave a 0.8B context with no room for the answer.
 

src/eval/scenarios.ts

@@ -59,7 +59,7 @@ export interface Scenario {
 }
 
 function searchQuery(calls: Invocation[]): string | null {
-  const search = calls.find((call) => call.name === 'web_search')
+  const search = calls.find((call) => call.name === 'web_search' || call.name === 'research')
   return search ? String(search.arguments.query ?? '') : null
 }
 
@@ -261,11 +261,10 @@ export const SCENARIOS: Scenario[] = [
     id: 'web-current-event',
     category: 'web',
     prompt: 'Who is the current secretary-general of the United Nations?',
-    expectTool: 'web_search',
+    expectTool: 'research',
     accept: matches(/guterres/i),
-    // The hardest of these under Wikipedia, whose lead extract describes the
-    // office and never names the incumbent: passing there needs a follow-up
-    // `read_page`. A web provider names him in the snippets.
+    // Wikipedia's lead about the office often never names the incumbent; this
+    // tool reads three pages, so the person page or a UN page can supply it.
     online: true,
   },
   {
@@ -344,7 +343,7 @@ export const SCENARIOS: Scenario[] = [
     // A price is looked up, never worked out. `how much is` was an arithmetic
     // keyword, so this reached the calculator with nothing to calculate.
     prompt: 'How much is a Big Mac in Japan?',
-    expectTool: 'web_search',
+    expectTool: 'research',
     accept: (answer) => /\d/.test(answer),
     online: true,
   },
@@ -353,7 +352,7 @@ export const SCENARIOS: Scenario[] = [
     category: 'web',
     // `today` was a current-date trigger, which answered this with the date.
     prompt: "What's today's news?",
-    expectTool: 'web_search',
+    expectTool: 'research',
     accept: (answer) => answer.trim().length > 20,
     online: true,
   },
@@ -364,7 +363,7 @@ export const SCENARIOS: Scenario[] = [
     // answer in English. The query has to keep the German word; translating it
     // to "chancellor of germany" is the 1inch failure in another language.
     prompt: 'Wer ist der Bundeskanzler?',
-    expectTool: 'web_search',
+    expectTool: 'research',
     acceptCall: (calls) => /bundeskanzler/i.test(searchQuery(calls) ?? ''),
     accept: matches(/merz|scholz|kanzler/i),
     online: true,

src/lib/tool-labels.test.ts

@@ -8,8 +8,8 @@ describe('describeTool', () => {
   })
 
   it('speaks in the past tense once it is over, however it ended', () => {
-    expect(describeTool('read_page', 'done')).toBe('Read a page')
-    expect(describeTool('read_page', 'error')).toBe('Read a page')
+    expect(describeTool('research', 'running')).toBe('Researching')
+    expect(describeTool('research', 'done')).toBe('Researched')
   })
 
   it('leaves a tool it does not ship under its own name', () => {

src/lib/tool-labels.ts

@@ -14,6 +14,7 @@ import type { ToolCall } from '@/types'
 const PHRASES: Record<string, { running: string; done: string }> = {
   web_search: { running: 'Searching the web', done: 'Searched the web' },
   read_page: { running: 'Reading a page', done: 'Read a page' },
+  research: { running: 'Researching', done: 'Researched' },
   calculator: { running: 'Calculating', done: 'Calculated' },
   current_time: { running: 'Checking the date', done: 'Checked the date' },
   weather: { running: 'Checking the weather', done: 'Checked the weather' },

src/skills/library.test.ts

@@ -33,7 +33,7 @@ describe('the shipped library', () => {
       ['current-date', 25, ['current_time']],
       ['summarize-url', 20, ['read_page']],
       ['lookup-term', 15, ['web_search', 'read_page']],
-      ['research-question', 10, ['web_search', 'read_page']],
+      ['research-question', 10, ['research']],
     ])
   })
 
