# Changelog

All notable changes to the Linux client since the 1.0.0 release. Format follows
[Keep a Changelog](https://keepachangelog.com/1.1.0/); versions follow semver, and every
release is tagged `vX.Y.Z` with `Cargo.toml` and `packaging/PKGBUILD` in agreement.

`Schema:` lines record the SQLite `SCHEMA_VERSION` (`pdfs-core/src/db/migrations.rs`) a
release ships. Migrations are forward-only — a database written by a newer client is a hard
refuse-to-open, not a downgrade, so rolling back a release means restoring the cache from
scratch (user data in `staging/` and `recovery/` is never touched by this).

## [Unreleased]

## [2.1.0] — 2026-09-23

The desktop app, the tray and the search prompt speak German, British English, Spanish, French,
Italian, Dutch, Polish and Brazilian Portuguese, picked from the system language or chosen in
Preferences.

Schema: **31** (unchanged).

### Added
- **Translations.** The desktop app, the tray and the search prompt are translated into German,
  British English, Spanish, French, Italian, Dutch, Polish and Brazilian Portuguese, and follow the
  system language. Counts use each language's plural rules, and dates use its own formats. The
  desktop files carry translated descriptions and search keywords. The `pdfs` CLI and messages
  from the daemon stay in English. [docs/TRANSLATING.md](TRANSLATING.md) describes how to improve a
  translation or add a language. A few labels on the Photos page (the filter tabs, the gallery
  subtitle and the heading shown when a filtered timeline is empty) are still in English.
- **Language preference.** Preferences → General → Appearance → Language overrides the system
  language for the app, the tray and the prompt, from their next start.

### Changed
- **Packages install translations.** The Arch, AUR, Debian and Fedora packages install the compiled
  catalogs under `/usr/share/locale`, and the release tarball carries them in `locale/`. Building
  the source packages now needs `gettext`.

## [2.0.0] — 2026-09-23

A redesign of the desktop app: a new shell with a Preferences dialog, keyboard-driven context
menus, a docked details pane, a rewritten photo viewer with infinite scroll, album management,
stop sharing and link expiry, a redesigned Shared with Me page, per-device restore, a tray that
shows the sync state, and a new app icon.

Schema: **31** (unchanged).

### Added
- **Pause and resume sync.** The Sync page, the tray and `pdfs sync pause [--for 1h]` /
  `pdfs sync resume` hold back queued uploads and mirror-folder reconciles. Reads through the mount
  keep working, writes keep being accepted and staged, and a timed pause ends by itself. The pause
  survives a restart.
- **See what is waiting to upload.** The Sync page lists the queue with each item's last error and
  next retry, with Retry now per item and Retry All; `pdfs sync queue` and `pdfs sync retry [id]`
  do the same from the CLI. Nothing in the queue is ever dropped from here.
- **Resolve sync conflicts.** `(sync-conflict …)` copies under My Files are listed on the Sync page
  next to the file they conflict with, with sizes, times and whether the content differs. Keep the
  original, keep the copy or keep both under a new name; whatever is not kept goes to Trash.
  `pdfs conflicts` and `pdfs conflicts resolve <path> --keep original|copy|both` do the same from
  the CLI. Files that are open or still uploading are refused until they settle.
- **Pause one folder.** A mirrored folder on the Sync page has its own pause button; while paused
  it is not reconciled in either direction and its mode cannot be switched, and resuming syncs it
  straight away. `pdfs sync pause --folder <id>` / `pdfs sync resume --folder <id>` do the same, and
  `pdfs sync list` marks paused folders. The pause is kept across restarts. On-demand folders
  upload through the shared queue, so only the global pause holds them back.
- **Bandwidth limits.** Preferences → General → Network caps upload and download speed in MiB/s,
  shared by every transfer and applied to running ones from their next block; downloads include
  reads through the mount. `pdfs sync limit --up 2M --down 0` does the same, and `pdfs sync limit`
  on its own prints the current caps. The caps are kept in `config.json` across restarts.
- **Tray states and menu.** The tray icon shows whether the drive is synced, syncing, paused,
  offline or needs attention. It asks the panel for attention while changes keep failing, and
  its tooltip gives the same line as the menu. A left click opens the app. The menu gains
  "N issues — View", which opens the Sync page, and "Sign In…" when signed out. "Hide Tray Icon"
  is remembered across logins (`tray_hidden` in `config.json`), and Preferences → Appearance
  → Show tray icon brings the icon back. "Stop Proton Drive…" asks for confirmation in the app,
  then stops the service and the tray. `pdfs-app` accepts `--page <name>` and `--confirm-stop`.
  A second launch raises the running window instead of opening another one.
- **Stop sharing.** Shared by Me → item menu → "Stop Sharing…" and `pdfs unshare <path>` (with no
  id) remove the public link, every invitation and every member in one go. Each step is tried even
  if an earlier one fails, and a partial result names what is still shared.
- **Restore another computer's folders.** Every computer's menu on the Computers page has
  "Restore to This Computer…", which syncs that computer's backed-up folders down to paths you
  pick; they keep syncing with that computer's copy, and this machine keeps its own identity.
  This computer's own restore moved from the header into its row's menu. `pdfs sync restore
  --device <uid>` does the same from the CLI.
- **Public link expiry.** The Share dialog's create-link form has an "Expires" choice (Never,
  1 day, 7 days, 30 days), and an existing link shows when it expires or that it has expired.

