io.github.Hlobo-dev/tengu-firm

Investor roster for a private company, resolved through the deal-investor relation; detail=full adds every investor column (AUM, dry powder, activity). Call this when the user asks 'who backed X / which VCs are on the cap table'; use /deals for the rounds themselves.

LIVE real-time overlay for a private company — the fast-moving complement to the (weekly, possibly STALE) private-markets profile, on a short TTL with per-field source provenance. ALWAYS call this for a private company to check whether it has GONE PUBLIC: it returns a `public_transition` block (new ticker, IPO date, SEC evidence) even when the snapshot still shows 'In IPO Registration'/no ticker — then pivot to the public-equity tools (price_snapshot / company_facts / sec_filings) for live data. Also surfaces live headcount / jobs / funding freshness when those providers are set.

Traverse a company's private-market graph one edge per call (relation=): competitors, investors, board, similar companies, affiliates, buyside targets, service providers, industries/verticals, news, financials, employee history and more. Call this when the user asks who backs, competes with, or sits on the board of a company.

Single deal / financing round by deal_id: deal size, type, VC round, pre/post-money valuation, and a synopsis. Call this when the user asks about a specific round ('what was the Series C?'); use deal_relations for the investors, lenders, and tranches behind it.

Traverse a deal's graph one edge per call (relation=): investors, tranches, debt lenders, sellers, service providers, bonds, loans, distribution beneficiaries. Call this when the user asks who funded, lent into, or sold in a specific round after fetching the deal itself.

Fund profile + performance: vintage, size, category, status, and returns (IRR/DPI/TVPI/RVPI/NAV/quartile) when loaded.

Traverse a fund's graph one edge per call (relation=): investors, LP commitments, team, portfolio holdings, close history, fund family, service providers, and returns time-series. Call this when the user asks what a fund holds, who committed capital, or how it has performed.

Investor profile for one VC/PE/family office by investor_id: type, AUM, dry powder, year founded, investment focus, and median valuation/round. Call this when the user asks who an investor is or how big/active they are; use investor_relations for portfolio, funds, and co-investors.

Traverse an investor's graph one edge per call (relation=): portfolio companies, funds raised, co-investors, LPs, board, deals, news, and investment focus by industry/year. Call this when the user asks what a VC/PE firm holds, who invests alongside it, or who its LPs are.

Limited-partner profile by lp_id: LP type, AUM, total/active commitments, allocation to alternatives, and openness to first-time funds. Call this when the user asks who an LP is or how much it commits; use limited_partner_relations for its fund commitments and mandates.

Traverse an LP's graph one edge per call (relation=): fund commitments, board, mandates, news, service providers. Call this when the user asks which funds an LP has committed to or what mandates it is running.

Person profile (founder/exec/board member) by person_id: role, board seats, affiliated deals/funds, education, and professional-profile link. Call this when the user asks who a founder or executive is; use person_relations to walk their full career and deal history.

Traverse a person's graph one edge per call (relation=): career positions, board seats, education, affiliated deals/funds, advisory roles. Call this when the user asks where a founder or exec worked before, what boards they sit on, or which deals they touched.

Catalogue of every queryable relation per private-markets entity type — the map of the private-capital graph. Call it first when unsure which relation= value a company/investor/fund/deal/person/LP relations tool accepts before traversing.

Search PRIVATE companies / investors (VC/PE) / funds / people / limited partners by name (prefix, case-insensitive), ticker, or CIK — relevance-ranked so the prominent entity is #1 (brand/AKA/former-name aware: 'Nubank'→Nu Holdings, 'Square'→Block). Use this FIRST for any private-company question (e.g. 'tell me about Stripe', 'who is Sequoia') to resolve the entity id, then call the company/dossier/realtime tools. type=all searches every entity kind. detail=full returns every column per hit (for rich tables).

INSTANT (sub-100ms) private-company typeahead — prominence-ranked with the SAME ranking as search, so the famous company is never truncated; each hit carries authoritative website/domain/logo_url plus sector, last-known valuation, ticker. Call this FIRST to resolve a name to company_id; use /search for multi-entity or detail=full.

