v1.3.0 — The second brain gets depth

The second brain gets depth release. v1.3 keeps private memory grounded in the current user’s own writing while making long sources, retrieval scope, progressive answers, deliberate automation inputs, and documentation gaps inspectable as that memory grows.

Added

• Chunk-aware Second Brain memory — long private notes, journal entries, portfolio theses, holding notes, and transaction notes are split into deterministic, lightly overlapping chunks. Stable internal keys preserve original citations; incremental reindexing embeds only changed chunks and safely prunes stale rows.
• Inspectable evidence scopes — asks can restrict retrieval to validated private source types. The UI exposes source filters, per-source index counts, and the evidence scope used for each answer.
• Incremental synthesis — compatible OpenAI-style providers stream grounded answers over authenticated NDJSON. Stable non-streaming and retrieval-only fallbacks remain complete, cited, and honestly labelled.
• API-key external note ingestion — read_write keys can idempotently upsert owner-scoped private notes at PUT /api/v1/notes/external, supporting deliberate inputs such as Hermes YouTube summaries.
• On-demand journal completeness review — the Trade Journal identifies absent rationale, outcomes, emotions, setup labels, and stale open-trade thesis updates, then links each result to its editor without inventing facts or advice.

Verification

• Backend: 817 tests passed, 70.15% coverage
• Frontend: 296 tests across 92 files
• Production build, compileall, fabricated-data guard, Docker Compose configuration, and host smoke approved

v1.2.0 — Research interrogates

[1.2.0] - 2026-08-31

The "research interrogates" release. v1.2 turns the private research layer
from passive storage into an adversarial assistant: it challenges a thesis with
asset-appropriate market evidence and the user's own notes, while the News hub
adds optional, transparent LLM sentiment without making inference automatic.

Added

• Adversarial Interrogate research for equities, crypto assets, and indices.
  The structured analysis pressure-tests the bull case, identifies missing
  evidence and disconfirming signals, and grounds its challenge in semantically
  related private notes when available.
• Crypto- and index-aware research context. Crypto interrogation uses
  tokenomics, dilution, TVL, fees, and on-chain context; index interrogation uses
  regime, breadth, constituents, and index-level market context instead of
  pretending either asset is an operating company.
• Crypto-native and broader ticker news. Keyless crypto RSS ingestion,
  crypto firehose routing, index news, and EU-aware search terms improve coverage
  outside US equities.
• Optional batched AI article sentiment. The backend scores and caches up to
  20 headlines in one structured LLM call and falls back per article to the
  classical lexical engine when the model is disabled or unreachable.
• News hub research workflow. Market-wide browsing, US/Europe/India/crypto/
  index slices, search and ticker scopes, sentiment drill-in, publisher-summary
  enrichment, and explicit on-demand AI scoring of visible headline batches.

Changed

• Asset-aware AI research — stock, crypto, and index briefings/interrogations
  now use domain-appropriate facts and prompts. Coins are evaluated through
  tokenomics/on-chain context and indices through regime/breadth context rather
  than being treated as companies with issuer fundamentals.
• On-demand News hub AI sentiment — the main market, ticker, and search news
  feeds can score visible headlines in batches of at most 20, overlay concise
  rationales, and clearly identify LLM, mixed, or lexical-fallback results.
• More portable structured LLM output — the requested JSON schema is carried
  in the prompt for providers that ignore response_format, and truncated
  structured responses receive one larger-budget retry. Frontend AI requests now
  allow enough time for local and reasoning models to complete.

Fixed

• Fresh AI regeneration and note grounding — Interrogate cache keys now
  follow each user's note state, and the UI's Regenerate action explicitly
  bypasses cached briefing/interrogation results. Unavailable insight generation
  is labelled honestly instead of being misreported as a lexical fallback.
• News routing and classification — crypto articles are tagged at ingestion,
  coin ticker requests use the crypto firehose, index headlines resolve, and EU
  company aliases no longer bias results toward unrelated US tickers.
• Portfolio and PostgreSQL regressions — projected dividends are included for
  regular distributors, sells reduce or remove their holding as well as updating
  cash, and screener materialised-store introspection works on PostgreSQL rather
  than relying on SQLite-only PRAGMA calls.

v1.1.0 — Portfolio becomes real

[1.1.0] - 2026-07-05

The "portfolio becomes real" release. v1.0 hardened the research half of the
north star (never show fabricated market data as live); 1.1 turns that same
integrity lens on the portfolio half — a real cash/transaction spine, honest
realized-vs-unrealized P&L, and the retirement of the old global, shared-across-
users portfolio (a privacy contradiction) in favour of a per-user one.