### Changed
- **New app shell.** Every page has its own header bar; the sidebar ends in a footer with sync
  state, storage and the account; the Settings page became a Preferences dialog (Ctrl+,), with
  transfers on the Sync page and developer details in About. The app follows the system accent
  colour; the Proton purple accent is a preference.
- **My Files header.** New folder and both uploads sit in one **New** menu; the view button switches
  grid and list, and its menu sorts by name, size or date, reversed or not, folders first or not.
  The choice is remembered. Build Thumbnails and Open in File Manager moved to the folder menu.
- **Browsing history.** Back and forward buttons (Alt+Left/Right, mouse back/forward) replace the
  single Up button; Alt+Up goes to the parent folder. Back leaves a search first.
- **Move to…** picks the destination from a folder browser instead of a typed path, and works on a
  whole selection. Rename selects the name without its extension.
- **Drop to upload.** Files and folders dropped on My Files upload into the folder on screen. The
  upload toast links to the Sync page.
- **Photos.** The main view is called Timeline. Right-clicking a photo opens a menu to open it,
  favourite it, select it or move it to Trash; Ctrl+A selects every loaded photo. An empty filter
  says what it filtered ("No videos in June 2024") instead of the first-run hint. A banner shows
  while a Takeout import runs and links to it. Pages load 200 photos at a time and keep loading
  until the window is full.
- **Trash.** Select several items (click, Shift/Ctrl+click, Ctrl+A) to restore or permanently
  delete them together; the permanent delete still asks first, with the count.
- **Shared by Me.** A row opens the item. Copy link stays one click away; Open, Show in My Files,
  Open Link and Manage Access… are in the row's menu.
- **Computers.** Rename, Use This Computer's Identity and Remove are in one menu per computer.
  Removing a computer deletes its backups, so the dialog now asks for its name to be typed. The
  page says which computer this machine backs up as, and says "computer" throughout.
- **Sign in.** Failures read as sentences — wrong password, too many attempts, no connection, no
  keyring, a skipped two-factor or verification step — instead of the raw error. The button spins
  while signing in, and the page links to account creation and password reset. The two-factor
  dialog mentions recovery codes.
- **Real menus.** Right-click menus in My Files, Photos, Shared by Me and Computers are GTK menus:
  arrow keys, Enter and Escape work, and the Menu key or Shift+F10 opens the one for the selection
  in My Files. Items follow one order — Open, Open With…, Play · Available offline · Share…,
  Copy Link · Rename…, Move To…, Versions… · Move to Trash. Right-clicking empty space offers New
  Folder, the uploads, Select All, Refresh and Open in File Manager. Open With… asks which
  application to use.
- **One selection in My Files.** The grid and the list share their selection, so switching views
  keeps what is selected, and dragging across empty space draws a selection box in both.
- **Details pane.** On a window at least 1000 px wide the details pane docks beside the files
  instead of covering them. It has Info, Offline, Sharing and Versions sections: who has access or
  whether there is a public link, how many versions there are, and Share…, Copy Link and
  Versions… buttons.
- **Search in this folder.** Inside a folder, a folder button beside the search box limits the search
  to that folder and everything below it. `pdfs search --in <folder>` does the same from the CLI.
  In the list view a Location column shows which folder each hit is in. Pinned files listed by the
  search prompt and `pdfs-query` show whether they are on disk, instead of always showing them as
  not downloaded.
- **Album management.** The Albums view has a New Album button, and an album's right-click menu
  offers Rename… and Delete Album…. Deleting an album keeps its photos in the timeline. A photo's
  menu and the selection bar offer Add to Album…, and inside an album, Remove from Album takes a
  photo out without trashing it. Albums shared with you only open. The CLI gains `pdfs new-album`,
  `rename-album`, `delete-album`, `add-to-album` and `remove-from-album`. This needs
  proton-drive-rs 0.6.6.
- **Albums view fixes.** The timeline's filter row no longer shows over the album grid. The header
  counts albums. Back from an album returns to the grid where you left it instead of reloading it,
  and going from Albums back to Timeline keeps the timeline's scroll position unless an album was
  opened in between.
- **Photo viewer rewrite.** The lightbox's controls fade out after two seconds without pointer
  movement. Photos zoom with Ctrl+scroll, a pinch, `+`/`-`, `1` for 100 % and double-click, and
  pan by dragging; `0` fits them again. The details panel docks beside the photo instead of
  covering it. Stepping past the last loaded photo loads the next page, and the counter shows the
  library's total rather than the loaded count. A video plays inline. Save a Copy uses the photo's
  real name and reports failures, and Open With… offers the app chooser.
- **Photos scroll without a Load More button.** The timeline loads the next page while you are
  still a screen and a half from the end, with a spinner under it while a page is on its way. A
  month scrubber on the right edge of the whole timeline, marked by year, jumps to the month you
  drag it to and follows the scroll otherwise.
- **Activity log redesign.** The feed is grouped by day (Today, Yesterday, weekday, date), with
  a time column and filter chips for All, Transfers, Changes, Sharing and Problems. A conflict the
  daemon logs again on every run now shows once, with a count. A "Needs attention" section on top
  lists the conflicts still waiting for a decision. Each logged conflict has a Resolve… button that
  opens the same dialog as the Sync page, and reads as resolved once the copy is gone.
- **Share dialog.** The Invite button stays off until every address looks like an email; a
  malformed one marks the field red and names the entry in a tooltip. Buttons show a spinner
  while their request runs, so a slow invite or link removal can't be sent twice.
- **New app icon.** A white cloud with the download arrow cut out of it, on a rounded Proton
  purple tile with a darker lower edge, drawn to the GNOME app-icon grid. It stays legible down to
  16 px on light and dark panels.
- **Shared with Me redesign.** Only sections with something in them are shown, and a page
  with nothing shared says so in one place. Pending invitations sit at the top, with an Accept
  button beside each. A row reads as the name over who shared it, when, and its size; what the
  share allows is a tinted pill ("Can view", "Can edit", "Can manage") rather than a bare
  role word that looked like a button. Open, Show in My Files and Leave… are in the row's ⋮
  menu, and folders show a chevron. Inside a shared folder the header shows the path from the
  share down, with a back button. Add Bookmark is an icon in the header.
- **US spelling throughout.** The app, the CLI and the docs say "Favorite" and "color", matching the
  `pdfs favorite` command.

## [1.11.1] — 2026-09-19

Schema: **31** (unchanged).

### Fixed
- **Double-click opens folders reliably.** Selecting a file no longer slides the details pane in
  by itself: the pane moved the files under the pointer between the two clicks of a double-click,
  so the folder did not open. The pane now shows on request, from the header's Details toggle,
  Alt+Enter or the item menu's Details, and then follows the selection. Closing it keeps the
  selection.
- **The details pane opens only when asked.** On a wide window it used to show at start, empty,
  and close again whenever you entered a folder. It now follows the header toggle alone. With
  nothing selected it says so instead of closing, so opening a folder no longer moves the files.
- **Context menus no longer have wide margins under some themes.** A menu opened on a file tile
  was placed inside the tile, so a theme's rules for tile contents (Catppuccin's
  `gridview > child box { margin: 12px }`) spaced out every section.
- **My Files behaves more like a file manager.** A click on empty space between the files clears
  the selection; Ctrl or Shift held keeps it. Backspace goes back while the focus is on the files.
- **An upload parked behind a temp-file rename no longer stays parked forever.** A create for a
  name that looks like a writer's temporary file waits for the rename to the final name. If that
  rename never came — the writer crashed, the file was deleted, or the name was never temporary —
  the upload sat in the queue permanently, holding its staged bytes. The queue now sweeps parked
  creates while it is idle: one whose file is gone is dropped, and one that has been parked for
  over an hour with nobody holding it open is uploaded under the name it has. An open file is
  left alone however old it is (B94).
- **A file whose real name ends in `.tmp` uploads normally.** `.tmp` and `.temp` were treated as
  writer temp-file suffixes, so ordinary files — a Google Photos Takeout `index.tmp`, for
  instance — were parked waiting for a rename that was never going to happen (B95).
- **Staged bytes that belong to nothing are cleaned up.** A staged blob that matches no queued
  upload and names no node used to be kept forever and warned about once per blob on every mount.
  Blobs older than 30 days are now retired with their size logged; younger ones are kept, but
  reported as a single line instead of one per blob (B96).
- **Refreshing the gallery repeatedly no longer wedges the control socket.** Each refresh scope
  now runs one at a time; asking again while one is running is answered immediately rather than
  repeating work that contends with it for the database. `pdfs diagnostics` also names the scope
  of a running refresh (B97).
- **Shutting down during a photo refresh no longer logs a wall of "undecryptable photo"
  warnings.** The metadata pass stops when teardown begins instead of running into the cancelled
  runtime. Nothing was wrong with those photos; the next refresh reads them again (B97).

## [1.11.0] — 2026-09-19

Schema: **31** (V30 `photos.content_hash` / `main_uid` / `group_key`, V31 `photos.resolved_at`); SDK bumped to `proton-sdk` /
`proton-drive-rs` **0.6.5**.

### Changed
- **The gallery lays photos out in justified rows.** Every tile is now its own photo's shape:
  photos are taken in capture order until they no longer fit the target height, and the row is
  then scaled to land exactly on the right margin. Nothing is cropped and nothing is
  letterboxed, which is what the square centre-cropped tiles used to do to portraits and
  panoramas. The last row of a day stays at the target height rather than being stretched, and
  Ctrl+scroll / Ctrl+± now changes that target height. Gaps and tile corners are smaller, so the
  page reads as photos rather than as cards.
- **Thumbnails stay on screen while you scroll.** A decoded thumbnail now lands in the texture
  cache even when the tile that asked for it has already been recycled, the cache is a real LRU
  holding 1500 textures instead of 600, and the rows just past the edge of the window are
  fetched ahead of the scroll direction. Scrolling fast through a large timeline no longer
  leaves a screen of blank tiles behind.
- **Refresh on the Photos page really refreshes.** Instead of only clearing the freshness stamp
  and letting the next read serve the stale page, it starts the timeline re-read and follows it:
  the page paints from what the daemon holds, then paints again when the new timeline lands. The
  timeline is also considered stale after 60 s rather than 5 minutes. `pdfs refresh photos`
  blocks until the re-read is done, as it reads.

