bvb-news

name

bvb-news

description

Monitor BVB company announcements, Romanian financial press, and corporate actions for anything that could move BET-Plus names. Runs on both morning and evening routines. Uses a mix of structured sources (per-symbol BVB announcement pages) and real-time web search for third-party coverage. Trigger whenever the pipeline needs current BVB news, when evaluating a specific stock, or when the user asks what's happening on the Romanian market.

BVB News

Two layers of sourcing:

1. Structured (authoritative): BVB per-symbol announcement pages — gives dates, headlines, upcoming corporate events straight from the issuer.
2. Real-time search (breadth): WebSearch for each holding and the broader market — catches third-party coverage, analyst notes, sector context the issuer page won't have.

Both layers run every run. The structured layer is the anchor; search fills in the rest.

Layer 1 — BVB Per-Symbol Pages (WebFetch)

For every symbol in the current portfolio holdings (store.getState().holdings) PLUS the BET-Plus watchlist (defined in trade-executor/SKILL.md), fetch:

https://bvb.ro/FinancialInstruments/Details/FinancialInstrumentsDetails.aspx?s=<SYMBOL>

Extract:

• Recent announcements (last 30 days) — dates, headlines
• Upcoming corporate events — AGM, earnings publication, ex-dividend, coupon payments
• Trading halts or suspensions if flagged on the page

Example (extracted by WebFetch with prompt: "List company announcements and upcoming corporate events with dates"):

17.04.2026 — Current report, contract disclosure (Art 234 lit i Reg 5)
16.04.2026 — AGOA held, no decisions adopted
29.04.2026 — AGA Ordinara Anuala scheduled
15.05.2026 — Q1 2026 financial results publication

What to extract / what to skip

• Keep: earnings dates, dividend announcements, AGM decisions with material content, share buybacks, rights issues, M&A, management changes, large contracts
• Skip: routine "contract disclosure Art 234 lit i" filings unless the WebFetch summary says the contract is material

Layer 2 — Real-Time Web Search

Run these searches each morning (evening can be lighter — skip per-symbol unless a position moved >3% that day):

Per-holding search (highest priority)

For each symbol in the current portfolio's positions (store.getState().positions):

WebSearch: "<company full name> news <current month year>"
WebSearch: "<SYMBOL> BVB news this week"

Example: "Romgaz news April 2026", "SNG BVB news this week"

Market-wide search

WebSearch: "BVB Bucharest Stock Exchange news today"
WebSearch: "bursa valori bucuresti stiri azi"
WebSearch: "BET index today"

Sector search (on rotating days to stay fresh)

One sector per day, cycling through: Energy, Banking, Utilities, Real Estate, Telecoms/IT, Pharma.

WebSearch: "Romania <sector> regulation <current month year>"

Dividend / corporate action calendar

WebSearch: "BVB dividende ex-date <current month year>"
WebSearch: "BET-Plus dividend calendar <year>"

Trusted Sources

Prioritize results from these domains when WebSearch returns ambiguous or low-quality hits:

• bvb.ro — official
• zf.ro — Ziarul Financiar, main Romanian business daily
• bursa.ro, profit.ro, economica.net — business press
• cursdeguvernare.ro, hotnews.ro (economy section) — general business press
• reuters.com, bloomberg.com, tradingview.com — international coverage for larger names
• bvbromania.substack.com — weekly retrospective (Mondays, high signal)
• asfromania.ro — regulator (sanctions, investigations)

Flag but treat cautiously: simplywall.st, stockanalysis.com — algorithmic content, useful for context but not for breaking news.

Analysis Framework

For each news item that clears the noise filter, produce:

{
  "symbol": "SNG",
  "date": "2026-04-19",
  "headline": "Romgaz proposes 606M RON dividend, reduced from prior year",
  "materiality": "high|medium|low",
  "direction": "positive|negative|neutral",
  "timeframe": "immediate|this_week|this_month",
  "source": "bvb.ro per-symbol page | cursdeguvernare.ro | ...",
  "url": "https://...",
  "actionability": "buy|sell|hold|watch",
  "rationale": "one sentence"
}

Materiality scoring

• High: earnings, dividend announcements, ex-dividend/record dates, M&A, management changes at CEO/CFO level, large contract wins, guidance changes, trading halts
• Medium: routine shareholder meetings with substantive decisions, insider transactions, analyst rating changes with new data, sector-wide regulatory moves
• Low: generic market commentary, price target tweaks with no new info, minor contracts

Drop "low" items from the briefing unless they're the only news on a holding.

