Release Notes

import { Aside } from ‘@astrojs/starlight/components’;

The full changelog lives in RELEASE_NOTES.md in the repository. Highlights of the current stable are below.

v5.16.1 — One public review loop

Documentation

README and website tell the same story. Holdings, local read, attention, and portable record now form one short path from first impression to install.
Product proof is current and explicit. Review Inbox, Plan & Protect, and risk claims sit beside captures from an isolated fictional profile.
Download decisions are complete before the click. Platform, architecture, prerequisites, unsigned first-launch behavior, SHA-256 verification, and the lack of Minisign authenticity are named together.
Trust wording is synchronized. Local storage, provider lookups, optional Claude context, update behavior, and the no-brokerage/no-trade boundary now agree across the README, landing page, and detailed guides.
Source upgrades keep the complete profile. One-line installs move writable state outside replaceable source code, migrate canonical or custom relative SQLite databases safely, stop on conflicts and non-portable paths, and retain the prior install as a recovery copy. Every launcher honors the same absolute profile pointer; a process-selected database remains selected on relaunch, and later code replacements roll back if setup fails.
Windows proves the source update journey before packaging. The release gate parses PowerShell, migrates a fictional profile, repeats the install, and checks the SQLite row, profile path, settings, backups, shortcut, and recovery count.
Verification matches the download journey. macOS filters the shared manifest to its DMG, and the docs distinguish integrity from unavailable Minisign authenticity.

This patch changes documentation, public assets, version metadata, and source installer profile migration. It does not change portfolio calculations, providers, schema, trade boundaries, or the v5.16.0 duplicate-recovery workflow.

v5.16.0 — Review Bundle with manifest

Save one review handoff. The active Month/Quarter Review Pack, data-health receipt, and saved target-plan snapshot leave together as one local ZIP with a JSON manifest.
One quote boundary. All included review, trust, and planning files reuse one in-memory quote response set; assembly writes no holdings, trades, daily snapshots, targets, settings, DCA rows, or backups.
Check every review artifact. The manifest carries app/format versions, UTC generation time, period and per-area quality context, missing and foreign-priced tickers, byte lengths, and SHA-256 digests for the four non-manifest review artifacts.
Integrity, not authenticity. The hashes help detect member corruption; they do not authenticate the ZIP or prove who created it.
Keep financial gaps honest. Foreign positions stay outside USD totals, missing values stay missing, and no FX conversion or new performance math is introduced. An unsaved target draft blocks the bundle.
Save real ZIP bytes. Browsers download the response; packaged apps use a native Save dialog and atomic binary writer. The bundle is capped at 8 MiB and is review material—not a restore file, tax form, recommendation, or trade instruction.
Keep imports responsive. Quote warming, provider validation, search fallback, Claude remapping, and narration run away from the request thread. Fresh file-backed state is reconciled before writes, and concurrent imports, manual adds, DCA buys, and reactivations cannot create two active rows for the same Portfolio and ticker.
Keep one runtime profile. A custom database must be paired with its data root and remain beneath it. Invalid split profiles fail before any directory, database, journal, setting, backup, restore, update, or legacy-copy write.
Migrate without guessing. Schema v7 adds the active-holding uniqueness index through the existing verified-backup migration flow. A read-only preflight stops on existing active duplicates, shows packaged-app recovery steps, and leaves portfolio rows, schema metadata, and the primary database/WAL payload unchanged. SQLite may refresh only transient shared-memory bookkeeping while inspecting a live WAL database; no financial row is deleted or merged.
Resolve duplicates without inventing a sale. The packaged recovery window shows every conflicting row and requires one keep choice per ticker. It creates a verified Backup Vault copy before archiving rejected rows, preserving their shares, cost basis, notes, and history without adding realized-trade activity.
No new external provider or FX path.

v5.15.0 — Review receipts

Filter the Review Inbox. All, Data gaps, Needs review, and On the radar expose their counts and narrow only the visible list. The Review badge and header retain the full total; nothing is dismissed or mutated.
Save a data-health receipt. The Trust tab exports generation time, overall quality, snapshot coverage, area counts, missing tickers, sources, caveats, and named foreign-currency exclusions as CSV. Saving runs a fresh read-only coverage check; partial remains partial.
Save a current target-course snapshot. Plan exports persisted basis points, current weights, descriptive drift, known USD value, valuation quality, and missing or foreign-priced tickers from a fresh read-only valuation. Unsaved target edits must be saved first; unavailable drift stays blank, and no trade is placed.
Return to the same Review context. The last tab, Inbox filter, and Month/Quarter choice are restored from local browser/app storage. Invalid values fail back to Inbox, All, and Month.
README, landing page, install examples, Review Orbit docs, and isolated fictional-demo screenshots now reflect the current app. Provider-backed market facts are current at capture time.
No schema migration. Installing over the latest published stable, v5.13.0, preserves local holdings, snapshots, backups, settings, keys, targets, and thesis text.

v5.14.0 mainline milestone — included in v5.15.0

This version was not published as a stable tag. Its dialog safeguards ship publicly as part of v5.15.0.

