martinkoutecky/tine v0.6.985

Tine v0.6.985

on GitHub

iPhone / iPad: join the public beta on TestFlight. A TestFlight build is published for most releases; it may lag this one by a few days.

Changed

• The search index now keeps every page current by one rule: each change
  carries the moment it was made, and an older change never overwrites a newer
  one. At launch Tine compares each file with the index without parsing
  unchanged pages, so after pages change elsewhere or while Tine was closed,
  only the changed, added or deleted pages are re-indexed; a larger change
  still rebuilds it once. A cold, stale, damaged or configuration-changed
  index is rebuilt in an unpublished file and appears only when complete. An
  index holding inconsistent entries is rebuilt once instead of retrying
  forever. Query results now list pages in file-path order, so a page created
  this session takes its place instead of appearing last. The fixes below
  marked GH #543 condense 118 individual entries; the full list is in
  docs/releases/v0.6.985-gh543-details.md (GH #543).
• The Guide's platform pages now describe the iOS TestFlight beta (themes
  work, plugins do not run yet) and F-Droid updates for Android.
• Spell-check language names in Settings now use the standard
  "Language (Region)" form, e.g. "English (United States)" and "Spanish
  (Mexico)" instead of "American English" and "Mexican Spanish", so regional
  variants of one language sort together. Contributed by mikob (GH #580).
• Diagnostic reports now record when Tine was closed through the
  unsaved-changes warning with the user choosing to discard, with the reason
  (a failed save or one still running) and the number of affected pages, but
  no page names. Before, such a session read as an ordinary clean exit
  (GH #540).
• Internal: compact-projection P1 moved tine-storage through v0.25.0 and
  v0.26.0. v0.25.0 deletes the Managed
  Storage half of the crate (durable batches, local journals, sealed digests,
  the oplog frontier and the managed layout — 32,640 → 11,863 source lines).
  The crate became the Direct Files durability and SQLite projection crate;
  nothing Tine used moved. At that stage, the projection schema was unchanged
  at version 29, so no rebuild was triggered. v0.26.0 then trims the crate's
  public surface to the methods Tine's Direct Files path calls (no user-visible
  change).
• Internal: Tine now pins tine-storage v0.27.0 and projection schema 30. The
  compact schema removes redundant cache data and uses one integer coordinate
  space for pages and blocks; an existing schema-29 projection is rebuilt once
  from the Markdown/Org graph, without changing graph files.
• Ctrl+K, the (( picker and unlinked-reference search use a bounded recent-
  result window. Inline Friendly search shares their index, canonical folding
  and exact matcher, but remains exhaustive up to its 5,000-result limit and
  preserves display order.
• A cold, stale, damaged or configuration-changed search index is built and
  checked in an unpublished file, then atomically becomes ready only as a
  complete image. Before the first ready image, Ctrl+K and the (( picker use
  the loaded graph snapshot while structured queries keep waiting; failure or
  cancellation preserves the previous coherent committed image.
• Internal: scripts/measure-projection.mjs runs graph_scale_bench --root <graph> on a real corpus and reports the compact-projection budget rows (S1 file/Markdown, S2 build write amplification, T1 build time, M1 peak RSS, U1 bytes per single-block edit, T2 Ctrl+K p95, T3 {{query}} p95) against scripts/projection-budget-policy.json; src/projectionBudget.test.ts runs it only when TINE_PROJECTION_CORPUS names a corpus. Anon baseline recorded 2026-09-19 (compact-projection P0a).
• Internal: every SQL statement tine-core sends to the search/query projection
  now passes through one door (query/projection_sql.rs), a blessed statement
  census exercises every shape against the pinned projection schema, and the
  storage contract lists the projection's tables as its schema of record
  (compact-projection campaign P0; no user-visible change).

Added

• The online Guide now includes three sample journal days, so the Journals
  view shows tasks, scheduled items, a live query and quick-capture notes
  instead of an empty page.
• Clicking N conflicts opens a Conflicts page listing every page that needs a decision, grouped by source, with how many blocks are left on each. Click a page to resolve it, or shift-click to keep the list in view beside it; rows disappear as pages are resolved. The conflict lists in Settings → Backups & recovery moved there, including Discard copy for a sync copy whose page no longer exists. Suggested by EllisMorrow (GH #536).
• Live exports now have shareable page and block links. With one pane and one
  tab, the browser address follows navigation; the page and block menus provide
  explicit copy actions, and opening a link restores its page or block in the
  read-only exported app (GH #182).
• Tine's desktop binary now has a documented command-line interface: --help
  and --version; explicit open and capture commands; create-only static
  and live graph exports with graph-relative output and an explicit --replace;
  and a read-only doctor check. Existing tine GRAPH and --capture
  invocations remain compatible. Linux packages install a generated man tine
  page (GH #567).
• Running with TINE_DEBUG=1 (or --debug) now records the search index's
  lifecycle: when a validation starts reading the graph, how long that read
  took, how many pages it queued, each worker turn's duration and rows, and when
  a staged build becomes ready. Without the flag nothing is recorded (GH #543).