@@ -427,7 +427,7 @@ describe('activating each shipped skill', () => {
     ['What year is it?', 'current-date', ['current_time']],
     ['What does https://example.com/pricing say?', 'summarize-url', ['read_page']],
     ['What is Stripe?', 'lookup-term', ['web_search', 'read_page']],
-    ['Who is the current secretary-general of the UN?', 'research-question', ['web_search', 'read_page']],
+    ['Who is the current secretary-general of the UN?', 'research-question', ['research']],
   ]
 
   it.each(cases)('materialises %s for %j with only its tools', (message, name, tools) => {

src/skills/research-question/SKILL.md

@@ -1,11 +1,10 @@
 ---
 name: research-question
-description: Answers a question about current or verifiable facts by searching the web, opening the most promising result and citing it. Use for anything recent, anything about a named person or organisation, and anything you would otherwise be guessing at.
+description: Answers a question about current or verifiable facts by researching it across several independent sources and citing them. Use for anything recent, anything about a named person or organisation, and anything you would otherwise be guessing at.
 jarvis:
   priority: 10
   tools:
-    - web_search
-    - read_page
+    - research
   keywords:
     - look it up
     - find out
@@ -53,61 +52,38 @@ jarvis:
     # thermometer must not be answered with a search.
     - '^\s*aktuelle[rs]?\s+(?!(uhrzeit|zeit|datum|temperatur|wetter|jahr|tag|woche|monat)\b)[a-zäöüß]{3,}\s*\??\s*$'
   exemplars:
-    # The failure this exists to catch: Wikipedia's lead names the office and
-    # not the person, so answering from the snippet invents or omits the name.
     - user: Who is the current secretary-general of the UN?
       steps:
-        - tool: web_search
+        - tool: research
           arguments:
             query: current United Nations secretary-general
           result: |
-            Searched 2026-08-26 for "current United Nations secretary-general".
+            Researched 2026-08-26 for "current United Nations secretary-general" across 3 sources, all read in full.
 
-            1. Secretary-General of the United Nations - Wikipedia
-               https://en.wikipedia.org/wiki/Secretary-General_of_the_United_Nations
-               The secretary-general is the chief administrative officer of the United Nations.
-        - tool: read_page
-          arguments:
-            url: https://en.wikipedia.org/wiki/Secretary-General_of_the_United_Nations
-          result: |
-            # Secretary-General of the United Nations
-            Source: https://en.wikipedia.org/wiki/Secretary-General_of_the_United_Nations
-
-            António Guterres of Portugal has served as secretary-general since 2017.
+            1. António Guterres - Wikipedia — https://en.wikipedia.org/wiki/António_Guterres
+               "António Guterres has served as secretary-general of the United Nations since 2017."
+            2. Secretary-General | United Nations — https://www.un.org/sg/en
+               "António Guterres is the ninth Secretary-General of the United Nations."
       answer: |
         António Guterres, since 2017.
 
-        Source: https://en.wikipedia.org/wiki/Secretary-General_of_the_United_Nations
-    - user: Who is the chief executive of Fictional Airways?
-      steps:
-        - tool: web_search
-          arguments:
-            query: Fictional Airways chief executive
-          result: |
-            Searched 2026-08-26 for "Fictional Airways chief executive".
-
-            1. Leadership — Fictional Airways
-               https://fictionalairways.example/leadership
-               Ama Osei has led Fictional Airways as chief executive since 2023.
-      answer: |
-        Ama Osei, chief executive since 2023.
-
-        Source: https://fictionalairways.example/leadership
+        Source: https://en.wikipedia.org/wiki/António_Guterres https://www.un.org/sg/en
     - user: Wer ist der Bundeskanzler?
       steps:
-        - tool: web_search
+        - tool: research
           arguments:
             query: Bundeskanzler
           result: |
-            Searched 2026-08-26 for "Bundeskanzler".
+            Researched 2026-08-26 for "Bundeskanzler" across 3 sources, all read in full.
 