A dialog keeps the keyboard inside it. FolioOrb’s eight dim-the-dashboard panels had each grown their own idea of what “modal” means, and three stopped the mouse but not the Tab key. Walking the real keyboard path: 6 of 8 Tab presses left the rename dialog, 7 of 8 left “Delete portfolio?” — the first press was already outside — and 2 of 8 left “Record a sale”. Both of those are confirmations for things that cannot be undone, and the dashboard behind them stayed fully live. All eight now hold the keyboard until you answer them, and the background is out of reach for a screen reader and the mouse too. Escape still unwinds one step at a time where a dialog has steps inside it.
Closing a dialog no longer loses your place. The focus return aimed at the control you clicked, and renaming or deleting a portfolio redraws the list that held it — so both flows dropped focus to the top of the page, every time. Each dialog now falls back to a landmark that outlives its trigger, and the return is checked rather than assumed.
A dialog opened on top of another one is usable. Marking the page unreachable also froze the other dialogs’ markup, so a second dialog arrived unfocusable while the first went on answering the keyboard. Dialogs now stack, and the keyboard goes back to the one underneath on close.
A tab strip no longer swallows a keypress. Review Orbit’s parked tabs were counted as stops, putting the wrap one control late, so once per lap a Tab press went nowhere. Only visible once one implementation served all eight dialogs.
No schema migration. Installing over v5.13.0 preserves local data, backups, settings and keys.

v5.13.0 — Checked at the edges

Faded-out panels leave the keyboard path. v5.12.0 gave five header panels one closing behaviour; other surfaces were hidden a different way — faded to invisible, with clicks switched off — which stops a mouse but not the Tab key. On a freshly loaded dashboard 14 of 47 keyboard stops were controls of this kind, sitting past the last visible one. The portfolio manager, the first-run welcome panel, the collapse pill, the hidden Senpai orb, the holdings table during a scan, and the loading splash now all leave the tab order while hidden, opening and closing exactly as before.
A junk symbol is refused instead of looked up. FolioOrb’s rule for a symbol — letters, numbers, ., - or ^, up to 10 characters — was applied when you add a holding but not when one is read back. 2 of the 12 addresses that take a symbol checked it; the rest passed whatever arrived to the market-data provider and into the log. All of them now check first, including the two that take a comma-separated list. Valid symbols are unchanged.
Undoing a DCA buy can only touch its own portfolio’s holding. The undo found its holding by internal row number alone, and that number carries no portfolio with it, so it could change a holding in a different book while reporting success. It now looks the holding up inside the plan’s own portfolio.
No schema migration. Installing over v5.12.0 preserves local data, backups, settings and keys.

v5.12.0 — Calm at every size

Closed header panels leave the keyboard path. Brand, Claude-key, live-feed, overflow, and cost details now synchronize their visible, ARIA, and inert state; Escape closes the active panel and returns focus to its trigger.
Senpai stops covering phone-width portfolio truth. Below 576 px the compact orb occupies a row after the dashboard, automatic activity stays compact, and the full line expands only after an explicit tap.
Review Plan fits 320 px. Its cards shrink with the sheet while wide portfolio and target tables keep their own horizontal scrollers.
Restore cancellation is layered. When an exact restore confirmation is open, the first Escape cancels it and returns keyboard focus to Restore; a second Escape closes Review Orbit. After Queue restore is submitted, its busy state blocks misleading Escape, Cancel, close, and tab actions until the request finishes. Explicit failures safely re-enable both choices; interrupted responses are reconciled with persisted Backup Vault state and otherwise stay visibly Status unknown.
Rehearsals cannot go stale silently. Changing the holding or cash replaces the old projection with Preview outdated, and late or out-of-order responses are ignored.
Shared service/client seams now carry portfolio scoping, valuation, realized-sales math, backup handling, DCA state, local exports, and Review Orbit behavior without changing the financial or recovery contracts.
Privacy docs now distinguish credential-only Anthropic availability checks, prompted Claude payloads, and publisher/CDN news-thumbnail requests.
No schema migration. Installing over v5.11.0 preserves local data, backups, settings, and keys.