Fixed

• Ctrl+K and the (( picker remain useful while the query index is
  unavailable: they search one already-loaded parsed snapshot in document
  order and rerun against ranked results when the complete index becomes
  ready. Structured and inline queries still wait for it. Building a large
  graph's index the first time is about 45% faster (10,000 synthetic pages:
  47.5 s to 26.2 s on Linux; performance-only, same schema, nothing is
  rebuilt), and opening a large graph no longer spends most of its time
  re-reading the index it is building. Search and page autocomplete answer
  within seconds of launch instead of after a half-minute read storm, and a
  search typed during launch validation reports "indexing" and retries instead
  of starting a second build (GH #543).
• Opening a large graph no longer parses every page beside the search-index
  check. Page lists, aliases, reference counts, block references, embeds,
  templates and the calendar answer from the index or wait for its check, and
  opening today's journal or editing, saving or deleting a page meanwhile no
  longer throws the check away. A named page such as a restored tab or a
  favourite opens straight from its file during launch indexing: 4 s to 1 ms
  on a 10,000-page graph. When two files claim the same page name, the file
  named for the page is now the one opened. A journal named by a date
  title:: under a journal title format without a year now lists as a journal
  with its calendar day (GH #543).
• Common actions on a large graph no longer make Tine re-read, re-parse or
  re-check the whole graph: renaming, merging or deleting a page, a
  journal-format migration, switching graphs or reloading settings during
  indexing, an edit landing as the index check finishes, and opening or saving
  while another edit is being indexed. Creating the first page or journal day
  after reopening checks the name against the index (416 ms to 23 ms on a
  10,000-page graph). Saving during indexing no longer queues extra lookups,
  reference counts or alias reads per save. Indexing runs from one background
  owner, so a page that always fails to index costs a few spaced-out retries
  instead of a whole-graph pass after every edit (GH #543).
• Search, queries and page lists no longer get stuck waiting for the index, in
  some cases for the rest of a session with nothing left running to do the
  work. This could follow a failed index update or indexing step, renaming a
  page whose content could not be read, creating or renaming a page while the
  index still listed a deleted page, two parts of the app reading the graph at
  once, or a rename or delete during a rebuild. Launch indexing now always
  reports that it ended, even when it fails; before, page aliases,
  block-reference counts and reference-only page names never appeared for that
  session. A search during a rebuild says the index is not ready instead of
  returning nothing (GH #543).
• Search now describes the files on disk. A rename, merge or rescue no longer
  leaves search answering from the page it used to be, and a merged-away page
  can no longer come back. Duplicate journal days, journal filename migration
  and migrated PDF highlight pages no longer leave duplicate or trashed text
  findable, and edits to either file of a duplicate day reach search. A merge,
  rescue, delete, move or save that fails part-way leaves search matching what
  is really there. A late parsed copy of the graph can no longer undo a save
  in search (the file on disk was never affected), and pages that change,
  open, fail or recover during the index check reach it (GH #543).
• One page file Tine cannot read (invalid UTF-8, unparseable, locked, or still
  arriving from a sync service) no longer blocks creating every new page,
  keeps the index "not ready", or makes page listings parse every page. Tine
  says once which file it could not read; only a new page whose name that file
  could hold is refused, and that refusal, like a rename the file blocks,
  names the file. Fixing, restoring or deleting it clears the record. The page
  keeps what was indexed for it until it can be read again. Exporting,
  printing, renaming or deleting on a graph whose files cannot be read now
  reports that instead of exporting an empty site (GH #543).
• Settings now apply in place. Changing the home page, the format for new
  pages, a display-only setting, dismissing the Guide notice, or editing a
  setting in config.edn outside Tine no longer reloads the graph, restarts
  indexing, cancels a print or closes an unsaved-edit recovery. Every outside
  change to config.edn is now picked up, and :hidden updates All Pages and
  page links without a restart. A setting changed while the graph reopens is
  no longer lost (a favorite added then could vanish from config.edn).
  Aliases, links and reference counts follow a reopened graph at once, and a
  link no longer opens a page that merely declares its name as an alias. (GH
  #543).
• Tine now reports when the system refuses to watch a graph folder (for
  example at Linux's inotify watch limit) and checks it for outside changes,
  including config.edn, every few seconds by polling instead of missing them
  silently, catching up once watching works; graph-wide settings such as the