Output Format

📰 BVB NEWS — [DATE]

🚨 BREAKING (if any)
[High-materiality items needing immediate attention — trading halts, surprise results]

HOLDINGS
  [SYMBOL]: [headline] — [materiality] [direction] → [action/watch]
  ...

WATCHLIST
  [SYMBOL]: [headline] — [materiality] [direction] → [action/watch]
  ...

UPCOMING EVENTS (next 14 days)
  [DD.MM] [SYMBOL] [event]
  ...

MARKET CONTEXT
[1-2 sentences on broad BVB news flow, index moves, foreign flows]

SENTIMENT SCORE: [-5 to +5]

Failure Handling

• WebFetch fails on a BVB page: retry once. Still fails → log "NO_STRUCTURED_DATA for " and rely solely on WebSearch for that symbol.
• WebSearch returns no results for a query: try one reformulation; if still empty, record "no news found for " — don't hallucinate headlines.
• Every source dark (rare): produce a skeleton briefing with "news layer failed this run" banner. The synthesis step should treat it as a "hold decisions" signal, not invent missing data.

Caching

Within a single run, cache all fetched pages and search results. Don't hit the same URL twice (e.g., the same holding appears in multiple search queries). Cache is in-memory for the run; nothing persists between runs.

Graphiti Sync

Every news item that clears the materiality filter (high or medium only) gets ingested into graphiti as a News episode. Low-materiality and dropped items are not ingested — they're noise, the graph doesn't need them.

Ingest format

For each kept item, call:

mcp__graphiti__add_memory(
  name: "News <symbol> <date> — <truncated headline>",
  episode_body: JSON.stringify({
    symbol: "SNG",
    date: "2026-04-19",
    headline: "Romgaz proposes 606M RON dividend, reduced from prior year",
    summary: "<1-3 sentence pre-extracted summary, NOT full article body>",
    materiality: "high",
    direction: "negative",
    timeframe: "immediate",
    source: "bvb.ro",
    url: "https://...",
    rationale: "Dividend down 38% YoY — windfall tax absorption continues"
  }),
  source: "json",
  source_description: "bvb-news item",
  group_id: "auto_trader",
  reference_time: <ISO timestamp of publication, NOT scan time>
)

Critical: reference_time is the news publication time, not now. Graphiti's temporal model needs the event time so retrospectives can correlate "news from 2026-04-19" with "trade opened 2026-04-20."

What gets extracted

Graphiti's LLM will produce News, Company, Catalyst (for dated events like AGM/earnings), and MacroEvent (for regulatory/rate decisions) nodes. The direction and materiality become attributes on the News node so downstream queries can filter.

Pre-extraction over raw article bodies

Do NOT pass full article text into episode_body. Passing the full body:

• Wastes extraction LLM tokens on boilerplate (cookie banners, navigation, footers)
• Produces noisy entities (random people quoted in the article)
• Inflates ingestion latency

Instead, summarize each item to 1-3 sentences during the bvb-news analysis pass (you're already producing the structured JSON above), and pass that. The full URL is preserved for any future retrieval.

Market context summary as its own episode

At the end of the run, ingest a single rolled-up MarketContext episode covering the day's broad news flow:

mcp__graphiti__add_memory(
  name: "Market context <date> <morning|evening>",
  episode_body: JSON.stringify({
    date: "2026-04-19",
    session: "morning",
    sentiment_score: 2,
    summary: "Energy names up on TTF strength; BNR holds at 6.50%; foreign flows neutral",
    notable_themes: ["Neptun Deep", "windfall tax watch"],
    regime_signals: ["risk_on"]
  }),
  source: "json",
  source_description: "Daily BVB market context",
  group_id: "auto_trader",
  reference_time: <session start ISO timestamp>
)

This lets the next session ask "what was the market mood when we opened ONE?"

Failure handling

If the graphiti call fails, log to stderr + the Telegram report and continue. Do not abort the news pipeline — the briefing still has to go out.

News items are not mirrored to bt-gateway — they exist only in graphiti. An outage during a single news run loses that batch. This is acceptable because:

• The underlying sources (bvb.ro per-symbol pages, WebSearch hits) are still authoritative
• Re-running this skill re-fetches and re-ingests
• News from a few hours ago becoming visible to graphiti a day late is a minor loss of context, not a correctness problem

If a multi-day graphiti outage occurs, manually re-run this skill with a wider lookback to catch up.

Caching applies to ingestion too

If the same news item appears in both the BVB structured page and a WebSearch hit, ingest it ONCE. Dedupe by url within the run.