v5.11.0 — Plan & Protect

Set a target course. Store exact basis-point targets across the eligible holdings in one Portfolio, then compare them with current allocation when every required USD quote is usable. Missing and foreign-priced rows block drift rather than being treated as zero.
Rehearse an external-cash buy. Preview shares, average cost, and projected allocation for a ticker you already own. The rehearsal is read-only, buy-only, USD-only, and neither a forecast nor a recommendation.
See every book without blending performance. The all-Portfolios pulse shows known USD value and completeness per book, isolates quote failures, and keeps its aggregate explicitly partial when any book is incomplete.
Export the record, not a tax answer. The annual realized CSV uses stored sale facts and reconciles displayed rows. It does not model lots, fees, holding periods, wash sales, or tax classification.
Carry human-readable records. A manifest-backed ZIP contains CSVs for Portfolios, holdings, sales, snapshots, and the DCA ledger. It excludes settings, secrets, AI caches, backups, and SQLite, so it is sensitive but not restorable.
Know whether the safety net is fresh. Manual-backup freshness stays separate from an opt-in, once-per-local-day automatic policy. Auto retention touches only verified auto-* snapshots.
Frozen smoke checks use an isolated data root, while native record exports stage and fsync a private sibling file before an atomic destination swap.
Schema v6 adds a nullable target field; existing local data remains intact.

v5.10.2 — DCA deletion follow-up

A refused DCA delete is no longer silent. The plan dialog now explains that applied buys must be undone first so their holding changes remain traceable. If the ledger refuses deletion, FolioOrb shows that reason in-app. Once no applied buys remain, deletion removes pending and skipped entries as documented.
No schema changes or migration. Installing over v5.10.1 preserves local data and keys.

v5.10.1 — Trust and flow patch

Browsing the Backup Vault is read-only. Listing, integrity checks, holding counts, downloads, and restore selection no longer create SQLite WAL/SHM sidecars. An absent vault stays absent, and a snapshot with a non-empty sibling WAL is refused as incomplete. Live database backup counts remain WAL-aware.
DCA decisions use in-app dialogs. Apply all, undo applied, skip all, change amount, and delete plan no longer depend on native browser prompts that WKWebView may not show. The dialogs validate inline, trap focus, cancel with Escape, and return focus to the opener. Ledger math and apply/undo behavior are unchanged.
Holdings reaches its rows on a phone. The compact toolbar no longer inherits a 320 px desktop flex basis after stacking vertically.
The welcome guide opens at the top. It focuses Close rather than auto-scrolling to its footer action, while keeping its modal focus and Escape behavior.
Every Review Orbit tab is visible. Phone-width tabs wrap instead of hiding Backups beyond a scrollbar-free strip; arrow-key navigation and tab semantics remain intact.
No schema changes or migration. Installing over v5.10.0 preserves local data and keys.

v5.10.0 — Architecture pass

Holdings can only be edited through the portfolio that owns them. The edit, delete and delete-trade endpoints resolved a row by id alone and ignored the portfolio the request was scoped to, so with more than one portfolio a change could land on the wrong book — and record its realized sale in the wrong ledger. All three now check ownership and answer 404 when the holding isn’t yours.
A slow endpoint no longer freezes the dashboard. Fifty-two endpoints were declared async while doing blocking work (Yahoo, SEC EDGAR, Claude, large scans), each holding the single event loop while every other request queued behind it. They now run in the worker threadpool, and the guard that missed them sweeps every router instead of a hand-written list.
A single NaN price can no longer blank every position weight, and with it the concentration score and what Claude is told about your book.
The action plan sees your sector and country exposure again — it had been reading keys nothing produced, and sending Claude an empty block.
One market-hours answer and one concentration scale, instead of two of each that could disagree on the same page.
No schema changes and no migration.
Also ships the user-flow resilience work below, which was prepared as v5.9.3 and never cut as its own tag.

v5.9.3 — User-flow resilience (shipped in v5.10.0)

Review Orbit owns the viewport. It stays fixed above the sticky navbar on compact screens and hides fixed dashboard utilities while open.
Holding details work from the keyboard. Native disclosure controls expose their expanded state, collapsed detail content leaves the accessibility tree, and ticker lookups stay scoped to the holdings table.
News failures are retryable. A failed feed offers an announced in-place retry and does not require a page reload to recover.
Focus returns somewhere visible. Closing Manage after entering from the first-run guide falls back to the named Manage control instead of focusing the hidden guide.
Help buttons have names. Icon-only tooltip triggers—including dynamically rendered ones—derive an accessible name from their help heading.
The Action Plan fits 320 px. Its header stacks cleanly with no page-wide horizontal overflow.
No portfolio math, trade behavior, or schema changed; existing data and keys are preserved.

