# Changelog All notable changes to this project are documented here. This project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html). > Entries before 0.19.0-dev7 describe the acquisition client as a specific private tracker and > name its settings by their current `TRACKER_*` names. The client was generalized then, and the > older setting names remain accepted as deprecated aliases. ## [0.21.0] - 2026-11-16 Five audit batches, one release. Each batch was its own pass with a green suite, and the matcher changes were verified against the real 123-book library rather than argued about. **The database stops losing things.** A re-save no longer wipes a book's match evidence, a pinned book's cover is found again, and the database runs in WAL mode. Backups change with it: `-wal`/`-shm` now sit beside the file while the app runs, so copying `audiobind.db` alone is not a complete backup — stop the app first, or use `sqlite3 audiobind.db ".backup copy.db"`. **An encode can no longer leave a half-written book.** Publishing is atomic across filesystems, ffmpeg's stderr goes to a file so a full pipe can never block an encode, the duration check now catches a dropped 15-minute file in a ten-hour book (it used to allow twelve minutes of drift), and zero-length chapters are dropped in every tier. **The pipeline stops redoing work.** A published file is honoured whatever the row's status, the work directory keeps its name across restarts, a LOW folder match is re-checked against the audio's own tags, and a failed second format rolls the first back out of the library. **The matcher stops trusting a single candidate.** A window-only match can no longer reach MEDIUM, coverage now measures the candidate it gates, a dead ASIN pin gives up after three checks, and an adaptation nobody asked for loses to the plain edition of the same book — which put four Riyria books back on their real editions at HIGH. Plus the audit's tail: a non-object catalogue answer is a failed lookup instead of a crash, a GET on a POST-only route answers 405 with `Allow`, an unknown `/api` path answers a JSON 404, and the app reports the version it was actually built as. The batch-by-batch detail follows. ### [0.21.0-dev14] - 2026-11-16 Support modules and the web layer: two fixes, two containment/regression proofs. The rest of the audit entries for these modules were already handled or did not hold up. ### Fixed - **A catalog answer that isn't a JSON object no longer crashes the lookup.** When Audible or Audnex answered with a bare list or string, `search`, `book`, `chapters` and `product_images` called `.get` on it and raised. `_get_json` now treats any non-object body as a failed lookup. It is not cached and not retried. The Discover queries still report it as "could not reach the catalog", not as "nothing found". - **A GET on a POST-only route is now 405 with `Allow: POST`.** Before, only `/logout` and the connection tests did this. A GET on `/api/account`, `/api/unit`, `/api/settings`, `/api/control`, `/api/library/abs-layout` and the wishlist actions rendered the page shell instead. One table (`POST_ONLY_ROUTES`) now covers them all. - **An unknown `/api/...` path now answers a JSON 404**, not the HTML page. ### Tests - The "Needs you" runtime group sentence is pinned at 35 % and 40 %, not only 25 %. The sentence already showed your configured tolerance. The new pin catches a reworded advice sentence that would silently stop the swap. - `/cover?path=` containment is proven over HTTP. Traversal, absolute paths, NUL bytes and a `cover.jpg` symlinked out of the output tree all get a 404 and never serve the file. The path is only ever a key into the state DB. ### [0.21.0-dev13] - 2026-11-16 Picking the right edition when an adaptation is in the running, gated on the same regression run. ### Fixed - **A plain edition beats an adaptation you didn't ask for.** 0.21.0-dev12 stopped the GraphicAudio dramatization from auto-converting, but it still won the pick. When the catalog also returned the plain edition, the two tied (same score, title 1.0). The tie went to the dramatization, and the tie drove the margin to 0. An adaptation the folder doesn't ask for is now handled the way a knockoff is: it never ranks above the plain edition of the same book, even when its "Book 1" subtitle out-scores the plain "Volume 1". It also never counts as that edition's rival for the margin, so the plain edition can reach HIGH. The reverse also holds: a folder that names an adaptation (`[Dramatized Adaptation]`, `GraphicAudio`, `Full Cast`, ...) gets the adaptation over the plain edition. An adaptation found with no plain edition beside it is still offered, still capped at MEDIUM. - **"Title - Series, Book N" folders find the bare title.