Advisory-firm profile by sp_id — the law firm, investment bank, auditor or consultancy behind private-market deals: type, employees, parent, and count of companies serviced. Call this when the user asks who advised a deal or how big an adviser is; use service_provider_relations for its client list.

Reverse adviser index for one service provider: every deal, company, investor, fund or LP the firm has advised (relation= picks the edge, limit default 50). PRIMARY tool for 'which deals did this bank run?' / 'who does this law firm work for?' — the counterparty-selection view the forward service_providers relation cannot express.

BATCH CROSSWALK — up to 100 symbols resolved in ONE pass, each returning its current identifiers (CUSIP9, CUSIP8, ISIN, SEDOL, ticker, estimate-vendor ticker), its issuer keys (gvkey, CUSIP6, regulator filer number), which rung matched and its status. This is the endpoint for mapping a whole portfolio or watchlist in a single call rather than N of them. Every symbol comes back with an explicit status — resolved, historical (renamed or delisted; the security and what it trades as now are both returned), ambiguous (shared by more than one current security: the candidates are listed and NO identity is guessed) or unknown (not carried). Symbols beyond the 100 limit come back as explicit over_limit entries, never silently dropped, and the response reports how many of the requested symbols resolved as a number and a percentage (measured 2026-08-02 on a 102-symbol large-cap list: 100 looked up, 99 resolved, 1 historical, 2 over_limit).

CROSSWALK COVERAGE — what the reference dataset actually contains, measured live and stated as numbers rather than adjectives: securities and issuers carried and how many are still active, identifier bindings held and how many are RETIRED, the percentage of securities carrying each rung, how many issuers bridge into the private-company graph and how those links break down by confidence grade, and how many dated issuer change events are on file. It also NAMES the rungs this crosswalk deliberately does not serve and why — a missing rung is a stated fact here, not a silent gap a customer discovers after integrating. Every source is listed with its own as-of date, its age in days and whether that age is beyond its expected refresh cadence, so a vintage component can never pass as current. Live-measured 2026-08-02: 77,313 securities (25,512 active), 58,179 issuers, 969,333 identifier bindings (209,715 current / 759,618 retired) across 9 indexed rungs, 26,189 issuers (45.0%) linked to the private graph. Call it before you buy, or before you build against it.

IDENTITY CROSSWALK for one symbol — every identifier the security and its issuer carry, which rung matched, and how confident that match is. Returns SECURITY-grain identifiers (CUSIP9, CUSIP8, ISIN, SEDOL, ticker, estimate-vendor ticker) kept deliberately SEPARATE from the ISSUER-grain keys (gvkey, CUSIP6, regulator filer number, entity id), because those identify a company and not a share line; plus the issuer's other listed securities, the dated timeline of every identifier this security has ever been bound to, and the bridge into the private-company graph with that link's confidence grade. Call it to join two data sets that key on different identifiers, to translate a CUSIP-keyed holdings file into tickers, or to find out what a symbol used to be. as_of=YYYY-MM-DD returns the identifiers that were IN FORCE on that date, not today's — FB resolves today to an ETF and Meta's 2012-02-01→2022-06-08 hold on the string comes back as a dated prior binding, never as the answer. A symbol shared by more than one current security is REFUSED with its candidate list rather than guessed; pass prefer_country to choose. Coverage is a number in every response: what THIS answer contains, and the corpus census (77,313 securities / 58,179 issuers / 969,333 identifier bindings / 26,189 issuers linked to the private graph, live-measured 2026-08-02). Identifier-history depth is uneven by construction — 38,850 of 77,313 securities carry a prior ticker — and every source block states its own as-of, its age in days and whether that age is past its expected refresh cadence.