v5.9.2 — No duplicate requests on load

A cold load no longer fetches the same endpoint twice. It was 29 requests across 26 endpoints; it is now 26 across 26. Sparkline history and the benchmark comparison share one request between the cached-paint and live-fetch passes, and the intelligence-mode handler no longer redoes its work when the Claude heartbeat re-applies a mode that has not changed. Refreshing still refetches everything.
One Claude call per payload. The AI insights loader had two callers that could fire together on a switch into Claude mode and no in-flight guard, so a single payload could cost two billed calls. It now has the same guard its local counterpart always had.
This closes the known issue listed under v5.9.1 below. No feature or schema changes.

v5.9.1 — Userflow fixes

The welcome guide is a real modal. The first-run guide blocked the mouse but left the dashboard behind it reachable by Tab, and never moved focus into itself. It now marks the background inert, moves focus to Add your first holding, traps Tab in both directions, and returns focus to its opener — matching Review Orbit and the portfolio manager.
Senpai settles instead of covering the numbers. On desktop the quip bubble stayed pinned open over the bottom-right of the dashboard, hiding Today’s impact contributions and part of the world-markets strip. It now fades back to the orb a few seconds after it speaks, as it already did on mobile. Clicking the orb brings the last line back.
Ticker validation works again. The ticker field’s pattern was not a valid regular expression under the stricter parser current browsers use, so it was dropped entirely and logged a console error on every load. Dashes and carets (BRK-B, ^GSPC) validate as intended; invalid input no longer slips past the field check.
Add-holding errors are announced. The inline error is now a live region tied to the ticker field, so screen-reader users hear what went wrong.

Known issues

A cold load still requests three endpoints twice (benchmark-comparison, analytics-insights, and the batched price history), which lengthens first paint. Queued for the next release.

v5.9.0 — The Review Orbit

One review workspace. Inbox, Data Trust Center, monthly/quarterly Review Pack, type-aware Watchlist Compare, thesis cadence, and the Local Backup Vault live behind the new Review control.
Attention without automation. Price gaps, pending DCA buys, upcoming earnings, calibration history, and thesis reviews are prioritised locally. Nothing places a trade or changes a holding.
Coverage before confidence. Prices, quote metadata, fund fees, dividend classification, top-holdings ETF overlap, theses, and stored history disclose their coverage and sources. Missing stays missing.
Verified recovery. Manual vault snapshots contain the SQLite database only, excluding API keys and .env. Restores are verified, applied only across a clean restart, and keep a verified safety copy of the pre-restore database.
Review packs tell the truth about returns. Print-ready HTML and CSV include current valuation, stored snapshots, realized trades, P&L contributors, and thesis attention; raw value change names its stored opening snapshot and is labeled as cash-flow affected, not time-weighted performance.
Additive schema v5. Optional thesis review timestamp/cadence fields do not affect valuation or verdict math. Installing over v5.8.0 preserves all existing data and keys.

v5.8.0 — The Workspace Remembers

Theme and place persist. Light/dark mode and the last dashboard zone restore across refreshes, relaunches, and portfolio switches; the saved theme is applied before first paint.
Research means research. Research-only tickers never fill the largest-holding tile, portfolio briefing, action-plan buckets, or invested-position projections. An all-research portfolio says it has no invested positions instead of narrating imaginary P&L.
No invented $100K projection. With no positions, personalized scenario cards stay hidden and the chart shows only the S&P 500 as an index labeled 100 = start.
A proper keyboard modal. Manage Portfolio takes focus, traps Tab, makes the background inert, closes with Escape, and returns focus to the opener.
320 px without page-wide overflow. The app bar, zone tabs, menus, cards, and internally scrollable holdings table stay inside the viewport.
Faster, truthful startup. Frontend libraries are vendored, portfolio/market cache paints immediately, duplicate cache misses collapse, and cold starts use a progress-linked compounding splash only when there is genuinely nothing ready to show.
Installing over v5.7.0 keeps the database, holdings, trades, DCA history, thesis notes, settings, and API key. No schema migration is required.

v5.7.0 — Reasons and Seasons

Your thesis, on every holding. Each expanded holding gets a box for the reason you own it, in your own words. Claude never reads or writes it, it round-trips through CSV, and a background refresh can’t clobber an edit in progress.
A dividend calendar. The income card now shows which months pay you, not just the annual total — twelve bars built from each payer’s real trailing ex-dates, so a monthly REIT and a quarterly stock land in their own months rather than an assumed rhythm.
Honest about what the months mean. They’re ex-dividend months — the cutoffs to own each payer — and the card says so, since cash lands later. A payer whose schedule can’t be read from history is named as unscheduled, never smeared across months nobody observed.
Installing over v5.6.0 keeps the existing database, holdings, trades, DCA history, settings, and API key. No schema migration is required.