Added

• Portfolio cash ledger (v1.1 "portfolio becomes real", spine) — cash is now
  derived from the transaction ledger as the single source of truth: a buy debits
  cash, a sell/dividend/deposit credits it, a withdrawal debits it, and fees always
  cost cash. New deposit/withdrawal transaction types (cash-only, no symbol).
  GET /api/portfolios, /api/portfolios/{id}, and /api/portfolios/{id}/analytics
  now report cash_balance and net_liquidation_value (holdings + cash).
• Portfolio transaction UI (v1.1) — the Portfolio Manager can now record
  buy/sell/dividend/deposit/withdrawal transactions (the endpoints existed but no
  UI ever called them), with a live cash-impact preview and a trade-history table.
  New Net Liquidation and Cash metric cards surface the ledger balance.
• Upcoming events in the Portfolio Manager (v1.1) — dividends, earnings and
  corporate actions for your holdings now show in the Manager's events calendar,
  not only on the legacy view / dedicated Dividends page.
• Deep analytics on the per-user portfolio (v1.1, consolidation) — the
  Portfolio Manager now carries everything the old view had: correlation matrix,
  benchmark-overlay equity curve, the full risk-metric set (Sharpe, Sortino, beta,
  alpha, information ratio, max drawdown), an upcoming-dividend tracker, and
  Brinson + factor attribution — all scoped to the selected portfolio via new
  owner-checked GET /api/portfolios/{id}/analytics/* and /attribution routes
  (with a primary alias for dashboards).
• AI Risk Assessment + backtesting in the Manager (v1.1) — the local-LLM risk
  narrative (on-demand) and the backtester (seeded from the portfolio's holdings)
  now live in the Manager.
• Sell from a holding row (v1.1) — each holding has a "Sell" button that
  pre-fills the transaction form (sell / that symbol / full size / current price)
  so a sale is attributed to the position instead of typed from scratch.
• Holdings CSV export in the Manager (it already had CSV import).

Changed

• The equity portfolio is now per-user, full stop (v1.1, consolidation) — the
  Portfolio Manager (/api/portfolios) is the only portfolio; the legacy
  manager|legacy view toggle is gone and every surface that read a portfolio
  (home, cockpit, HUD, launchpad, correlation dashboard, risk / stress / factor /
  dividends / reports) now reads the caller's own primary portfolio via
  GET /api/portfolios/primary.

Removed

• The global, user-less legacy portfolio (v1.1, consolidation) — the shared
  Holding table and its /api/portfolio, /api/portfolio/analytics/*,
  /api/portfolio/holdings* and /api/portfolio/{id}/attribution endpoints are
  deleted. On a shared instance every user saw and edited the same holdings — a
  contradiction of the "private terminal" thesis; portfolios are per-user now.
  (The global /watchlists/items feed is unaffected.)
• Tax-lot accounting (the TaxLot model, /api/portfolio/tax-lots* endpoints
  and the Tax Lot Manager UI) — a jurisdiction-specific rabbit hole where a wrong
  number is worse than none; deliberately shelved.

Fixed

• Import from Legacy / bulk add crashed with a 500 (v1.1) — adding a holding
  auto-generated its lot_id from a second-resolution timestamp, so adding the
  same symbol more than once within a second (as the legacy import does) collided
  on the (portfolio_id, symbol, lot_id) unique constraint. Auto lot_ids now
  carry a random suffix, and a genuine duplicate lot_id returns a clean 409
  instead of an unhandled 500.
• Realized P&L was proceeds, not gains (v1.1) — analytics summed
  shares * sell_price - fees, overstating realized P&L by the entire cost basis.
  Now computed by replaying the ledger chronologically (running average cost per
  symbol): shares_sold * (sell_price - avg_cost) - fees, dividends excluded
  (they remain in dividend_income_ytd). New pure backend/services/portfolio_pnl.py.
  Annualized return now uses net liquidation (holdings + cash) as current equity.
  The "Total P&L" card is relabelled "Unrealized P&L" to match what it shows.
• Multi-portfolio analytics annualized-return crashed on portfolios whose
  transactions carried bare YYYY-MM-DD dates (naive vs. tz-aware datetime
  subtraction); inception dates are now normalised to UTC.
• Multi-portfolio analytics used India defaults for risk metrics (risk-free 0.06,
  benchmark NIFTY50); now 0.04 / S&P500 in line with the de-India defaults.

v1.0.0

[1.0.0] - 2026-06-29

The first tagged release. A hardening milestone: a coherent, honest,
installable product where every advertised feature works or is explicitly
labelled degraded. All release buckets are complete — integrity (A), de-India
defaults (B), robustness (C), the single-source version contract (D), and
ship-with docs (E). See docs/wiki/Roadmap.md → Release plan for the full
gate and docs/wiki/Limitations.md for the honest keys/limitations map.