IDENTIFIER HISTORY for one symbol — every identifier this security has ever been bound to, with the dates each binding started and ended, plus the dated events where the ISSUER changed its name, ticker or CUSIP. This is the part of a crosswalk a current-state table cannot give you: it answers 'what was this on 2015-06-30', 'when did this CUSIP change' and 'what was this company called then'. Returns the timeline grouped by identifier rung (each entry with valid_from, valid_thru and whether it is still current), the issuer change events with what changed at each one, and — with as_of — the exact set of identifiers in force on that date. Use it to back-map a historical holdings file, to audit identifier drift in your own data, or to explain a ticker that no longer exists. STALENESS, stated rather than implied: the issuer change log is a VINTAGE snapshot that lags the live security master (as-of 2026-04-20 when this shipped, 104 days old, newest change event 2026-01-30). Its age in days and its expected cadence are in every response, and any disagreement between it and the live spine is named explicitly rather than blended into one confident answer. Coverage is a number: 969,333 identifier bindings on file, 759,618 of them retired, 38,850 of 77,313 securities carrying at least one prior ticker.

REVERSE IDENTIFIER LOOKUP — give it a CUSIP, ISIN, SEDOL, ticker, estimate-vendor ticker, gvkey, regulator filer number, entity id or private-company-graph id and it returns the security (or, for an ISSUER key, every security under that issuer) plus which rung matched and whether that binding is still current. Exactly one identifier per call. RETIRED identifiers resolve by default — 759,618 of the 969,333 bindings on file are retired, which is precisely why a stale identifier in your own data still lands, and it comes back with the date it was replaced and by what. Issuer keys NEVER nominate one security: an issuer can carry a common line, other share classes and a foreign listing, so all of them are returned and the grain is stated. Use it to turn a CUSIP-keyed custodian file into tickers, to find the ticker behind a regulator filing, or to audit whether an identifier in your own data has gone stale. Rung coverage, measured 2026-08-02: CUSIP9 and CUSIP8 100% of 77,313 securities, ISIN 79.4%, SEDOL 74.2%, regulator filer number 67.8%, estimate-vendor ticker 37.2%.

Returns-correlation mesh around a seed ticker — nodes = tickers, edges = |rho| >= 0.6, from the 36-month realised-returns sidecar (depth expands the neighborhood). PRIMARY tool for 'what moves with X?' and finding hedge or pair candidates. Caveat: needs 36-month overlap; a freshly-listed or thinly-traded seed may be absent.

Read ONE specific page you already have the URL for — call ONLY after a search returned the link, never to discover pages. Returns the page's readable body as clean markdown with title, author, published date, and a domain trust tier; handles paywalls. Read at most the 2-3 most load-bearing links per question. NOT a search tool: tengu_v3_research_web_search finds the links this tool reads.

Citation-rich research synthesis (grounded LLM). Returns a concise synthesized answer plus the list of source URLs that grounded it. Use when the user wants the answer GROUNDED with explicit sources (e.g. 'summarize NVIDIA's last earnings call and link the transcript'). Distinct from web_search — this returns prose + citations, not a list of headlines.

Multi-mode web search in parallel across link, cited-synthesis and social-sentiment coverage. Returns synthesized_answer, social_sentiment, citations, and a deduplicated list of links. Call this BEFORE relying on model knowledge for time-sensitive questions (markets, earnings, regulation, breaking news, macro events). 15-min LRU cache.

Real-time X/Twitter sentiment narrative. Pass ticker=NVDA for a focused fintwit read on a name, or query=... for a free-form social-media question. Returns the narrative answer with quantified bullish/bearish ratio and any source URLs social-search grounded against. Use when you want the *vibe* on a name right now (retail sentiment, breaking rumours, unusual social activity), not the news article list.

Outcome-resolver status

Consensus portfolio VaR + CVaR (USD) on the live top-decile shadow book — Cornish-Fisher + t-copula Monte-Carlo + filtered-historical-simulation blended, with a liquidity-adjusted VaR. Call this when the user asks how much the model portfolio could lose. Caveat: 1-day horizon only (horizon_days_served=1); multi-day is not scaled.

Structured extracts from the latest 10-K/10-Q/8-K via the SEC EDGAR JSON API: balance_sheet_summary, cash_flow_summary, shares_outstanding (basic+diluted+4Q trend), filing_date, filing_url, accession_number. Call for numbers straight from the latest filing. MVP: text_sections are deep-links only — use web_search on filing_url. 24h cache.

Recent and pending self-healing actions the system has taken or queued. Call when the user asks whether the system/data pipeline is healthy or why data looks missing/stale.