v5.6.0 — Inside Out

Insider activity (SEC Form 4). Recent open-market buys and sells by a company’s own officers and directors, in each stock’s expanded detail. Grants, gifts, and tax withholding are shown but never counted as conviction; funds have no insiders and say so.
Financials over time (SEC XBRL). Revenue, net margin, and diluted EPS by fiscal year from filed numbers, with revenue history stitched across the 2018 GAAP tag change. Missing metrics stay blank, never a fabricated zero.
Dividend income view. Annual cash your holdings pay you at your position size, blended yield, and an ex-dividend heads-up. Non-payers are named, never counted as $0.
Filings in plain English. 8-K item codes are labelled by what they report (“Results announced,” “Officer or director change”) instead of raw numbers.
Fixed: low-yield dividend yields read 100× too high. yfinance reports dividend yield in two fields 100× apart; a sub-1% yield rendered as tens of percent. Normalized once at the source — Apple now reads 0.32%, not 32%.
Installing over v5.5.0 keeps the existing database, holdings, trades, DCA history, settings, and API key. No schema migration is required.

v5.5.0 — Straight From the Source

The filings, unfiltered. The News tab shows what your companies actually told the SEC — 8-K, 10-Q, 10-K — pulled from EDGAR and linked to the source document. Funds and crypto are named as non-filers, never shown as companies that filed nothing. No Claude key needed.
The curve, and where fear actually sits. The market backdrop reads the US Treasury yield curve (2s10s, 3m10y) and reports where the VIX sits in its own five-year range. An inverted curve nudges verdicts toward quality; flat, normal, and steep are reported and left alone.
What your funds cost you. Expense ratios as real dollars per year and over a decade, with the growth assumption stated next to the projection. A fee that can’t be read stays unknown — never a quiet $0.
Whether your ETFs are the same bet. Pairwise overlap across your funds’ top 10 published holdings, labeled for exactly what it measures.
Earnings with the bar attached. The radar carries the consensus EPS estimate and the recent beat record alongside the date.
Fixed: fund fees read 100× too high. An expense ratio could be scored in the most expensive tier when it was among the cheapest, dragging that fund’s quality score. Ratios are now normalized once, at the source.
Installing over v5.4.2 keeps the existing database, holdings, trades, DCA history, settings, and API key. No schema migration is required.

v5.4.2 — Honest Valuation, Safer Ledgers

Return math now matches the full investment. Portfolio total-return percentage includes the cost basis of both open and already-sold shares, matching the realized + unrealized gain shown beside it.
Bad quotes never become a confident read. Zero, non-finite, malformed, or missing prices are labeled unavailable; incomplete valuations do not write daily snapshots or generate portfolio-level Claude briefings, analytics narration, or action plans.
Applied DCA buys stay traceable. Undo applied buys before deleting their plan so the contribution ledger and holding mutation cannot drift apart.
Cleaner architecture and public docs. Lifecycle, valuation, DCA, and narrative caching now have focused service interfaces, while the landing page and docs share a sharper visual system.
Installing over v5.4.1 keeps the existing database, holdings, trades, DCA history, settings, and API key. No schema migration is required.

v5.4.1 — Per-Portfolio Verdict History

Your track record follows the portfolio. The Signals “how did my past calls age?” report card and calibration stats now scope to the portfolio you’re viewing, instead of blending Add / Trim / Hold calls across all of them.
Verdicts logged before this update are kept and attributed to your default portfolio (additive schema v4 migration). Installing over any 5.4.0 keeps everything in place.

v5.4.0 — More Than One Portfolio

Multiple portfolios. A switcher in the top bar lets you create, rename, delete, and switch between portfolios — give your taxable account, IRA, and experiments their own scoreboards.
Everything re-scopes. Value, P&L, holdings, every analytics chart, news, DCA plans, and the AI briefings/action-plans all follow the portfolio you’re viewing. The Manage panel names the one you’re editing.
Cleanly separated. Each portfolio’s data — including its cached AI narratives (namespaced per portfolio) — stays its own; switching never shows another portfolio’s content. “My Portfolio” is always present and can’t be deleted; new portfolios start empty.
Installing over any 5.3.x keeps your holdings as the default portfolio.

v5.3.1 — Accurate Sales, Working Links

Record a sale at the real price and date. Reducing a holding now lets you enter the actual sale price and date (pre-filled with today’s, but editable) — so a sale you made last month books into your realized P&L and year-end recap correctly, in the right tax year. Leave the price blank to use the live market price.
External links work in the desktop app. Links like console.anthropic.com (to get a Claude key) and the docs now open in your real browser instead of a dead in-app frame.
Installing over any 5.3.0 keeps everything in place.