Added

• Private second brain (RAG) — flagship: ask-anything research grounded only
  in your own writing (journal, theses, position/transaction notes), with inline
  citations; provider-agnostic embeddings (Ollama nomic-embed-text default,
  sentence-transformers fallback) and a dialect-aware vector store (pgvector on
  Postgres, numpy cosine on SQLite). POST /api/brain/ask|reindex, GET /api/brain/status + a Second Brain page.
• Generic notes capture — one-step note capture anywhere (stock/crypto detail,
  watchlist rows, news, Portfolio Lab, Notes hub) feeding the second brain.
• Crypto fundamentals — tokenomics + on-chain TVL/fees (CoinGecko + keyless
  DefiLlama); GET /api/v1/crypto/fundamentals/{symbol}.
• EUR display currency — cross-rates-driven, instrument-currency-aware display
  engine (USD/EUR/INR) with correct per-currency symbol/locale/compaction.
• Dividend calendar (real data) — market-agnostic calendar/history/aristocrats/
  income via the corporate-actions service (Yahoo chart events=div + FMP), with
  labelled next-ex-date projections for regular distributors.
• Scheduled reports + PDF generation backend; provider-agnostic LLM layer
  (OpenAI-compatible, Ollama default); economic data terminal (calendar +
  macro dashboard).

Changed

• Postgres-first persistence (SQLite opt-in); pgvector Postgres image
  (pgvector/pgvector:0.8.3-pg16-trixie, no migration).
• De-India data layer — unified instrument_master (US/EU/crypto sources),
  accent-insensitive search, EU Yahoo-suffix routing; market-overview/ticker-tape
  show US indices. NSE/BSE F&O stays supported.
• External-client robustness (1.0 bucket C). A shared retry-with-jitter +
  per-client circuit breaker (backend/shared/http_resilience.py) backs off on
  429/5xx instead of hammering a depleted provider, and response caching now
  covers Finnhub and Yahoo as well as FMP (the shared multi-tier cache
  incl. the persistent SQLite L3) to stop free-tier quota burn; 429/5xx never
  cached. Finnhub no longer force-appends .NS to bare symbols.