- **Restore puts back the whole tree.** The trash is a flat list, but you delete a shape: a
  folder, and often things you had already deleted inside it. Restoring a folder now also
  restores the items that were trashed inside it, and restoring one of those items first
  restores the trashed folders above it — otherwise the file came back into a folder that is
  still in the trash, which is to say nowhere you can see it. Parents are restored before their
  children, and each node is still restored individually, so one node the server refuses does
  not cancel the rest.

### Added
- **A RAW and its JPEG are one photo.** A camera that writes both put two tiles of the same shot
  in the timeline; now it puts one, badged `RAW`, and the lightbox has a control that switches
  between the files. The Raw tab still lists every raw file individually, because that is what
  the tab is for. Grouping follows the server's own photo relation where it has one — which
  also collapses live photos and bursts — and otherwise pairs files with the same name taken on
  the same day where one is raw and the other is not. Deleting the tile deletes the whole shot,
  and the confirmation says how many files that is.
- **Delete from the gallery.** Select photos with the Select button, with Ctrl+click or with
  Shift+click, then Move to Trash — or press Delete in the lightbox. The tiles leave the grid
  immediately with an Undo toast, and the photos land in Proton's trash, where they stay
  recoverable. A photo the server refuses to trash comes back into the grid.
- **A photo deleted on your phone disappears here.** The daemon now follows the photos volume's
  own event stream, so a photo trashed on another device leaves the gallery within about ten
  seconds instead of waiting for the timeline to go stale.

- **`pdfs diagnostics`.** Reports what the daemon is doing right now: every worker thread with the job
  it holds and for how long, queue depths per lane, the control requests still in flight, resident
  size and the pending-op count. Answers even while the daemon is wedged, which is the only time
  anyone asks (B91).
- **A stall watchdog.** The daemon warns in the journal when a worker has held one job for two
  minutes, when a control request has run that long, or when a queue is not draining — and it now
  answers the systemd watchdog, but only after a real round trip over its own control socket. A
  daemon that stops answering is restarted instead of sitting there looking healthy (B91).

### Fixed
- **A Pixel raw capture groups with its JPEG.** Pixel names the two files of one shot
  `PXL_….RAW-01.COVER.jpg` and `PXL_….RAW-02.ORIGINAL.dng`, so their stems differ and the pairing
  rule never saw them as one shot. The part-of-a-shot marker is now cut off before the names are
  compared, and the grid shows the cover JPEG once. The grouping is computed by a timeline
  refresh, so an existing library pairs up after the next one.
- **The Raw tab no longer decodes a raw it does not have to.** A raw file that was written next
  to a JPEG of the same shot is now given that JPEG's thumbnail, which the server usually already
  has. No raw download and no raw decode for those tiles.
- **Photos can be moved to Trash from the gallery again.** Photos live on the photo volume, which
  the mount never holds, so the write check found no node to consult and refused every photo with
  "trash access: not allowed". Membership of our own timeline now answers for them; an album photo
  on someone else's volume is still refused.
- **A photos refresh no longer fails with "Resource temporarily unavailable" (B92).** Re-reading a
  twelve-thousand-photo library resolves the nodes in batches of 200 and takes minutes, but the
  daemon held the control reply until it had finished — well past the client's 120 s read bound.
  The caller therefore saw `io: Resource temporarily unavailable (os error 11)` for a refresh that
  was running perfectly, and the gallery reloaded from the old timeline. The daemon now answers as
  soon as the refresh is running, and the new `PhotosRefreshStatus` request reports when it is
  done. Each refresh also logs its photo count, batch count and duration.
- **A worker no longer panics during shutdown (B91).** The FUSE worker pool is joined before the
  tokio runtime is dropped, so a worker still inside a job does not find a runtime that is going
  away underneath it.
- **A permanently rejected block upload is no longer retried (SDK 0.6.5).** The SDK now knows
  which API errors are worth replaying: 4xx is permanent, except 404, 408 and 429.
- **Long downloads no longer fail on expired block URLs (SDK 0.6.5).** A download whose block
  URLs lapse mid-transfer re-mints them once instead of failing, on the whole-file path and on
  public links, as reads already did.
- **A long rate-limit answer can no longer look like a hang (SDK 0.6.5).** A server-supplied
  `Retry-After` is now capped at 120 s.

## [1.10.2] — 2026-09-16

Fixes a daemon that could not be stopped or restarted. No schema change.
Schema: **28**; no SDK change.

### Fixed
- **A dead network mount in your home directory no longer wedges the daemon (B90).** The
  "This computer" file indexer walked `$HOME` and descended into other FUSE mounts — an sshfs
  whose server had gone away, an unresponsive rclone. Stat'ing one of those blocks
  uninterruptibly, so the indexer thread stopped responding to SIGKILL, the daemon could never
  exit, `systemctl --user stop` timed out, and the half-dead process kept the cache database lock
  that the next start needed: every restart then failed with "another Proton Drive daemon is
  already using cache.db". The indexer now skips mountpoints below your home directory by path,
  without touching them. Nothing changes about which files are indexed — those mounts were always
  excluded from the results, just not before they had been stat'ed.

## [1.10.1] — 2026-09-15

Dependency upgrade release, clearing both open Dependabot advisories. No schema change.
Schema: **28**; SDK bumped to `proton-sdk` / `proton-drive-rs` **0.6.4**.

### Security
- **`lru` IterMut Stacked Borrows advisory.** The SDK now depends on `lru` 0.18.
- **`atty` unaligned read advisory.** The tray moved to `ksni` 0.3, which no longer pulls
  `dbus-codegen` → `clap` 2 → `atty`.

### Changed
- **Keyring backend is `keyring-core` + `dbus-secret-service-keyring-store`** (keyring 4's split).
  Same libdbus Secret Service as before; stored sessions written by earlier releases are still
  found, so no re-login is needed.
- **The tray waits for a late StatusNotifierWatcher** instead of failing to start when it beats
  the desktop's tray host onto the session bus.
- Upgraded `gtk4` 0.11, `libadwaita` 0.9, `webkit6` 0.6, `rusqlite` 0.40 (bundled SQLite),
  `fuser` 0.18, `directories` 6, `sha1`/`sha2` 0.11, `base64` 0.23.

## [1.10.0] — 2026-09-15

One-request move-and-rename in the drain, and who shared an item with you. No schema change.
Schema: **28**; SDK bumped to `proton-sdk` / `proton-drive-rs` **0.6.3**.

### Added
- **Shared with me says who shared an item, and when.** The GUI row subtitle and `pdfs
  shared-with-me` show the inviter and invite date for a share root. The invitation's signature is
  checked against the claimed inviter's keys; when it does not verify, the GUI shows a warning
  and the CLI prints one, since the sender may be forged. `DirEntry` gains `shared_by`,
  `shared_at` and `shared_by_unverified` (defaulted, so older front-ends and daemons interoperate).

### Changed
- **A queued move-and-rename lands as one request.** `drain_rename` used to rename in the source
  folder and then move, with a rename-out-of-the-way step when the destination held the name. It
  now calls `move-multiple` with a target name, so there is no half-applied state between the two
  halves and no second call for the encrypted-name requirements to go stale against (B46). A
  rename that stays in the same folder still uses the rename endpoint.

### Fixed
- **Moving a node with no signature email no longer fails.** Such nodes (created anonymously, e.g.
  through a public link) need the passphrase signed by the mover; SDK 0.6.3 sends that signature.
- **A node trashed through the newer API field is recognised as trashed.** The SDK now reads
  `TrashTime` as well as the legacy `Trashed` field, so the drain no longer tries to rename a node
  that is already in the trash.

## [1.9.3] — 2026-09-01

The daemon's wait-for-login loop, and the latency that fixing it introduced. No schema change.
Schema: **28**; SDK unchanged at `proton-sdk` / `proton-drive-rs` **0.6.2**.