v5.3.0 — Honest When Things Go Wrong

Never a scary $0. If market data can’t be reached, the dashboard keeps your last-known values and shows an honest “unavailable” status instead of $0 with a green “synced” check — and it no longer writes that $0 into your performance history (which left a permanent fake cliff in the P&L/drawdown charts).
Honest about your Claude key. Saving a key now verifies it actually reaches Anthropic before claiming “connected,” you can disconnect Claude from the key panel, and the offline setup steps point at the one-click panel instead of “edit .env and restart in a terminal.”
No accidental sales from a typo. Reducing a holding’s share count now asks before booking a realized sale, and mid-typing keystrokes no longer book phantom sales.
Smaller edges filed down. First-run/empty portfolios get an “Add your first holding” prompt; remove/world-markets/news failures surface clearly instead of silently; stale hero tiles clear on an empty portfolio; and Cmd/Ctrl shortcuts no longer double-fire.
Installing over any 5.2.x keeps everything in place.

v5.2.1 — DCA, Polished

Bulk actions ask first. “Apply all” and “Skip all” now confirm with the count and dollar total, and each plan gets an Undo applied action that reverses a whole backfill in one move.
Pause means pause. Resuming a paused plan no longer retroactively books the buys skipped while it was paused — it picks up from the resume date.
Sharper edges filed down. Undoing a buy that empties a holding now retires that holding (no $0 leftover); an exact-duplicate plan is blocked; same-ticker plans are told apart in the bucket; large backfills render lightly; and double-taps can’t fire a duplicate action.
Installing over any 5.2.0 keeps all plans, holdings, and settings (additive schema v3).

v5.2.0 — Auto-Invest, On Your Terms

DCA auto-invest plans, simulated locally. The most-asked question — “can it sync my broker’s auto-invest?” — answered the local-first way. Set a ticker, a dollar amount, and a cadence (daily / weekly / monthly); each interval books a buy at that day’s real close into a review bucket. Weekends and holidays snap to the next trading day.
Nothing moves until you say so. Every booked buy waits for you to Apply it (shares and average cost update), Skip it, or leave it. Applies are reversible with Undo; skips can be Restored. A past start date backfills the full history, with a double-count guard if you already hold the ticker.
Catches up after you’ve been away. Reopen the app after a week and it fills in every missed buy, idempotently — never double-booking. A badge shows how many are waiting.
Installing over any 5.1.x keeps all holdings, settings, and .env. Still no brokerage connection; still not financial advice.

v5.1.0 — Looking Back, Honestly

Year-end realized recap. The Realized gains tab now opens with a year-by-year recap of your closed trades — realized P&L and return, sales and tickers covered, winners vs losers, and the best and worst position of the year. Reads every stored trade from your local database; no live quotes. Switch years with the toggle.
A verdict report card. The Signals tab grades FolioOrb’s own past Add / Trim / Hold calls by how each holding has done since the call — an overall “aged well” rate, a per-action breakdown, and a ledger with a ✓/✗ per call. A look-back, not a forward bet; small samples are noisy and it’s not financial advice.
Installing over any 5.0.x keeps all holdings, settings, and .env.

v5.0.0 — FolioOrb

FolioOrb is the native desktop app for local-first portfolio intelligence.

FolioOrb everywhere — app window, installers (FolioOrb-macOS-arm64-*.dmg, FolioOrb-Windows-x64-*-Setup.exe), docs, and repository at github.com/udhawan97/FolioOrb.
Data stays local in the normal per-user app data directory for your platform.
Privacy posture stays clear — FolioOrb is local-first, Claude-optional, never places trades, and has no FolioOrb telemetry or cloud account.

v4.5.2 — Reliability patch

A full-codebase bug audit and the fixes it surfaced — no new features, just sturdier numbers and safer data.

Bad market data no longer poisons your analytics. A zero or negative close (a halted or delisted ticker, a data glitch) used to contaminate annualized return, volatility, and correlation with NaN; those days are now treated as flat. Correlation reports “no data” honestly when a frozen price series makes it undefined, instead of faking a 0.0.
Trim verdicts score the right way round — a cheap stock is no longer handed a high-confidence trim.
A failed rollback can never leave you with no database. Restore now stages and re-verifies the backup before touching your live database, and tells you clearly if data restored but settings didn’t.
Watchlist edits never fabricate P&L, the batch price-history endpoint validates its period, plus assorted updater, cache, and logging hardening.