Cross-asset regime signals — the regime read taken across asset classes rather than from single tickers. Call this when the user asks about the broader market regime or wants cross-asset confirmation of a single-asset view.

Latest fused signals across all voters — the combined signal after voter aggregation, up to `limit` names (default 100). Call this when the user asks what the system's current signals are overall; use signals_mtf for one ticker's timeframe alignment.

One EQUITY cross-horizon ensemble row for a ticker. Despite the legacy `mtf` path name, the warehouse does NOT emit independent per-timeframe rows, so this tool cannot confirm 1m/4h/1w agreement and must not be cited as multi-timeframe evidence. Crypto requests fail closed.

Active veto state — which risk, regime, or circuit-breaker (CB) vetoes are currently in force over signals. Call it to know whether signals are being suppressed before trusting any signal read; PRIMARY for 'why isn't the system acting on <ticker>?'.

Apex Equity Intelligence — single-call omnibus brief for 'what do you think about $TICKER'. Fuses 12 vendor calls in parallel (daily + hourly charts, company facts, income/balance statements, snapshot, insider, congress, lobbying, gov-contracts, WSB, patents, off-exchange). Returns: 'fundamentals' (12 scalar fields — name, sector, market_cap, last_price, day_change_pct, latest_revenue, latest_net_income, latest_eps, total_assets/liabilities/equity, filing_date), 'intel' (8 scalar fields — insider/congress 30d buy-sell counts, ttm_lobbying_usd, ttm_gov_contract_usd, wsb_7d_mentions+sentiment, patent_filings_recent, avg_dark_pool_pct_30d), 'flags' (notable patterns: insider_cluster_buy/sell, elevated_dark_pool, retail_attention_spike, etc.), 'chart' (primary daily 120-bar candlestick PNG with RSI/MACD/BB indicators, palette-quantized to keep the tool result under 32KB), and 'charts' (a list of ADDITIONAL charts beyond the primary — currently just the hourly intraday chart; do NOT expect the daily here, it's only in 'chart'). 'components' carries raw counts of items behind each digest. 'vendor_errors' is non-empty when one of the parallel fetches failed; the rest of the payload is still usable. For full per-section detail (complete insider trade list, full income statements, etc.) call the dedicated tools tengu_v3_fundamentals_*/tengu_v3_intel_* — apex_equity is the digest, not the firehose.

One-call technical read on a ticker: fuses candlestick chart (RSI/MACD/BB), GEX, max pain, options flow/volume, insider and congressional trades, and off-exchange volume into a signal list, aggregate bull/bear stance, and embedded PNG chart. Call this FIRST for 'how does the chart/setup look?' — one round-trip replaces ~8 calls.

Actionable long/short plan for a ticker: entry, stop (recent swing or 1.5x-ATR proxy), 1R/2R/3R targets, position size for a given risk_pct, plus a thesis citing supporting signals (trend, flow tilt, insider, congress) and an embedded PNG chart. Call this when the user asks 'how would I trade X'; bias=auto picks direction from TA stance.

Realised slippage stats (mean/median/p95)

Recommended venue mix (lit/dark/algo)

Catalogue of the 21-strategy library — one entry per strategy. Call this when the user asks which strategies exist or what the system can run, or to resolve a strategy name before drilling into its evolution or signals.

Historical metrics for a named strategy

Latest signals from a named strategy

SSE stream of risk / guardrail alerts

SSE stream of live 1-minute OHLCV bars for EVERY US-listed ticker (no subscribe step needed — the all-ticker feed is always on). Each frame closes one 1-minute candle ~1s after the venue closes it: open/high/low/close/volume/vwap + bar_period_s=60. The live-candle feed for charting; quiet outside ~04:00-20:00 ET, which is correct. For chart HISTORY use REST aggregates, not a stream.

Effective decision-generator status across the optional in-process worker and dedicated processor heartbeat: input signals, emitted decisions, deployment state and local gate counters. Call when decisions are missing.

SSE stream of decision-lifecycle events

Per-user scoped SSE event stream

Live-data ingest daemon status: provider, subscribed symbols, connection uptime, tick/bar counters, last error. Call when the user asks why live data looks missing/stale.