** A trailing ` - , Book N` tail (also ` - - Book N`, ` — Book N`, Volume / Vol.) is now dropped as an extra title reading. This is what finds the plain *Rise of Empire* for a folder whose misspelt author ("Micheal") finds nothing. - Regression run (`docs/match-after-tiebreak.csv` against `docs/match-after-adaptations.csv`): four rows changed, all Riyria. - *Rise of Empire*: B09P49F6J7 MEDIUM → B0081IIZ6Y HIGH (margin 0.978 → 0.572). - *Theft of Swords*: B09PQ951KX MEDIUM → B007KSJKI8 HIGH (margin 0.022 → 0.519). - *Heir of Novron*: B09P4BYYX5 LOW → B008UCYCSQ HIGH (margin 0.058 → 0.269). - *The Crown Tower*: B00DZ0G088 LOW → HIGH, same ASIN (its dramatization had tied it; margin 0 → 0.378). Tiers: HIGH 69, MEDIUM 9, LOW 6, NONE 7. ### [0.21.0-dev12] - 2026-11-16 Audiobook adaptations, gated on the same regression run. ### Fixed - **An adaptation the folder didn't ask for is never HIGH.** Three Riyria Revelations folders, originally converted from the plain Tim Gerard Reynolds editions, now matched the GraphicAudio dramatizations. *Rise of Empire* did so at HIGH. Its catalog search returned the dramatization alone (the folder misspells the author "Micheal", so the author search finds nothing), which gave it a margin of 0.978. A catalog hit whose title or subtitle says *Dramatized Adaptation*, *GraphicAudio*, *Graphic Audio*, *Full Cast* or *Audio Drama* is now capped at MEDIUM, as a foreign-language hit is. The exception is when the folder name, the scanner's title reading or the embedded-tag title carries any of those markers. With the default `AUTO_CONVERT=high`, such a hit waits for your click instead of converting unseen. It is not dropped to LOW, because title and author do match. The panel note reads "catalog hit is a Dramatized Adaptation and your folder does not say so", and the match evidence stores the phrase (`unasked_adaptation`). An ASIN you pin is never affected. - Regression run (`docs/match-after-adaptations.csv` against `docs/match-after.csv`): one row changed. *Rise of Empire - Riyria Revelations, Book 2* went from HIGH to MEDIUM on B09P49F6J7 (title 1.000, author 0.944, margin 0.978). No ASIN changed. Tiers: HIGH 65, MEDIUM 11, LOW 8, NONE 7. ### Measured, not changed - **A title reading without the " - Series, Book N" tail.** Searched live, the bare titles do return the plain editions. But each then ties its dramatization exactly (both score 1.0, both Audnex-corroborated). The tie-break keeps the dramatization, and the tie brings the margin to 0. All three books went to LOW on the dramatization and none switched to the plain edition, so the reading was not kept. The fetched responses stay in `docs/match-catalog-cache.json`. - **Theft of Swords and Heir of Novron already retrieve the plain edition.** It loses on score: the dramatization's subtitle "Riyria Revelations, Book 1" matches the folder's "- Book 1" better than the plain edition's "Volume 1". Those two stay at MEDIUM (margin 0.022) and LOW. ### [0.21.0-dev11] - 2026-11-16 Matching, gated on a before/after regression run over the real library. ### Added - **Match regression harness.** `docs/match-regression.py` grades every library row of a state database with the matcher, against a frozen catalog (`docs/match-catalog-cache.json`), so two runs see the same Audible data and the diff measures the code. Ingest rows and keyed part-bucket rows can't be rebuilt from the database alone; they are skipped and each one is listed. The baseline from 0.21.0-dev10 is `docs/match-baseline.csv`, and this release's run is `docs/match-after.csv` (91 books graded, 32 skipped). Before: HIGH 66, MEDIUM 11, LOW 7, NONE 7. After: HIGH 66, MEDIUM 10, LOW 8, NONE 7. No book changed ASIN. One changed tier, listed under Fixed. - **Matching fixtures** (`tests/test_match_fixtures.py`), written before the fixes: Dune vs Dune Messiah, summarized and study-guide knockoffs, a German edition, a box set named for its series, a series echo, an author-less folder. Cases that still fail are listed in `KNOWN_GAPS`, and the test checks that they still fail. A fix has to promote its case out of that list. ### Fixed - **A window-only match is never MEDIUM.** HIGH already refused a catalog title that is only one run of words from the folder name (under 60% of it). Both MEDIUM branches allowed it, and MEDIUM is one setting away from converting automatically. Such a match is now LOW. A foreign-language hit can still reach MEDIUM, its cap. In the library, *Heir of Novron - Riyira Revelations Book 03* moves from MEDIUM to LOW on the Dramatized Adaptation (B09P4BYYX5). That title covers 3 of the folder name's 6 words, and its subtitle supplied the rest of the score. - **A dead ASIN pin gives up after 3 checks in a row.