### Fixed
- **An enabled service with no saved session no longer polls the keyring 28,800 times a day.**
  `wait_for_session` slept a flat 3 s between checks, and a check is not just a sleep — `auth::load`
  builds a fresh `keyring::Entry`, so every iteration opened a new D-Bus connection to the Secret
  Service and negotiated a session on it. The wait now backs off, and "not logged in" is said once
  at `info` instead of on every pass, which also stops the journal filling with tens of thousands
  of identical lines (`#21`).
- **Signing in wakes the parked daemon instead of leaving the mount up to an interval late.** The
  backoff had no wake-up path: a unit started at boot reaches the ceiling within ~90 s, and
  `enable --now` is a no-op on an already-active unit, so the GUI login and the tray's Connect left
  it sleeping. `service::enable_start` now runs `enable` + `restart`, which interrupts the sleep and
  still starts a unit that is down, and `pdfs login` restarts the unit when it is already active —
  deliberately without touching `enable`, so a headless setup driving `pdfs daemon` by hand is left
  alone. The ceiling also drops from 60 s to 15 s, still an 80% cut in keyring/D-Bus wakeups
  (28,800/day → 5,760/day), bounding the worst case for paths that never reach systemd.

### Changed
- The wait is extracted into `wait_for_session()`, taking the load and sleep calls as parameters,
  so the schedule is testable without a keyring and without real time passing.

## [1.9.2] — 2026-08-28

Repair for photos a Takeout import filed under the day they were imported, an `fzf` launcher front
end, and a gallery that survives 1,600 photos on one date. No schema change. Schema: **28**; SDK
`proton-sdk` / `proton-drive-rs` **0.6.2**.

### Added
- **`pdfs redate-photos`: repair photos filed under the day they were imported.** A Takeout whose
  metadata sidecars did not match imports with no capture time at all, and the SDK then seals
  "now" into every revision — a decade of photos lands on one date. The date is still in the file
  name, so this finds photos whose name disagrees with their stored capture time by more than a
  day and rewrites them: download, re-upload at the right time, carry favourites and album
  memberships across, then **trash** (never delete) the original — Proton has no API for editing a
  sealed capture time. `--dry-run` reports without touching anything, `--from`/`--to` aim the run
  at the days an import landed on, and `--limit` caps it. Followed with `pdfs redate-status`,
  stopped with `pdfs cancel-redate`.

- **`pdfs-prompt --fzf`: launcher search that updates as you type.** A dmenu-style launcher
  (fuzzel, rofi, wofi, …) is handed one static list and filters it itself, which is why `--dmenu`
  has to make searching two steps — type, Enter, *then* results. `fzf` can re-run a command on
  every keystroke, so the new mode re-queries the daemon per keystroke (debounced to match the GTK
  prompt) and keeps `--disabled` on so the daemon's ranking, not fzf's matcher, decides the order.
  Abbreviation and typo tolerance therefore work here exactly as they do in the built-in window.
- Started without a tty — from a WM keybinding — `--fzf` spawns a terminal for itself (foot,
  ghostty, kitty, alacritty, wezterm or xterm, whichever is installed), each with the app-id/class
  `pdfs-prompt` so a single window rule floats it. Override with `"prompt": { "terminal": [...] }`,
  where a `{cmd}` token places the command and its absence appends it. `"prompt": { "mode": "fzf" }`
  makes it the default front end for an existing keybinding.

### Fixed
- **Google Photos Takeout: sidecars with a localized suffix are matched again.** Google localizes
  the `.supplemental-metadata` it appends to a photo's JSON sidecar — a German export writes
  `.ergänzende-Metadaten` — and the scanner only knew the English spelling. No sidecar matched, so
  no photo had a capture time, so every photo in the export was stamped with the moment it was
  uploaded. Matching now hangs on the photo name in front of the suffix rather than on the suffix
  itself, which is locale-agnostic, and covers the misspelt variants Google has also shipped.
- **A photo with no usable sidecar now takes its date from its file name.** `PXL_20260818_171030868.jpg`,
  `IMG_20230219_171030.jpg`, `Screenshot_20230101-102950.png`, `IMG-20230219-WA0001.jpg` and
  `Screenshot from 2023-01-01 10-29-50.png` all carry the date the photo was taken. Read as UTC, so
  it can be hours out — but never the years out that "the moment it was imported" is. A sidecar
  still wins.

- **An opened file or folder no longer dies with the window that opened it.** Everything launched
  through `open_with` was started in the caller's process group with the caller's stdio, which was
  invisible from a resident GUI but fatal from `--fzf`: the prompt exits the moment you choose
  something, its terminal closes, and the pty hangup killed the application it had just started —
  so an `@dir` rule like `alacritty --working-directory {}` opened nothing at all. Opened children
  now get their own process group and no inherited stdio. Affects every front end; only `--fzf`
  could observe the bug.

### Changed
- **The gallery virtualises at the row, not the day.** The Photos ListView had one item per day,
  and a bind built every one of that day's tiles — fine for a normal day, ruinous for the ~1,600
  photos a Takeout import drops onto a single date, which was 1,600 widgets constructed on the main
  thread before the section could be shown. Days are now flattened into heading rows and grid rows,
  so a bind is a handful of widgets whatever the day holds. A resize or a zoom step rebuilds the row
  model (the column count decides which photos share a row) and scrolls back to the photo that was
  at the top.
- Updated to `proton-sdk` / `proton-drive-rs` 0.6.2 (photos save-to-timeline, `enumerate_events`).

- The daemon side of both launcher front ends — search, pins, mountpoint lookup and the
  open-a-result policy — moved out of `pdfs-gui/src/dmenu.rs` into a shared `pdfs-gui/src/query.rs`,
  so `--dmenu` and `--fzf` cannot drift on what a result set is or what "open" means. No behaviour
  change to `--dmenu`.

## [1.9.1] — 2026-08-21

Thumbnails everywhere the GUI lists files, a status bar under the Files browser, and the
repository's contribution/branding files. No schema change. Schema: **28**; SDK unchanged at
`proton-sdk` / `proton-drive-rs` **0.6.1**.

`exiftool` (`libimage-exiftool-perl` / `perl-image-exiftool` / `perl-Image-ExifTool`) is a new
runtime dependency, declared in the `.deb`, `.rpm` and `PKGBUILD` packaging. It is only used to
pull the embedded preview out of a camera RAW; without it every other image format still gets a
thumbnail.

### Added
- **Thumbnails for ordinary Drive files**, not just the Photos timeline — painted in Files (both
  the grid and the Name column of the list), Shared, Shared by me, and Trash. Standard raster
  formats (`bmp`, `gif`, `jpeg`, `png`, `tiff`, `webp`) are decoded in the daemon; camera RAWs
  (`arw`, `cr2`, `cr3`, `dng`, `nef`, `orf`, `raf`, `rw2`, and the rest of the concrete list) are
  served from their embedded preview via `exiftool`. Previews are cached beside the existing photo
  thumbnails, outside the cache-budget scan, and tagged with the node's modification time — so a
  new revision of an image can never be shown with its predecessor's thumbnail.
- Thumbnails for visible rows are requested opportunistically as tiles scroll in, and the daemon
  drops work for listings the user has already navigated away from. The listing generation is
  issued by the daemon (`ReserveFileThumbGeneration`), not seeded from the front-end's clock.
- **Build a whole folder tree ahead of time** from the Files toolbar: one recursive local-thumbnail
  job per daemon, with live scan/found/completed/failed progress and a Cancel that appears only
  once the daemon has acknowledged the start. The job outlives navigation — it is deliberate work,
  separate from the opportunistic per-tile requests.
- **A Dolphin-style status bar under the Files browser**: the current listing's counts, a Zoom
  slider for the icon grid, and account storage usage. Zoom resizes the realised grid cells rather
  than rebuilding the model, so dragging the slider does not tear down the selection; the default
  size reproduces the previous fixed 72 px tiles. Capacity stays out of the bar until a real
  reading lands, and the quota fetch (60 s TTL) now also runs while Files is on screen instead of
  only in Settings.
- Repository housekeeping: `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md`, `SECURITY.md`, issue and pull
  request templates, a project logo, and workspace crate metadata (`repository`, `homepage`,
  `readme`, `keywords`, `categories`).