  journal title format take effect there too. Deleting or renaming a page
  outside Tine, or changing a setting, no longer rescans the whole graph
  folder; only a removed folder that held pages still does. Restoring a backup
  no longer processes restored pages twice, a sync tool redelivering a
  duplicate journal day's file unchanged no longer counts as an edit, and an
  outside edit the watcher first missed no longer stays out of search (GH
  #543).
• Windows and graph switches keep to their own graph. A graph change or
  finished launch in one window no longer makes another window reload, rebind
  or stop waiting, and after switching graphs the previous graph's conflict
  banners and favourites order no longer show on the new one. Closing a window
  releases its index at once; switching back to a graph Tine is still
  releasing, or opening one another Tine instance is still indexing, now waits
  and then indexes instead of leaving search without an index. An HTML export
  stops if its window switched graphs, and failing to update the recent-graphs
  list (for example on a full disk) no longer fails the open or skips indexing
  (GH #543).
• The indexing progress bar now shows while search waits for the index to be
  repaired, during a later rebuild in the same session, during whole-graph
  page reads (also when the index is ready, for example while finding unused
  assets), and for the whole build, including the first pass that reads every
  page, where Ctrl+K's progress line vanished for minutes and re-ran the
  search repeatedly; it now says it is indexing until it knows how many pages
  there are. It no longer reports a refreshed graph finished before it starts
  indexing, hides for a second on a repaint, freezes behind a page listing, or
  goes blank while the index is being written (GH #543).
• The app no longer freezes while Quick Capture suggests pages during
  indexing, while the journal title format changes, or during graph-wide
  listings (journal conflicts, filename migrations, asset trash, installed
  plugins, spell-check dictionaries) and large asset reads, imports or backup
  pruning; that work now runs off the main thread. Exporting query results or
  the graph to HTML during indexing waits for the index and then exports,
  instead of showing the wait as a failure. Linked and unlinked references in
  the right sidebar refresh after a rename, backup restore or journal-format
  change, moving between pages during indexing no longer shows linked
  references as failed, and a pinned block is removed at launch only when it
  no longer exists (GH #543).
• A search index file damaged badly enough that SQLite will not open it is now
  deleted and rebuilt from your Markdown (nothing in the index is a source of
  truth); several searches failing at once no longer rebuild it twice; and a
  damaged index no longer keeps answering or keeps a CPU core busy. A query
  with a very long (or …) or (and …) list (about a thousand terms) now
  answers instead of failing and rebuilding the whole index over and over. The
  debug report records why a query was refused instead of "other", still as a
  short fixed code, never anything from your graph. Performance: the record of
  pages changed during indexing no longer grows without bound (GH #543).
• Exports used the favorites and workflow the graph had when it opened,
  ignoring changes made since.
• When the graph's pages could not be read, the unused-assets list offered
  every asset for trashing, including media still in use. It now reports the
  error instead.
• Advanced queries written the way Logseq's documentation and many
  config.edn files write them now run: [?b :block/marker "TODO"], a marker
  or date variable narrowed by contains?, =, not=, <, <=, > or
  >=, the block's journal page and journal day, and
  (not [?b :block/scheduled ?d]). Before, these were refused as
  unsupported, and a not or or Tine only partly understood ran as the part
  it understood, so an "overdue" query could list scheduled tasks. Such a
  not/or is now left out whole, so a partial answer can only be wider than
  the query, and says so (GH #542).
• Opening a graph no longer rebuilds its search index from scratch on every
  launch. The Journals view reads its first days before the index has been
  checked, and those reads were given index positions that collided with the
  ones already stored, so the index was thrown away and rebuilt each time,
  a large, repeated write on every open (a ~40-second open was reported on an
  Android phone). Pages read at launch now keep
  their stored place, and unchanged ones are not re-indexed at all (GH #550).
• Two more pieces of whole-graph work are off the path to the first journal
  on every open. Today's journal is found from its date instead of by listing
  every file in the graph first, and the automatic launch backup now waits
  until startup has finished (the background indexing is done and the app has
  been quiet for a few seconds) instead of copying the whole graph one second
  after open, where it competed with the first page for the disk (GH #550).
• The online Guide at tine.page showed emoji in a monochrome outline font.
  Published pages now use the browser's color emoji font; the desktop editor