** An ASIN you set that has no catalog data used to hold the book in review with the same message on every re-check, forever. Now, on the third check in a row with no catalog data, the book is matched normally and graded normally. It gets a note that the pin was set aside, and it is never converted automatically: you asked for one edition, and the search may have found another. **Convert anyway** still works. The pin is still tried first on every check, and the count goes back to zero when the pin works again, when you set an ASIN (the same one again counts), or when you unlock it. The count is stored in the database (new `units.pin_failures` column), so it survives a restart, and a re-check never resets it. - **Window coverage measures the same candidates as the score it describes.** It used to read the subtitle even when the title floor score did not, because the title alone was under `SUBTITLE_GATE`. This had no effect today, because `TITLE_FLOOR` and `SUBTITLE_GATE` are both 0.60. The two no longer have to stay equal. ### Measured, not changed - **The title/author blend.** `t_sim` (weighted Dice) and `a_sim` (SequenceMatcher with containment) are on different scales. Across the 791 scored candidates with an author, `a_sim` is bimodal (330 are exactly 1.0, the 10th percentile is 0), while `t_sim` is spread out (mean 0.47, sd 0.29). Four recalibrations changed no book's tier or pick: author scored with `title_sim`, 0.70/0.30, 0.75/0.25, and z-scoring `a_sim` onto `t_sim`'s distribution. The blend only ranks candidates and sets the margin, and `grade()` gates each score on its own. The blend stays at 0.60/0.40. ### [0.21.0-dev10] - 2026-11-16 Pipeline flow. ### Fixed - **A LOW folder match is re-checked against the audio files' own tags.** An ingest folder name is a release name and can match the wrong book at LOW. The embedded-tag retry only ran when the folder matched nothing, so that wrong match stood. It now runs on LOW too. The tags' match replaces the folder's only when it grades higher, or at the same tier scores at least 0.10 better. Otherwise the folder match stays, and a different book the tags pointed to is named in a note. "matched on the audio file's embedded tags" now appears only when the tags' match was actually used. - **A crashed encode's work folder is found and removed.** The folder was named from Python's `hash()`, which changes with every process, so after a restart the same book got a new folder and the old one stayed in `WORK_DIR` forever. The name is now a stable SHA-1 prefix of the book's key, and the folder is emptied before each encode, so a crashed run's partial file can never be taken for new output. Folders left by older versions are not removed automatically; once no encode is running, anything in `WORK_DIR` can be deleted. - **A published file is never encoded twice or forgotten.** Two cases still encoded a duplicate into the current layout. One: a `ready` row (its file was already published) whose file sits under an older layout. Two: a converted book that a re-check moved to review or held. That re-check also cleared the row's output path while the file stayed on disk. Now any row whose recorded file still exists is left unencoded, a re-check keeps the output path, and such a row is recorded `ready` ("already published") when it would otherwise convert. A `converted` row is still left exactly as it was. **Convert anyway** still re-encodes. - **M4B + MP3 publishes both files or neither.** Both formats already encoded before either was published. But if the mp3's publish failed (say, the disk filled during a cross-filesystem copy), the m4b stayed in the library, and the next scan took it as already published, so the mp3 was never made. A failed publish now removes the files, `cover.jpg` and empty folders that this attempt created. Nothing that was there before is removed. ### [0.21.0-dev9] - 2026-11-16 Encode correctness. ### Fixed - **A finished book reaches the library whole or not at all.** Publishing moved the encode with `shutil.move`, which across filesystems is a copy and a delete: a copy that died (disk full, I/O error) left a truncated m4b at the library path. A cross-filesystem publish now copies to a temporary file in the destination folder, fsyncs it and renames it into place; any failure removes the temporary file and leaves the old file (if any) untouched. `cover.jpg` is written the same way, and only once the audio is in place. On one filesystem it is still a single rename. - **A chatty ffmpeg can no longer hang an encode.