-            1. Bundeskanzler (Deutschland) – Wikipedia
-               https://de.wikipedia.org/wiki/Bundeskanzler_(Deutschland)
-               Friedrich Merz ist seit Mai 2025 Bundeskanzler der Bundesrepublik Deutschland.
+            1. Bundeskanzler (Deutschland) – Wikipedia — https://de.wikipedia.org/wiki/Bundeskanzler_(Deutschland)
+               "Friedrich Merz ist seit dem 6. Mai 2025 Bundeskanzler der Bundesrepublik Deutschland."
+            2. Bundeskanzler.de — https://www.bundeskanzler.de
+               "Friedrich Merz führt die Bundesregierung."
       answer: |
         Friedrich Merz, seit Mai 2025.
 
-        Source: https://de.wikipedia.org/wiki/Bundeskanzler_(Deutschland)
+        Source: https://de.wikipedia.org/wiki/Bundeskanzler_(Deutschland) https://www.bundeskanzler.de
 ---
 
-Search first, then answer from the results. Open a result with `read_page` when the snippet does not name the answer. Answer in the language you were asked. Always end with the source URL.
+Call `research` once. Answer from the quoted passages, in the language you were asked. Cite more than one source URL when several came back.

src/tools/builtins.test.ts

@@ -71,6 +71,16 @@ describe('web_search', () => {
     expect(description('langsearch')).toMatch(/Search the web/)
   })
 
+  it('refuses an empty research query without asking the network', async () => {
+    const fetchMock = vi.fn()
+    vi.stubGlobal('fetch', fetchMock)
+    const tool = createBuiltinTools(DEFAULT_WEB_ACCESS).find(
+      (candidate) => candidate.schema.function.name === 'research',
+    )!
+    await expect(tool.execute({ query: '  ' })).rejects.toThrow('query must not be empty')
+    expect(fetchMock).not.toHaveBeenCalled()
+  })
+
   it('stamps today on the results so a current-events answer has a date', async () => {
     vi.useFakeTimers()
     vi.setSystemTime(new Date('2026-08-26T15:00:00'))

src/tools/builtins.ts

@@ -1,5 +1,6 @@
 import { evaluateExpression } from './calculator'
 import { memory } from './memory'
+import { researchQuestion } from './research'
 import { defineTool, type Tool } from './types'
 import { DEFAULT_WEB_ACCESS, readPage, searchWeb, type SearchProvider, type WebAccessConfig } from './web'
 import { weatherReport } from './weather'
@@ -136,6 +137,23 @@ export const currentTime = defineTool(
   },
 )
 
+function createResearch(config: WebAccessConfig): Tool {
+  return defineTool(
+    'research',
+    'Search the web, read the three most independent results and return quoted passages from each. Use for current events, people, organisations, or anything you would otherwise be guessing at.',
+    {
+      type: 'object',
+      properties: { query: { type: 'string', description: 'The question to research' } },
+      required: ['query'],
+    },
+    async (args) => {
+      const query = String(args.query ?? '').trim()
+      if (!query) throw new Error('query must not be empty')
+      return researchQuestion(query, config)
+    },
+  )
+}
+
 /**
  * The network tools close over the current provider settings, so they are
  * rebuilt when those change. Every tool ships in every deployment: none of them
@@ -147,7 +165,14 @@ export const currentTime = defineTool(
  * who asked not to be remembered.
  */
 export function createBuiltinTools(config: WebAccessConfig, options: { memory?: boolean } = {}): Tool[] {
-  const tools = [createWebSearch(config), createReadPage(config), calculator, currentTime, weather]
+  const tools = [
+    createWebSearch(config),
+    createReadPage(config),
+    createResearch(config),
+    calculator,
+    currentTime,
+    weather,
+  ]
   return options.memory === false ? tools : [...tools, memory]
 }