  on Linux keeps the monochrome font that avoids the WebKitGTK emoji
  crash (issue #76, unchanged).
• On a Mac whose web engine is older than Safari 15.4 (for example Big Sur
  with Safari never updated), Tine now says so at startup and explains that
  updating Safari fixes it, instead of opening with every page failing to
  display. The iOS build now requires iOS 15.4 (GH #572).
• On Android storage that refuses Tine's atomic "rename without overwriting",
  a settings write could leave logseq/config.edn hidden under a .retired
  name, after which the graph failed to open with "Invalid argument". Tine now
  checks the rename works before touching config.edn, and a graph already
  affected opens again with its settings restored (GH #538).
• The Journals view no longer shows "No journal entries found" when the graph
  folder holds something that is not a page, such as a symlink, an editor lock
  file, or a file the system cannot read (for example an iCloud file that is
  not downloaded). Such an entry also no longer stops today's journal from
  being created or pages from being renamed, and a day that cannot be opened
  no longer hides all the others (GH #385).
• A graph no longer opens with every page blank because one item in its folder
  cannot be read (for example a locked file, an unreadable folder, or a file
  with an unusual name). Tine now skips that item and shows the rest. Opening a
  large graph also no longer freezes the app for several seconds while it
  checks for conflict copies (GH #332).
• The default journal template could be applied again over text typed into a
  nested block of today's journal, replacing it. Text anywhere in the journal
  now counts, so the template never overwrites it (GH #550).
• With a custom journal title format, opening a graph no longer tries to write
  the journal template over today's journal delivered by a sync app, which
  caused a hidden save conflict on every launch (GH #550).
• A save conflict no longer ends up with a comparison that can never load. If
  Tine lost track of the conflict in the split second after it happened (for
  example because another page was deleted or renamed at that moment on
  Windows), the comparison failed, and after a restart it failed every time.
  Tine now compares your draft with the file as it is on disk instead, as it
  already does for any conflict it restores after a restart (GH #490).
• A page whose first bullet is a numbered-list item keeps saving. Making the
  first bullet of a new page a numbered item and then typing its text could
  leave the file holding only logseq.order-list-type:: number, with every
  later save of that page failing (reason code: unknown) until Tine was
  restarted. Tine had taken the empty numbered item for page properties; it now
  stays a list item (GH #540).
• A save Tine refuses in order to protect what is already in a file (a page
  header, merge-conflict markers, an Org file it cannot reproduce) is no longer
  retried and reported as reason code: unknown "after 3 tries" while the page
  silently stops saving. One message stays up until the page saves again, says
  the edits are kept in this window, and Review unsaved opens the draft to
  copy it or open the page (GH #535, GH #546).
• Typing a page's alias after [[ or # now offers it, and the row says
  which page it is an alias of. Choosing it inserts the alias as you wrote it,
  and the link opens the real page (GH #558, GH #482).
• One page that cannot be saved no longer blocks every rename in the graph.
  A rename used to need every open page saved first, so a single page whose
  save kept being refused made every rename fail until you deleted and
  recreated that page. Now such a page stops a rename only if the rename would
  change it, or its unsaved text mentions the old name, and the message names
  it. After a rename Tine reloads only the pages the rename changed, so
  unsaved edits on every other page are kept (GH #535).
• Renaming a page that many other pages link to is much faster in a large
  graph. Every file the rename rewrote re-listed the whole pages folder to
  check for names that differ only in case or accents, so the cost grew with
  the number of linking pages times the size of the folder; the folder is now
  listed once per rename. On an 8,000-page graph, a rename touching 800 pages
  went from about 2.6 s to 0.8 s of file work, and time now grows in step with
  the graph. The rename still reads every file once, and the app still reloads
  the graph afterwards (GH #406).
• Editing a page's properties no longer raises a save-failure toast while you
  are still typing them. Page properties written as the first bullet of a page
  are saved into the file's own header, but Tine kept treating them as an
  ordinary bullet afterwards, so a half-written key:: value line proposed a
  change the save path is required to refuse — surfacing mid-edit as
  "Couldn't save … after 3 tries — (reason code: unknown)". Tine now follows
  those properties into the file, and an invalid page header is reported when
  you finish editing it instead of on every autosave; the page saves by itself
  as soon as the properties are valid again (GH #546).