Fixed

• Silent-mock integrity sweep (1.0 bucket A, part 1). Audited the backend for
  code that fabricated plausible market data when a real source was unavailable
  and presented it as live. Established a standard degraded: {reason, source}
  marker (backend/shared/degraded.py) + a reusable DegradedBanner so the UI
  always flags non-live data. Fixed the 7 critical sites:
  • Market status (/reports/market-status): removed the hardcoded
    index/commodity values + random jitter (NIFTY/SP500/gold/silver/crude…);
    missing quotes now stay null and the response is flagged degraded.
  • Heatmap treemap: stopped seeding synthetic price/change/volume per tile;
    tiles with no live quote render neutral and the response is flagged degraded.
  • Tape (/tape/*): removed synthetic bars and invented buy/sell order
    flow; only real trade-feed data is shown, otherwise empty + degraded.
  • Historical OHLCV (historical_data_service): removed the synthetic
    random-walk fallback so backtests/charts never run on fabricated history.
  • Macro indicators: the no-FRED_API_KEY fallback is now flagged degraded
    (previously only the calendar carried a sample flag, the macro dash didn't).
  • Yield curve / 2s10s (fixed_income): no-key/error paths return empty +
    degraded instead of a hardcoded curve.
  • Bonds (/bonds/*): replaced the hardcoded India-only bond universe,
    spreads, and ratings with empty + degraded (no live source wired yet).
• Silent-mock integrity sweep (1.0 bucket A, part 2). Extended the audit to
  the high-severity sites and a few the first pass missed:
  • Hotlists (/api/hotlists): the page 401'd because it used a raw fetch()
    that bypassed the bearer-token client — now routed through the authed api
    client. The endpoint itself was 100% fabricated (hardcoded prices + seeded
    sparklines) → now empty + degraded; default market de-India'd to US.
  • Insider (/api/insider/*): removed the auto-seeded fabricated India
    insider universe (source="SEEDED"); serves only ingested trades, else
    empty + degraded. Deleted the dead, mock-only insider_monitor.py.
  • ETF (/api/etf/*): removed the hardcoded screener (incl. NIFTYBEES),
    the AAPL/MSFT holdings placeholder, and the random fund-flows; empty +
    degraded instead. ETF FE components routed through the authed api client
    (same latent 401 as hotlists).
  • Shareholding: the no-source fallback no longer fabricates a "100%
    public" cap table — returns zeros + degraded.
  • Sector rotation (RRG): removed the hashed-seed synthetic price series;
    empty + degraded when the price source is unavailable.
  • Charts (/chart/{ticker} + indicators): removed the synthetic
    random-walk fallback; empty + degraded instead of a fake series.
  • Crypto correlation: drops assets without real return history rather than
    back-filling a synthetic sine/cosine series; flags degraded when partial.
• Silent-mock integrity sweep (1.0 bucket A, part 3) — real Binance crypto
  microstructure + derivatives. Replaced the last synthesized-crypto sites with
  real Binance public-REST data (new backend/core/binance_client.py +
  backend/services/crypto_derivatives_service.py):
  • Crypto heatmap depth (/v1/crypto/heatmap): depth_bid/ask_notional +
    depth_imbalance are now the real best bid/ask from the spot order book
    (bookTicker, one call, cached), not volume*price. Symbols Binance doesn't
    list read as empty depth; no order-book coverage at all flags degraded.
  • Crypto derivatives (/v1/crypto/derivatives): funding_rate_8h is the
    real last funding rate (futures premiumIndex) and open_interest_usd is
    real (openInterest × mark price), no longer derived from change_24h.
    24h liquidations have no Binance REST endpoint, so they are read only from the
    live forceOrder WebSocket stream (BinanceDerivativesState); until that
    runner is wired they read 0 and the response is flagged no_live_source
    instead of being fabricated. DegradedBanner wired into both crypto tabs.
• Silent-mock integrity sweep (1.0 bucket A, part 4) — real order-book depth.
  The /api/depth/{symbol} REST route and the /ws/depth stream returned a
  fully synthetic order book for every market — prices, sizes, spread and
  imbalance were all derived from a sha256(symbol:market:levels) hash (a site
  the earlier audit missed). Both now serve real depth via unified_fetcher. fetch_depth:
  • Crypto → real Binance spot order book (/api/v3/depth, new
    BinanceClient.get_order_book); the FE panel detects crypto symbols so a
    crypto detail page no longer requests depth as an equity market.
  • India → already had real Kite/NSE depth in fetch_depth, but the route
    bypassed it; it's now wired through (real).
  • US / EU equity → no free Level-2 source, so the book is empty +
    degraded (no_live_source) instead of fabricated; OrderBookPanel shows
    the degraded banner. Real US/EU L2 is a roadmap item (Interactive Brokers).
  • Depth quantities widened to floats (crypto books are fractional); the
    synthetic orderbook_service generator was deleted.
• De-India defaults (1.0 bucket B). Western-oriented defaults across the app
  (India markets stay supported, just not the default): watchlist symbol search
  passes the selected market through instead of forcing NSE; home/sidebar F&O
  widgets default to SPY (sidebar label → "EQUITY ANALYTICS"); the portfolio
  risk/attribution/lab benchmark defaults to S&P500 (FE + backend route &
  service defaults) and the default risk-free rate is 0.04 (was 0.06).
  NIFTY50/SENSEX remain selectable benchmark options. The F&O heatmap's
  empty/error copy is now market-neutral ("F&O options data feed") instead of
  hard-coding "Kite API key required".
• Relative Strength stub de-India'd + degraded (1.0 bucket A). Every /rs/*
  endpoint (rankings, sector-rs, chart, new-highs) previously returned hardcoded
  fake Indian rows (RELIANCE/TCS/INFY…) presented as live; they now return empty
  • degraded and the FE defaults the universe to S&P 500. The real RS
    computation is a tracked post-1.0 follow-up.
• Scheduled-report 422 on missing email (1.0 bucket A). POST /api/reports/scheduled rejected a blank email; it now defaults to the
  authenticated user's account email and only errors when there's genuinely no
  address.
• Index detail page (1.0 bucket A). Navigating to a market index (^GSPC,
  ^NSEI, …) rendered the full equity tab set — Financials/Peers/Valuation/
  Earnings/Shareholding — all blank, since an index has no issuer fundamentals.
  An index now shows only the chart-backed overview + notes, with an explicit
  "fundamentals don't apply to an index" note; price/chart/performance (which do
  apply) are unchanged. isIndexSymbol also keeps ^NSEI from being treated as
  an NSE equity.
• FMP corporate-actions migrated to /stable. corporate_actions.py was
  missed in the original /stable migration and still used legacy /api/v3
  path-segment URLs that 404'd (dividends/splits/IPO calendar). Now uses the
  stable /dividends, /splits, /ipos-calendar shapes, skips crypto, and
  caches 404s as empty to stop log noise...