No migration or .env change required.

v4.5.1 — Export That Actually Exports

Fixed: CSV export and the import template now download in the desktop app. The packaged app is a native window with no download chrome, so Export CSV and Download template used to open the file inline as raw text — no Save dialog, no back button. Both now route through a native Save As… dialog and write a real .csv (UTF-8 with the BOM Excel expects). Browsers download exactly as before. Either way, you get a file, not a dead end.
Website polish — Senpai gets his own animated spotlight above the footer, a refreshed one-liner pill, and CSV import is now called out in the workflow walkthrough.
Installing over v4.5.0 keeps all holdings, settings, and .env. Nothing about your data changes.

v4.5.0 — The Spreadsheet Release

Export your holdings as CSV. One click in the portfolio manager downloads your active holdings and watchlist as a clean, Excel-ready CSV — with every cell escaped against spreadsheet formula injection. The file you get is the import template.
Import holdings, two ways. Local (always available, no API key) does a strict, exact-schema parse of the template — deterministic, free, offline, and never gated. Claude assist (when a key is set) maps almost any brokerage export onto the FolioOrb format for you, and every mapped row still passes the same strict validation before it touches your book. Clean template files skip Claude entirely.
A per-row report either way — added, skipped (duplicates are skipped, never overwritten), or errored with a plain-English reason — plus a Senpai-narrated recap in Claude mode. Safety as usual: a 256 KB / 200-row import cap and content-type checks.

v4.4.1 — Software Update, done right

Fixed: the updater falsely showing “You’re offline.” The packaged app’s bundled OpenSSL pointed at a build-machine certificate path that doesn’t exist on your Mac, so every update check failed TLS verification and was reported as “offline” even when you were connected. Now fixed at the root — the checker verifies against a CA bundle shipped inside the app — and failure states are told apart instead of lumped together (“Couldn’t securely check for updates,” “GitHub rate limit reached,” etc., each with its own diagnostic).
Real in-app updates on macOS, not just “open the DMG.” Update Now downloads, verifies, backs up your data, swaps in the new app, and relaunches automatically. Windows keeps its one-click silent install.
A consent-first in-app update system. FolioOrb checks quietly for new versions, shows a calm indicator when one’s available, and never downloads or installs without an explicit click. Check any time from Check for Updates… in the app menu or Settings → Software Update.
Your holdings are protected at every step. A verified backup is taken before any update or migration; if the backup can’t be made or fails verification, the update pauses rather than risking your data. Restore previous version… rolls back safely, always snapshotting current data first so nothing is lost either way.
Trustworthy downloads. SHA-256 verified against published checksums, with optional minisign authenticity signing. Release notes render as text, not raw HTML.
Built in eight phases, then hardened through repeated rounds of adversarial review — against both source and the actual installed app — that caught and fixed the offline/TLS root cause plus nine other real issues, including a data-loss gap in backup verification, an XSS path in release-notes rendering, and a macOS “quit unexpectedly” crash on every exit. 256 dedicated tests, pylint 10.00.

v4.3.4 — Quieter Hot Paths

Performance release — no feature changes. Targeted fixes to the code that runs most often, so the dashboard stays smooth as portfolios and interactions scale up. Installing over v4.3.3 or earlier keeps all holdings, settings, and .env.
Currency formatting no longer allocates per value — every dollar figure was built with a freshly constructed Intl.NumberFormat, which is far more expensive than the formatting itself. The currency and world-market formatters are now built once and reused. Output is byte-for-byte identical.
The correlation heatmap stops rebuilding its canvas on hover — the hover redraw was reassigning the canvas’s pixel dimensions every time, which reallocates and clears the whole backing store even when the size is unchanged. It now only resizes the bitmap when the dimensions actually change.
Hover and resize work is coalesced to one pass per frame — the heatmap’s mousemove handler and the two dashboard-zone indicator resize handlers are throttled to requestAnimationFrame, so at most one layout read/redraw runs per frame.
Fixed misleading “AI failed” warnings when no Claude API key is configured — running key-free (the default Local Intelligence mode) logged every briefing/insights/action-plan request as a warning-level failure, even though it’s the expected, harmless case. These now log at debug with an accurate message; a genuine failure with a key present still warns.
Fixed the greyed-out Engine toggle opening the wrong panel — tapping the disabled “Local Intel” control (no API key set) opened the passive intro card instead of the interactive Connect Claude AI panel. It now opens that panel directly — paste a key, hit Save & Connect, no restart — and closes the menu first so it isn’t covered.
Verified: the full 381-test suite passes, pylint holds at 10.00/10, and the theme, spacing, typography, and animations are untouched — these changes affect only how existing work is scheduled and routed, not what is drawn.

v4.3.3 — One Range, Every Section

Feature release. The Overview time-range switcher (Today / 1M / 3M / 6M / 1Y) gains a 1W option and now drives the sector/movers panel, the portfolio briefing, and the allocation focus read — not just the P&L card. Installing over v4.3.2 or earlier keeps all holdings, settings, and .env.
One shared range. All four sections read from a single selected time range instead of managing their own state, so switching to 3M means Insights, Briefing, Allocation, and P&L all narrate the same three months.
Fixed a real inconsistency caught during testing: the hero P&L card needed weeks of the app’s own usage history to compute longer ranges, so a newer portfolio could show “1M P&L: –” directly above a movers panel already showing a real number for that month from actual market price history. The hero card now falls back to the same price-history calculation so both agree.
Performance: all five non-day ranges come from one new endpoint in a single request, cached per holdings set so revisiting a range is instant with no network call. Rapid switching never flashes stale data, each section fails independently with its own inline retry, and manual refresh now refreshes range data too.

