Historic News Skill

Research and annotate price-moving news events on the chart. Uses the shared AgentSwarm service to spawn a research team.

The team

Role	Mandatory	Tools	Added when
Researcher	✅	search_web, fetch_news, fetch_url	Always — does the actual web queries for historic news
Analyzer	✅	—	Always — parses raw findings into structured NewsEvent objects with timestamps, categories, impact ratings, and direction
QA	✅	—	Always — filters out unsubstantiated events, checks timestamps are plausible, drops duplicates
Macro Researcher	⏳ optional	search_web, fetch_policy	Added when the asset is a commodity, currency, or index (where macro news dominates price action)
Regulatory Researcher	⏳ optional	fetch_policy, fetch_url	Added when the asset has active policy considerations (crypto, pharma, defense)

Flow

1. Team Planner picks which researchers to include based on asset class
2. Researchers run in parallel — each does 2-4 targeted web searches for news in different periods (pre-rally, drawdown, range-bound, etc.)
3. Analyzer merges findings into a structured event list with:
   • timestamp (unix seconds)
   • headline, summary, source, url
   • category — earnings | regulatory | macro | product | sentiment | geopolitical | technical
   • impact — high | medium | low
   • direction — bullish | bearish | neutral
   • price_impact_pct — rough estimate of the event's price effect (optional)
4. QA reviews: drops duplicates, flags timestamps outside the chart's data range, ensures each event has a credible source
5. Final event list is pushed to the store → chart renders vertical markers → Historic News tab lists them

Output

• On chart: colored vertical line + dot at each event's timestamp. Red for bearish, green for bullish, orange for neutral. Hover for tooltip.
• Bottom-panel tab: timeline list of events sorted newest-first. Click an event → chart zooms to that time range + selected article detail appears on the right.

Input — interactive directives

The skill parses your message for steering directives so you can drive the swarm:

You say	What happens
historic news for this chart	Researches the chart's own asset over its date range (default)
fetch oil news on this chart	Researches oil, plots markers on the loaded chart (different asset!)
plot AAPL news on the BTC chart	Researches AAPL, tags markers for the BTC chart specifically
show macro news on all charts	Broadcasts — same news set renders on every open chart
earnings news for TSLA	Topic-filtered — only earnings queries run
regulatory + macro news for AAPL	Multiple categories combined

Recognised category keywords: earnings, regulatory/SEC/policy, macro/fed/fomc/inflation/cpi, product/launch, geopolitical/war/sanctions. Recognised broadcast keywords: all charts, every chart, multiple charts, broadcast.

Requires a dataset loaded on the chart (the skill queries around the chart's visible date range and uses the chart's symbol as the default plot target).

How real news is fetched

The researcher does NOT hallucinate from training data. The processor:

1. Generates 6-9 search queries from the topic + date range + categories
2. Calls web_search() (DuckDuckGo, rate-limited) for each
3. Aggregates real results — real titles, real URLs, real snippets
4. Feeds those real findings to the analyzer agent

The analyzer is then explicitly told: "use the EXACT URL from the search result. NEVER invent URLs." So every Reuters/Bloomberg link in the output is a link the search engine actually returned.

Example

User: "Find historic news events that moved AAPL"

→ Team: Researcher + Analyzer + QA. Researcher queries for AAPL earnings, product launches, and macro shocks in the chart's date range. Analyzer extracts ~15-25 events with timestamps + summaries. QA drops 3 unsubstantiated ones. Final: 20 events plotted on chart with category-coloured markers. Bottom panel shows timeline.

User clicks the "Q1 earnings surprise" event → chart zooms to that date; right panel shows the full article summary + link.