### Fixed
- A remote thumbnail miss no longer suppresses local generation. The two negative results are now
  held in separate bounded caches: "Proton has no thumbnail for this node" and "this exact revision
  was decoded locally and could not be read" mean different things, and only the second is a
  permanent verdict for that revision.
- A RAW whose preview could not be extracted because of the *environment* — `exiftool` missing, a
  failed download — is retried instead of being remembered as undecodable. Only bytes conclusively
  outside the supported formats are cached as a negative.
- A recursive build that finds a uid an opportunistic tile job already owns waits briefly for it
  rather than either duplicating the work or blocking on it forever, and a cancellation racing a
  waiter's registration is caught by a backstop instead of leaving the waiter parked.
- Starting a build for a root that is already building attaches to that job; starting one for a
  *different* root is refused explicitly rather than silently reporting the running job's progress
  as the caller's.
- Decrypted camera RAWs are staged for `exiftool` in an owner-only (`0700`) directory of their own,
  removed as soon as the preview is out and swept at cache open after an unclean shutdown.
- The RAW extension list had `rw` (not a format) and was missing `raw`; both classification helpers
  now read from one shared table instead of two hand-maintained `matches!` arms.
- The tray and desktop-entry icons were wrong/missing; `.desktop` files and the packaged SVG were
  corrected.

## [1.9.0] — 2026-08-18

Google Photos import, and a pass over the desktop app's interaction gaps. No schema change.
Schema: **28**. Builds against the published `proton-sdk` / `proton-drive-rs` **0.6.1** — the album
writes this release needs are in the registry crates, so the local `[patch.crates-io]` shim the
feature was developed against is gone and no local path dependency remains.

### Added
- **Import a Google Photos Takeout export into Proton Photos** — `pdfs import-google-photos
  takeout-*.zip` (add `--dry-run` to see what it would do, `--wait` to follow it), or the import
  button in the Photos header, which multi-selects the export's zips. Albums in the export are
  recreated as Proton albums and their photos filed into them; year folders and `Archive/` import
  into the timeline; `Trash/` is skipped. Capture times and favourites come from Google's per-photo
  JSON sidecars, including the truncated and `(N)`-swapped sidecar names Google produces.
- Photos already on the account are **not re-uploaded**: the import asks the photos volume which of
  the export's names exist (batched, 150 per request), reads and SHA-1s only those files, and skips
  the ones whose content digest matches too — Proton's own duplicate rule. A library already partly
  synced by Proton Photos on a phone therefore imports only what is missing, and an interrupted
  import resumes by being run again. Google keeps a full copy of a photo in every album folder it
  appears in; those fold into one upload that joins several albums.
- The import streams each photo straight out of its zip, so an export needs no scratch space to
  unpack into, and reports as a job in `pdfs transfers` and the GUI's activity view. `pdfs
  import-status` shows the counts, `pdfs cancel-import` stops it — a cancelled run still files what
  it uploaded into its albums.
- **An Import page in the GUI** (Settings → *Import from Google Photos*, or the import button in
  the Photos header). Archives are staged by drag-and-drop onto the page or through the file
  chooser, listed with their sizes so the set can be checked before anything is sent, and launched
  as either a Scan (dry run) or a confirmed Import. The page shows live progress off the import's
  job, keeps the finished run's report on screen, and can stop a run in flight; because the import
  outlives the page, it is polled wherever the user navigates and its completion is reported as a
  desktop notification.

- SDK (`proton-sdk-rs` **0.6.1**): album **writes** — `ProtonPhotosClient::create_album` and
  `add_photos_to_album`, ported from the TypeScript SDK (upstream C# has no album write API), plus
  batch duplicate detection (`find_duplicates_many`, `name_collisions`). The workspace deps moved
  `"0.6.0"` → `"0.6.1"` and `Cargo.lock` re-resolved to the registry crates. Adding a photo from an
  album *shared with us* (a different volume) is still unported and fails that photo's outcome.
- **Multi-select in Files.** Both views select a batch (click-drag, Ctrl-click, Shift-click); a bar
  slides in reporting the count and offering Move to Trash and offline on/off for the whole
  selection. Delete acts on everything selected, Esc clears it, and right-clicking inside a
  selection offers the batch actions rather than the one row under the pointer. Actions that need a
  single subject — rename, move, share, revisions — stay single-entry, and the details pane steps
  aside for a batch instead of speaking for it.
- **Undo on trash.** "Moved *X* to Trash" now carries an Undo that restores the batch, single files
  included.
- **The empty states are a place to act**: an empty folder offers Upload files / New folder, and an
  empty photo timeline offers Upload photos / Import from Google Photos, instead of describing what
  the user could go and do elsewhere.
- **Ctrl+F works from any page**, navigating to Files and focusing the search box, rather than being
  silently inert everywhere but Files.
- Staged Takeout archives survive closing the window; archives that have since been moved or deleted
  are dropped from the restored set.

### Fixed
- Settings showed the local cache as "X of 0 B used" when no cache budget was set. An unset budget
  means *unlimited*, so it now reads "X cached — no limit set" rather than implying a zero cap.
- The cache-budget editor fired a `SetCacheBudget` round-trip — and a toast — on every step, so a
  run of `+` clicks applied a series of caps the user never asked for and stacked a column of
  toasts. It now waits for the value to settle.
- A long pin list pushed everything below it off the end of Settings. It shows the first six with a
  "Show all N" row for the rest.

## [1.8.2] — 2026-08-17

Closes the two findings 1.8.1's live verification produced (B85, B86) and finishes the data-safety
audit entries that had been sitting as "fixed in code, verification pending" since 1.0 (B49–B55).
No schema change. Schema: **28**.

The read path no longer assumes a revision is a series of 4 MiB blocks; it takes the geometry from
the revision. The B85 investigation also **contradicted its own bug report** — see Notes, because
the conclusion recorded there was that reads of such files were losing data, and they were not.

### Fixed
- A folder switched from mirror to on-demand now shows the files the mirror pass itself uploaded.
  The sync engine created nodes on the remote without invalidating the parent's cached child
  listing, and that listing is a database flag rather than an in-memory cache — so it survived a
  daemon restart, and the files stayed invisible until the folder was switched back. Every remote
  mutation the engine makes (folder create, file upload, and the three trash paths) now drops the
  parent's cached listing by uid, which reaches a folder that no mount has resident — the case a
  mirror folder always is (`docs/BUGS.md` B86).
- `pdfs ls` and `pdfs refresh` accept a path under any mount the daemon owns, not only the primary
  one. Both previously answered "is not under the mountpoint" for anything inside a secondary
  on-demand mount, which left a folder serving a stale listing with no manual way to re-enumerate
  it. A path under nothing this daemon has mounted still gets the same error as before.
- A file whose blocks are not 4 MiB is no longer read twice over. Block offsets came from
  `index * 4 MiB`, so on a revision laid out differently every block the client asked for
  straddled two of the revision's own, and each was fetched and decrypted to answer for one — then
  cached at boundaries the next read would not ask about. Offsets now come from the revision's
  recorded block sizes, and the on-disk cache, the in-memory ring and the in-flight fetch key
  address a block by the byte range it covers rather than by its index (`docs/BUGS.md` B85).