v4.3.2 — Scrolling, Finished

Performance release — no feature changes. Second half of the v4.3.1 scroll fix, this time aimed at portfolios with a real number of holdings in them. Installing over v4.3.1 or earlier keeps all holdings, settings, and .env.
Fixed a sparkline redraw bug — each holding’s 7-day trend canvas was repainting every time it scrolled into view, even when its price history hadn’t changed. It now only redraws when the data actually changes, verified pixel-for-pixel.
Fixed the real cost — switching away from the Holdings tab hid it with visibility: hidden, which keeps a hidden element fully “in play” for layout. With enough holdings, that table’s column-width math is expensive, and it was being recomputed on every scroll frame on every tab — not just Holdings. It’s now skipped entirely while off-screen (content-visibility: hidden) and restored instantly on switch.
Result: roughly 4× faster scrolling on Overview and Holdings with a 30-holding portfolio, with Analytics and News both meaningfully smoother too.

v4.3.1 — Smooth Scrolling in the Desktop App

Performance release — no feature changes. The native app scrolled sluggishly on macOS; v4.3.1 fixes it. Installing over v4.3.0 keeps all holdings, settings, and .env.
Fixed the universal scroll killers — the ambient background moved off the scrolling page onto a fixed, compositor-cached layer (background-attachment: fixed repainted the whole gradient every frame), and the drifting background orbs lost a heavy blur(40px) filter that was re-rasterized each frame. Both help the browser and from-source runs too.
Desktop-app rendering profile — inside the system WebView (WKWebView / WebView2) the app now drops backdrop-filter and freezes a few always-on ambient animations, which those engines render expensively. Frosted surfaces fall back to near-opaque fills, so the look holds up; the in-browser experience is left at full fidelity.
Result: ~3× smoother scroll frame rate on the overview and analytics views in throttled testing, with janky frames cut by roughly two-thirds.

v4.3 — FolioOrb Goes Desktop

One-click desktop installers — no Python, no terminal. Download a native app for macOS (Apple Silicon) or Windows (x64) and launch it like any other app. The FastAPI server runs in-process behind a native window
Automated release pipeline — every tagged release builds, smoke-tests, and publishes the .dmg and .exe to GitHub Releases with a SHA256SUMS.txt, so a broken build can never replace a good download
Rolling latest-main builds — every merge to main refreshes a prerelease with the newest installers, available under “Development builds” on the site for early testers
Download-first website — the landing page detects your OS, links straight to the current installer, and shows live release version, date, and checksums
Honest trust story — early builds aren’t code-signed yet, so the install guides walk through the expected macOS Gatekeeper / Windows SmartScreen warnings and how to verify your download
Local data stays put — the installed app keeps your database and .env in the per-user data directory (~/Library/Application Support/FolioOrb on macOS, %LOCALAPPDATA%\FolioOrb\FolioOrb on Windows), never inside the app bundle

v4.2 — Meet Senpai, and Never Get Lost on Day One

Senpai — the dashboard orb formerly known as “dashboard pet” / “Portfolio Butler” is now named Senpai everywhere: ids, classes, JS, localStorage, and the visible label
Tips & tricks — Senpai’s quote rotation now surfaces genuinely useful one-liners (Research mode, hold-type icons, keyboard shortcuts) about 1 time in 4
First-run welcome guide — a one-time modal on a fresh install with zero holdings, covering how to add a holding, what Research mode means, and what the four hold-type icons do — sourced straight from the same tooltips used elsewhere in the app
Docs site fix — 5 internal links that 404’d on GitHub Pages (missing the site’s base path) are fixed, and a Documentation link was added to the app’s nav menu

v4.1 — The Dashboard Finally Knows Its Own Key

In-dashboard API key configuration — paste a Claude key from the nav, validated and written to .env, reconnected without a restart
Live token cost tracking — the cost HUD now shows real accumulated input/output tokens and actual spend instead of a cache-based estimate
First-click holding expand — rows expand their intelligence panel on the first click, with auto-refresh keeping prices current
Overview sector graph rework — proportional weighted fills instead of absolute percentage bars, plus an overflow note when more sectors are held than the strip can show
Assorted fixes: font resize regression, a scroll-interception bug, and ticker management now triggering a fresh quote load without a full reload

See the full RELEASE_NOTES.md for prior versions and complete technical detail on each change.