/news-scan

Cosa fa. Raccoglie le ultime notizie rilevanti su uno o più titoli.

Quando usarlo. Per restare aggiornato o capire un movimento.

Argomenti. TICKER... [1-30d] [1-25] [deep]

Esempio. /news-scan NVDA 7d 10

Cosa restituisce. Un elenco di notizie con data, fonte e sintesi.

Come funziona

Quando lanci /news-scan, Dexter non esegue codice fisso: invia all’AI la prompt qui sotto — il segnaposto {arg} viene sostituito con ciò che scrivi dopo il comando — e l’agente la esegue chiamando da sé gli strumenti necessari. Vedi Come funzionano i comandi per le convenzioni comuni (chiamate in parallelo, anti-allucinazione, fasi, widget di chiusura).

Strumenti che può usare: get_company_news, web_search, dispatch_agent.

Flusso/visualizzazione: può lanciare sub-agenti di ricerca in parallelo (dispatch_agent).

Prompt completa

È la prompt esatta inviata all’AI quando usi il comando (sostituendo {arg}).

News scan for {arg}.

Parse "{arg}" (space-separated tokens):
- The token "deep" (case-insensitive) = enable DEEP mode (topic fan-out). It is NOT a ticker.
- Every other bare word token (e.g. AAPL, MSFT) = a TICKER. One or more are required. If "{arg}" is blank, literally "{arg}", or contains no ticker, ask the user for at least one ticker and stop.
- A token like "14d" / "1d" / "30d" (digits followed by "d") = look-back PERIOD in days → recency_days.
- A bare integer (e.g. "15") = NUMBER of articles → limit (the overall budget).
- Tokens may appear in any order.
Defaults when omitted: recency_days=7, limit=10.
Clamp: recency_days to 1..30 (if more is requested, use 30 and say so); limit to 1..25 (if more, use 25 and say so).

Then pick exactly ONE branch:

## A) One ticker, no "deep" — quick linear scan
Call get_company_news(ticker=<TICKER>, recency_days=<resolved>, limit=<resolved>, text=true) and web_search for the ticker over the window. Write the report directly.

## B) One ticker + "deep" — topic fan-out
Choose 3-5 DISTINCT topic angles relevant to <TICKER> (e.g. earnings/guidance, contracts/government, analyst/price-targets, regulatory/legal, product/partnerships — tailor to the company). Emit ONE dispatch_agent per topic IN THE SAME TURN so they run in parallel. Each call: subagent_type="news", description "<TICKER> <topic>", and this self-contained prompt:
  "News scan for <TICKER>, focus: <TOPIC>. Window: last <resolved> days. Call get_company_news(ticker=<TICKER>, topic=<TOPIC>, recency_days=<resolved>, limit=<per-topic>, text=true) and a targeted web_search for '<TICKER> <TOPIC>' over that window. Return a digest: for each item source, date, and a 1-sentence summary; flag material-impact items; end with a sentiment for THIS angle (bullish/neutral/bearish) plus one-line reasoning."
Use <per-topic> = max(5, round(<resolved limit> / number-of-topics)). Do NOT fetch news yourself. Merge the digests and DEDUP overlapping items (same URL/headline across topics); present up to <resolved limit> unique items.

## C) Two or more tickers, no "deep"
Emit ONE dispatch_agent per ticker IN THE SAME TURN (parallel). Each call: subagent_type="news", description "<TICKER> news", self-contained prompt that calls get_company_news(ticker=<TICKER>, recency_days=<resolved>, limit=<resolved>, text=true) and web_search for that ticker over the window, and returns a structured digest (source, date, 1-sentence summary; material flags; sentiment + reasoning). Do NOT fetch yourself; assemble the per-ticker digests.

## D) Two or more tickers + "deep"
Emit ONE dispatch_agent per ticker IN THE SAME TURN (parallel). Each sub-agent prompt instructs IT to cover 3-5 topic angles internally (earnings, contracts/government, analyst, regulatory, product) via get_company_news(topic=...) + targeted web_search, dedup, and return a per-ticker digest with an overall sentiment. Keep depth INSIDE each ticker's sub-agent — do not spawn a ticker×topic matrix. Do NOT fetch yourself; assemble the per-ticker digests.

## Report (all branches)
State the resolved window, article count, and mode (quick/deep) at the top. Cover press releases, earnings, strategic updates, analyst commentary, breaking news. For each item: source, date, 1-sentence summary. Flag news with potential material impact and assign a per-ticker sentiment score (bullish/neutral/bearish) with reasoning.