### Changed
- The block cache records each block's plaintext start alongside its `(mtime, size)` tag, and each
  revision's block table in a small sidecar beside its blocks. Blocks cached by an earlier release
  carry no start; they are honoured wherever the old 4 MiB assumption and the new plan agree on
  the offset, which is every file this client uploaded — so the existing cache survives the
  upgrade intact and is quietly refetched only for the files it was placing wrongly.

### Notes
- **B85's diagnosis was wrong, and the entry has been corrected rather than closed.** It read a
  set of journal lines as proof that this client was requesting wrong byte ranges and being saved
  only by B84's repair path. `RevisionReader::read_at` does not work that way: it plans over the
  revision's real block sizes and clamps every block to the range asked for, identically in the
  0.6.0 the client builds against and in every earlier version. A read can therefore never return
  more bytes than it requested — yet every quoted line shows exactly that (`blen=4194304
  got=4939506`), and the one place in the client that logs that shape can only reach it on the
  opposite condition. The numbers were either transcribed with two fields swapped or captured from
  a build not in the history; they want re-capturing from the journal before anything is concluded
  from them. What was real is the assumption itself, and that is what this release removes — a
  cost fix, not a correctness one. Nothing here changes what a read returns.
- The first read of a file still plans on the 4 MiB assumption, and learns the real geometry as a
  side effect of opening the reader it needs anyway. Resolving a reader merely to *plan* would
  reintroduce the per-read key derivation the block cache exists to avoid, so the first read of a
  non-4-MiB file pays one straddling fetch and then self-corrects.
- Seven data-safety audit entries (B49–B55) moved from "fixed in code, verification pending" to
  verified. The fixes were all genuinely present and needed no change; what they lacked was the
  injected-failure testing each entry asked for, and that is what was written — unwritable
  directories and `chmod 000` subtrees for permission failures, a real `/dev/shm`↔`/tmp` device
  boundary for the cross-filesystem fallbacks, SQLite's `max_page_count` for a full disk, an abort
  trigger for a failed insert, and truncated and unpublished sidecars for the crash boundaries.
  Two seams were added to make failure arrangeable at all: `stage_write_at`, mirroring the
  `preserve_write_at` that already existed, and `walk_local_tree` lifted out of the sync engine so
  the local scan can be driven against a real tree without a daemon.
- **Two of those entries are not fully closed, and say so.** B51 and B52 concern crash consistency,
  and nothing short of a machine actually losing power tests that; the crash *boundaries* are
  covered by reconstructing on disk exactly what a crash at each point leaves. B54 also diverges
  from its required fix on purpose: it has no "deliberate delete" confirmation path, so a total
  local wipe is refused outright rather than offered for confirmation. That is stricter than asked
  for, and the cost is that a user who really did delete everything must remove and re-add the
  sync folder.
- B86's fix routes only `ls` and `refresh` to the owning mount. The other path-addressed control
  requests — pin, rename, move, delete, sharing — still resolve against the primary mount only.
  That is a larger change and was not what blocked the repro; it is recorded as an open follow-up
  (`docs/BUGS.md` B88).
- Three findings were opened rather than fixed: the geometry above is learned one read late, so a
  non-4-MiB file pays a single wasted fetch on its first read (B87); the remaining path-addressed
  control requests are still primary-only (B88); and the container-level `CLAUDE.md` still
  describes the SDK dependency as 0.1.11 when 1.7.1 bumped it to 0.6.0, which is what sent the
  B85 investigation to the wrong source tree to begin with (B89).

## [1.8.1] — 2026-08-16

Some files read short through the mount and reported success — an application got a truncated
copy with nothing anywhere saying so (`docs/BUGS.md` B84). Found because GNOME's file indexer
kept complaining that PNGs under `~/ProtonDrive/Pictures` were corrupt; it was right. The rest
of this release is the long-standing architectural backlog from the July concurrency audits
(B25, B32, B39, B40, B43, B44), each verified against the current code before being fixed.
Schema: **28** — migration V28 adds `sync_entry.local_mtime_ns`.

### Fixed
- A read no longer trusts a revision whose recorded block sizes disagree with the file's size.
  Such a file leaves the block/range path entirely: it is fetched through the whole-file,
  manifest-verified download — the same path `pdfs pin` uses, and the reason pinning was the
  workaround — and served from the resulting cached blob. One download per file however many
  reads discover the problem at once, and files above 512 MiB fail `EIO` with a note to pin
  them rather than pulling down gigabytes for one read.
- A block that comes back shorter than the file's size implies is never served, cached or kept
  in memory. Assembling a read out of short blocks is what quietly produced the truncation:
  the missing bytes were simply not appended, the kernel handed userspace a short read, and
  every tool reads a short read as end-of-file.
- A synced folder no longer reports the same deletion as an error on every pass. Deleting a
  local file that something else had already deleted counted as a failure, which kept its
  baseline and made the next pass try again — forever. A path that is already gone is the
  state the sync pass wanted (`docs/BUGS.md` B42).
- A mirror folder no longer misses an edit made in the same second as its last sync. Change
  detection compared whole seconds, so rewriting a file in place without changing its length
  read as "unchanged" and never uploaded — the remote silently stayed behind until something
  else about the file moved. Baselines now record the modification time to nanoseconds
  (`docs/BUGS.md` B25).
- A sync download no longer overwrites an edit that finished while it was running. The
  destination is re-checked in the last moment before the download is published, and a local
  file that has moved on since the pass planned for it is kept as a conflict copy — the same
  resolution a both-sides-changed conflict gets (`docs/BUGS.md` B39).
- An upload interrupted by a writer no longer reports itself as synchronized. The baseline now
  records the content that actually went up rather than whatever is on disk when the transfer
  ends, so a revision torn by a concurrent write is replaced by a clean one on the next pass
  instead of standing as the file's content on Drive (`docs/BUGS.md` B40).
- Listing a directory that changes while it is being read no longer skips or repeats entries.
  `readdir` resumed by index into a listing rebuilt on every call, so a create or a trash
  between two pages shifted everything after it. The listing is now frozen for the life of the
  directory handle (`docs/BUGS.md` B43).
- Opening a file for writing while a previous close is still publishing no longer discards that
  close's revision. The base is re-checked at the moment the write handle is installed, rather
  than trusted across a copy that can be the size of the whole file (`docs/BUGS.md` B32).
- Unmounting now stops and joins the daemon's background workers instead of leaving them
  running. The drain, sync engine, conflict sweep, online probe, local indexer and control
  listener share a stop signal and interruptible waits, so an in-process remount no longer
  starts a second set on top of the first (`docs/BUGS.md` B44).

### Notes
- The size `stat` reports was always the correct one; it is the reader's belief about where
  blocks end that was wrong, so nothing about reported sizes changes.
- Verified live on the affected account: four multi-block PNGs that had been reading short now
  deliver their full size, and the journal shows each one repaired. The diagnostics also
  answered the open question, and not as expected — the revision metadata is fine; those files
  simply do not have 4 MiB blocks, and this client assumes they do. Filed as `docs/BUGS.md`
  B85. Until that is fixed, a file with non-4-MiB blocks is downloaded whole on its first read.
- Migration V28's new column is nullable rather than converted from the existing seconds. A
  baseline row written before this release has no sub-second time to recover, and inventing
  zero for it would make every already-synced file compare as changed and re-upload an entire
  mirror on first start; `NULL` keeps those rows comparing exactly as they did, and each gains
  a real value the next time its path syncs.
- Three audit entries closed without a code change, having been fixed already or being correct
  as they stand: sync-folder removal is serialized against reconciliation (B41), the drain's
  database contention was removed by the read-only connection pool and the claim column (B30),
  and the refusals for symlinks, hard links, device nodes and `chmod` are deliberate — Drive
  cannot represent them, and a local-only representation no other Proton client can read would
  be worse than a clean refusal (B31).
- Six of this release's fixes were then driven live against the installed 1.8.1 on a real
  account: the same-second edit uploads (B25), three rapid rewrites of a 64 MiB file publish the
  third one byte for byte (B32), a write racing a download survives as a conflict copy (B39), an
  upload torn by a concurrent write heals itself on the following pass (B40), a `getdents64`
  walk with a 160-byte buffer against a directory mutated between pages returns every entry
  exactly once (B43), and a restart of a daemon with six mounts and a busy drain stops in 110 ms
  and comes back at the same thread count (B44). B42 could not be driven from a single machine —
  see B86 below.
- One new bug came out of the verification: an on-demand mount serves a folder listing that
  predates the sync engine's own uploads, so files the mirror pass created are invisible until
  the folder is switched back (`docs/BUGS.md` B86). Nothing is lost — they are on the server and
  come back on the next mirror pass — but they cannot be seen or named while the folder is
  on-demand, which is also what blocks B42's live verification.