• "Export to PDF…" is no longer offered on Android and iOS, where it could not work: a mobile WebView cannot print, so the option silently did nothing. Choosing it from the command palette or a shortcut now says PDF export needs the desktop app instead of opening a dialog that leads nowhere (GH #560).
• "Export graph to HTML" now says why it exported nothing when a graph has no public pages, instead of reporting "Exported 0 pages" with no explanation. Only pages marked public:: true are exported, as in Logseq (GH #560).
• Saving a page no longer fails when the file is also hard-linked somewhere else. Graphs managed by git-annex (annex.thin mode links every page into .git/annex/objects/) or by a deduplicating tool refused every save with a physical-resource alias error, so the graph could be read but never edited. Tine now refuses only the case it can actually name - two pages inside the graph sharing one file - and only where it already knows the whole graph, which is page creation (GH #571, GH #555).
• A page that is also hard-linked elsewhere (git-annex, a deduplicating tool) can now be renamed; before, such a graph could be saved but every rename failed with a physical-resource alias error. Opening a graph no longer fails when a crash-recovery copy of a page has picked up a second link, and edits made outside Tine to such pages no longer force a full re-index. Two pages inside the graph sharing one file are still refused, with both named (GH #571, GH #555).
• A graph inside a Cryptomator or VeraCrypt vault (and any other user-space
  volume) can be created and opened again. Windows asks the volume driver to
  resolve a path to its final name, and these drivers are allowed to answer
  "this volume has no recognized file system" — which Tine treated as a broken
  path and refused, after it had already created the graph folders. Tine now
  falls back to the plain absolute path when the driver cannot answer; the
  no-follow open that actually protects the graph tree is unchanged. Saving on
  such a volume was affected by the same call and is fixed with it (GH #561).
• A query whose results are pages — {{query (page-property tags x)}} and any
  other page query — now shows those pages in every display view. Table and
  Board said "No results" and Search reported a count of 0, because those three
  views read the block results a page query never produces, while List read the
  pages. The page results were always there: the summary panel went on counting
  them correctly beside the header's 0. Changing the view now changes only how
  the same matches are drawn, and the count is the count of what is on screen
  (GH #547).
• Typing a [[ page link on a large graph no longer re-reads the whole
  reference index on every lookup. Offering page suggestions needs the set of
  page names the graph links to, and the index answered that by handing back
  one row per link — 110,000 rows to name 10,010 pages on a 10,000-page
  graph — which took about 1.4 seconds every time it was asked, whatever was
  typed. That set is now remembered until the graph changes, so a lookup takes
  about 10 milliseconds. Page autocomplete, the quick capture picker, the query
  sheet's page field and the settings page picker all share the one answer,
  and a newly linked page still appears immediately
  (REG-QUICKSWITCH-REFERENCE-NAMES-MEMO-001).
• The first [[ page lookup after an edit is no longer slow either. Remembering
  the set of linked page names fixed every repeat lookup, but the first one
  after each edit still had to read the whole reference index. The index now
  keeps those names in an order it can read straight through, without consulting
  the pages themselves, so collecting every linked name on a 10,000-page graph
  takes about 20 milliseconds instead of 1.3 seconds. Two consequences worth
  knowing: the first launch after this update rebuilds the graph's search index
  once, which takes longer than a normal start, and the finished index uses
  about 5% more disk. Saving is unaffected
  (REG-QUICKSWITCH-REFERENCE-NAMES-INDEX-001).
• Searching a large graph no longer reads every block for every keystroke.
  Ctrl+K asked SQLite to materialise and rank every block in the graph and
  then keep the matches, and it did that twice per keystroke: once for the
  block results, and again to work out which pages had matching content. On a
  10,000-page graph each keystroke cost about 4.5 seconds whether it matched a
  hundred blocks or none at all. Both reads now start from the substring index
  and visit only the blocks that can possibly match, bringing a selective
  search down to about 1.7 seconds. A needle common enough to match more than
  twenty thousand blocks still uses the old plan, which is genuinely the
  faster one at that point. The answer is unchanged: the exact text predicate
  still decides every row (REG-FRIENDLY-BLOCK-CANDIDATE-BOUND-001).

Download an installer for your platform below. Windows and Linux ship both x64 and ARM64 builds (match your CPU). Windows users who prefer no installer can grab the portable Tine_*-portable.zip for their architecture. Windows builds are currently unsigned, so Windows may show a warning on first launch. The macOS build is signed and notarized. On macOS, if Tine repeatedly asks to access Documents, see the workaround in the README.