## [1.8.0] — 2026-08-16

A queued write whose node had left the local tree was retried every five seconds forever and
counted as an ordinary pending upload — 28 of them, holding 18.4 GiB, had been doing that for
30 days on the machine that found it (`docs/BUGS.md` B83). Schema: 27 — migration V27 adds
`pending_op.access_deferred_since`.

### Fixed
- The drain no longer reads "I have never heard of this node" as "you may not write this
  node". A missing authority is now resolved against the remote: the node is re-interned if
  it is still there, and reported as a real failure if it is gone. Ops that had been stuck
  invisibly for weeks now say what is wrong with them on the first pass.
- A create or mkdir resolves its *parent* as the authority, so the refetch asks about the
  folder rather than the node it has not made yet.

### Added
- Access deferrals are bounded. A deferral still costs no attempt and records no error for
  the first five minutes — the case it exists for is a share downgraded mid-queue and undone
  a moment later — but past that window each recheck records a failure, so the op enters the
  ordinary backoff and appears in `failing_ops` and the `pdfs status` error text like any
  other stuck operation. The first deferral of a run is logged at `warn` rather than `debug`.

### Notes
- No staged data is touched by any of this. A reported failure keeps its `pending_op` row and
  its staged blob exactly as before; the change is in what the user is told, and in the queue
  no longer waiting on a permission change that was never coming.

## [1.7.1] — 2026-08-16

Write-back throughput follow-up to 1.7.0, same audit
(`docs/plans/audit-2026-08-15-perf-locks.md`). Schema: 26 — migration V26 adds
`pending_op.claimed_at`.

### Added
- Parallel write-back drain: three workers share the queue through a `pending_op.claimed_at`
  claim column (migration V26), so a large upload no longer holds every queued rename, trash
  and small write behind it. Ordering is preserved per node — the claim query never offers
  two workers ops for the same uid — and claims left by a crashed run are cleared at open.
- An upload already on the wire is cancelled when a newer write, a trash or a discard
  supersedes it, instead of spending the uplink on a revision the next operation replaces.

### Changed
- The revision debounce is adaptive: it widens toward how long that node's last upload
  actually took (bounded to 2–60 s), so a file saved faster than it can be sent supersedes
  in the queue rather than mid-upload.
- Workspace deps bumped to `proton-sdk` / `proton-drive-rs` 0.6.0.

## [1.7.0] — 2026-08-16

Performance and robustness pass over the mount, from the architecture audit in
`docs/plans/audit-2026-08-15-perf-locks.md`. Schema: 25.

### Added
- Sequential-read prefetch (`Core::prefetch_ahead`): per-file sequential detection, depth
  doubling 2→8, reset to 0 after a seek, skipping blocks already cached, and bounded by its
  own permit budget so it can never queue ahead of a demand read.
- Single-flight block fetching (`BlockFlight`) keyed on `(uid, mtime, size, block)`, so a
  demand read, kernel read-ahead and prefetch of the same block share one download+decrypt.
- In-memory block LRU (`BlockRing`) in front of *both* block paths; a block read off disk is
  promoted into it, so the following ~31 kernel reads of that 4 MiB block cost nothing.
- Read-only SQLite connection pool (`ReadPool` / `Db::read`) with every `SELECT`-only method
  routed at it, so a read no longer waits on the write connection.
- `nodes.path` persisted and backfilled, replacing the recursive `path_of` walk.
- Single-writer `flock` on the database; a `SQLITE_CORRUPT`/`NOTADB` file is quarantined as
  `cache.db.corrupt-<ts>` (WAL sidecars too) and rebuilt rather than failing the mount.
- `VACUUM` and the deep `integrity_check` now ack immediately and run as background jobs;
  `pdfs doctor` polls for the verdict.
- Staging reconcile at mount: staged blobs no queued operation refers to are re-queued, or
  named in the log when they cannot be addressed.
- Queue/staging visibility — `Status` gained `parked_uploads`, `failing_ops`,
  `failing_error`, `staged_bytes`, `staged_oldest_secs`; `pdfs status` prints them when
  non-zero.
- `statfs` implemented against the account quota; FUSE `init` negotiates `max_readahead`,
  `max_background`, `PARALLEL_DIROPS` and `CACHE_SYMLINKS`; `blksize` raised to 1 MiB.

### Changed
- `State` no longer writes to SQLite under the inode lock — mutations queue into an outbox
  applied after the lock is released, so interning a 5000-child listing no longer commits a
  transaction with the whole mount frozen.
- Re-listing an unchanged folder no longer touches its subtree; when a subtree really moved,
  one recursive walk carries both the new path and the inherited trashed flag.
- `release(2)` no longer gap-fills from the network: it stages an incomplete blob from local
  caches only and the drain thread finishes it off the dispatch loop.
- A partial write over an undrained one is merged instead of refused.
- Size-upgrade waiters park on a queue instead of occupying a worker thread, capped at 8
  folders in flight; past the cap `stat` answers provisionally.
- Block-cache disk writes moved off the read's critical path; no more fsync per cached block.
- Cache LRU touches buffered in memory and flushed in one transaction.
- Search: sync-folder list and pin lookups hoisted out of the per-hit loop.
- SQLite tuned (`cache_size`, `mmap_size`, `temp_store=MEMORY`, `wal_autocheckpoint`) and
  three new indexes added.
- Mount registry lock released before FUSE teardown/join; kernel notifications batched under
  the state lock and flushed after it.

### Fixed
- **Data loss:** `ENOSPC` is now propagated from `write`/`fsync`/staging instead of being
  swallowed, with emergency eviction wired up and scratch marked durable before any staging
  error; a failed `record_pending_write` fails `release` rather than dropping the write.
- **Data loss:** the conflict sweep no longer discards queued operations after the trash
  round trip.
- **Data loss:** `rescue_scratch` moves a blob with an unreadable sidecar to `recovery/`
  unaccompanied instead of fabricating a sidecar for it.
- Own-sealed revisions are persisted (7-day TTL), so a restart no longer reopens the
  transient-download fork window fixed in 1.1.0.
- Depth-capped `is_pinned` (cycle-safe), daemon write timeout, and every client-supplied
  `limit` clamped.
- `recover_fsynced_writes` no longer scans the recovery directory on every idle pass.

## [1.6.0] — 2026-08-12

Search-index correctness and ranking (`docs/plans/search-index.md`). Schema: 21 —
migration V21 rebuilds `nodes_fts`.

### Fixed
- **B80** — 29% of the account was missing from the search index. `node_is_indexable_tx`
  required a null-parent ancestor, which only the My Files root has, so every node under a
  device folder was dropped from `nodes_fts` on its first upsert after the v16 backfill
  (2,705 of 9,252 nodes on the account this was diagnosed against). The rule now matches the
  backfill, migration V21 replays it, nodes under a trashed folder stay out, and the walk is
  depth-capped.
- **B81** — the prompt labelled every Drive hit "My files", including device folders; pin
  rows also took `is_dir` from `Pin::recursive`.
- **B82** — a `parent_uid` cycle hung the daemon on any short search: the recursive CTEs in
  `path_of` and the pin walk use `UNION ALL`, which does not deduplicate. `path_of` is now
  depth-capped at 256 and returns the truncated path.

### Changed
- Local search ranking: whole-query prefix matches get their own lane (newest first) ahead
  of trigram rank, which on a home directory full of dependency trees was burying the
  obvious answer — 73% of the 500 best-ranked candidates for "test" came from one Go module
  cache. The single-character lane, which recovers a typo that destroys every trigram, stays
  behind it.

## [1.5.0] — 2026-08-11

Schema: 20.

### Added
- **`pdfs-prompt --dmenu`** — the same search through an external dmenu-style launcher
  (fuzzel, rofi, wofi, tofi, bemenu, dmenu) instead of the built-in GTK HUD, so the prompt
  matches a tiling-WM desktop. A launcher filters a fixed list rather than re-querying per
  keystroke, so searching is a two-step loop; `--query` skips the first step, fuzzel and rofi
  get file-type icons. Configurable as `"prompt": { "mode": "dmenu", "menu": [...] }`, with
  `--gtk` overriding for one invocation.
- **Configurable open-with rules** (`pdfs-core/src/opener.rs`): an ordered list of name/class
  patterns in `config.json` under `open_with`, each mapped to argv and optionally wrapped in
  a terminal, first match wins, falling back to `xdg-open`. Lets "open a text file" mean
  `nvim` in a terminal without breaking every graphical caller. Absent config reproduces the
  old behaviour exactly.

### Fixed
- The prompt and dmenu openers hand applications the cached file while matching rules
  against the *original* Drive filename, so e.g. `.md` files stop being routed by their
  cache-path extension.

## [1.4.0] — 2026-08-10

Schema: 20.

### Added
- **File version history** (`pdfs-fuse/src/revisions.rs` + a GTK versions dialog): list a
  file's revisions, restore one, delete one, or write an old one out to a local path. A
  restore re-points the file server-side, so no content crosses the wire and nothing enters
  the drain queue; the daemon drops its cached content and repopulates on the next read.
- **Persistent SDK entity cache** (`pdfs-core/src/sdkcache.rs`): the SDK's decrypted node
  metadata now survives a daemon restart instead of re-walking the tree. It lives in its own
  `sdk_cache.db`, deliberately off the daemon's hot `cache.db` mutex, holds no key material,
  needs no migrations, and is cleared on logout.
- Per-node outcomes for batch operations (`pdfs-core/src/batch.rs`): `trash`/`restore`/
  `delete`/`move` report one result per node, so a node the server rejects no longer fails
  the whole batch.
- Photo viewer page in the GUI and richer photo metadata in the details pane.

### Changed
- SDK telemetry routed into `tracing`, so its transfer/block-storage/request spans land in
  the daemon log instead of a no-op sink.
- A file that fits in one block uploads as a single atomic request rather than the
  draft/block/commit sequence.

## [1.3.0] — 2026-07-31

Schema: 19.

### Added
- **Photo albums** — `PhotoAlbums` / `AlbumPhotos` control requests, daemon-side `albums.rs`
  and a GUI albums page (the former gallery view, renamed). Albums other people share with
  you are included; their photos are persisted per album because a shared album's photos
  never appear in your own timeline. Served stale-while-refreshing like the timeline.
- **GUI "Locations" page** — one row per mount spec: the primary `~/ProtonDrive` mount
  alongside every folder this computer backs up, mirrored or on-demand. Called *Locations*
  rather than *Mounts* because a mirror folder is a plain directory with no FUSE session.
  It absorbs the old Computers page's actions and the mountpoint chooser from Settings.
- `role` on shared-node metadata (`viewer`/`editor`/`admin`), defaulted for wire-compat with
  older clients and daemons.
- FUSE acceptance coverage for the permission-based mounts.

## [1.2.1] — 2026-07-31

Schema: 18.

### Fixed
- Migration bug that could leave the `ProtonDrive` mount unwritable after upgrade.
- Permission error on on-demand mounts.
- B78.

### Changed
- Release pipeline: `.deb` signing added and `.rpm` signing fixed.

## [1.2.0] — 2026-07-29

Sharing and multi-mount release. Schema: 18.

### Added
- **Shared with me**: shares others sent you are listed, verified and mountable, including on
  the primary mount.
  They appear under a synthetic `Shared with me` directory on a virtual volume
  (`pdfs-fuse/src/virtual.rs`), with name collisions and over-long components resolved by
  its own naming rules.
- Per-share access model (`db/share_access.rs`): the effective role of each shared root is
  persisted and kept resident in mount state, so interning an inode never costs a query. The
  database stays the restart/offline authority — every role change is written through before
  the cache updates. A share you cannot write is presented POSIX read-only.
- Unified location model (`pdfs-core/src/mounts.rs`, `db/mounts.rs`): one presentation shape
  for every local Proton Drive location, joining device sync state with a new `mount` table.
  `MountMode` is `Mirror` (full local copy, reconciled both ways) or `OnDemand` (FUSE session
  fetching on access), with an `Unknown` fallback so a mode written by a newer client does
  not break an older one.

## [1.1.1] — 2026-07-27

Schema: 16.

### Fixed
- Trash refresh did not pull updates.
- Packaging: `PKGBUILD` was left at 1.1.0 by the 1.1.1 version bump, failing the release
  workflow's tag/manifest/PKGBUILD agreement check.

## [1.1.0] — 2026-07-26

Conflict-handling and concurrency release. Schema: 16.

### Added
- Conflict sweep (`pdfs-fuse/src/sweep.rs`, 5-minute cadence): removes conflict copies proven
  identical by size + `content_sha1`, surfaces divergent ones once as an activity entry, and
  never removes what it cannot prove.
- Sweep mode and `.pdfsignore`-driven conflict handling; ignore rules are rebuilt per sync
  pass so edits take effect without a restart.
- Substantially extended FUSE acceptance suite.

### Fixed
- **B69** — spurious `(sync-conflict)` copies. Revision identity now uses the server's
  `revision_id` rather than `(mtime, size)`, which the server re-stamps on seal; the planner
  gained `AdoptBaseline` so an equal-size no-baseline pair is recorded instead of forked on
  first reconcile.
- **B70** — browser downloads into an on-demand mount had their in-flight temporaries sealed
  as canonical revisions. Transient names are recognised and held until the name is finalized.
- Assorted race conditions and concurrency bugs on the write/drain paths.

### Changed
- Workspace deps bumped to `proton-sdk` / `proton-drive-rs` 0.2.2.

## [1.0.0] — 2026-07-22

First stable release: FUSE files-on-demand mount, sync daemon under `proton-drive.service`,
`pdfs` CLI and the GTK4 app, on top of the pure-Rust Proton Drive SDK. Schema: 16.

### Fixed
- The outstanding FUSE defects tracked in `docs/BUGS.md`, plus a truncate defect, validated
  by a new POSIX compliance suite for the filesystem.

[2.1.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v2.1.0
[2.0.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v2.0.0
[1.11.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.11.1
[1.11.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.11.0
[1.10.2]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.10.2
[1.10.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.10.1
[1.10.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.10.0
[1.9.3]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.9.3
[1.9.2]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.9.2
[1.9.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.9.1
[1.9.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.9.0
[1.8.2]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.8.2
[1.8.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.8.1
[1.8.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.8.0
[1.7.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.7.1
[1.7.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.7.0
[1.6.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.6.0
[1.5.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.5.0
[1.4.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.4.0
[1.3.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.3.0
[1.2.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.2.1
[1.2.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.2.0
[1.1.1]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.1.1
[1.1.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.1.0
[1.0.0]: https://github.com/narrrl/proton-drive-linux/releases/tag/v1.0.0