** ffmpeg's stderr went to a pipe drained by a thread; if that reader stalled, the 64 KB pipe filled and ffmpeg blocked forever. stderr now goes to `ffmpeg-stderr.log` in the work folder and its tail is read after ffmpeg exits, so the failure message still carries the reason. The drain thread is gone. Stop still stops. - **No zero-length chapters.** An Audnex chapter with `lengthMs` 0 (or two chapters on one offset), a source file that measured 0 s in per-file mode, or an embedded chapter tag past the end of its file each produced a chapter with `start == end`. Such chapters are now dropped; the chapter before still runs to where the next one starts. ### Changed - **The dropped-file check is tighter.** The merged output must match the summed source durations within 30 s or 0.2% (72 s on a ten-hour book). The old allowance was 2 minutes or 2% (12 minutes on ten hours), wide enough to pass an encode that had lost a whole 15-minute file. ### [0.21.0-dev8] - 2026-11-16 Data safety in the state database. ### Fixed - **A re-save no longer wipes a book's detail.** Saving a row with an empty `detail` used to replace the stored one, so any caller that built a fresh row lost the match evidence, the catalog cover URL and the re-check banner flag without a word. An empty `detail` now means "not supplied" and the stored blob is kept; clearing it on purpose is `State.clear_detail(path)`. No code path cleared it that way today. `alternates` is left as it was: its only writer is the matcher, which always stores a JSON list (`[]` when there are none), never an empty string. - **A pinned book's cover is found again.** `detail_for_asin` only looked at the matched ASIN, so a book pinned with an ASIN override but never matched had no detail, and the cover proxy had no image URL for it. It now looks at the matched or pinned ASIN, as `has_asin` already did. ### Changed - **The database runs in WAL mode** (`synchronous=NORMAL`, 30 s busy timeout). Reads and writes no longer block each other inside SQLite, so the app's lock is held for less time. An existing database converts by itself on first start and stays converted. **Backups:** `audiobind.db-wal` and `audiobind.db-shm` now sit next to `audiobind.db`. While the app is running, recent writes may live only in the `-wal` file, so copying `audiobind.db` alone is not a complete backup. Stop the app first (SQLite then folds the `-wal` back in), or take a live copy with `sqlite3 audiobind.db ".backup copy.db"`. ## [Unreleased] - The ASIN link opens the right Audible page again — the old address form stopped working. - The Settings number-field unit labels read correctly now ("1 credit", "1 book", not "1 credits"). - A non-English listing no longer skips the title, author and margin checks: it has to clear them like any other, and is then capped at MEDIUM — so a weak foreign hit grades LOW instead of MEDIUM. - The "why this match" panel names the grade it was given and the margin behind it, and the big number reads "title-only score" when your folder names no author. - A runner-up with the same score but a different author is no longer called "another edition" — the caption says to check which one you have. - A title that only contains another one no longer scores as a near match: "Dune" inside "Dune Messiah" and "Blue Moon" inside "Once in a Blue Moon" used to score 0.85 on that alone. Containment now counts by the share of words covered, and only when the shorter side has two words or more. - Titles are compared word by word now, not as plain strings: a leading "The" / "A" / "An" does not count against a match ("Midnight Line" is "The Midnight Line"), small words like "of" count less than real ones, and a typo still scores ("Steeleart" vs "Steelheart"). - A subtitle can no longer earn the title credit by itself. It only counts once the catalog title does, so a "…Hunting Lee Child's Jack Reacher" subtitle never makes a stranger's book a match for "Lee Child". A catalog title with an "aka" in it is scored by each of its names too, so "Count to Ten" matches "Count to Ten aka Private Delhi". - The catalog search asks for the cleaned-up title as well as the name on the folder. Audible's title search is strict, so "Private 20 - Private Dublin" found nothing and the pick came from an author-only search. The search now also asks for "Private Dublin", the title without rip noise ("chapterized", "64kbps", "read by …"), and title plus author. When every title search comes back empty it makes one looser keyword search. "Private 18 Private Rome" becomes "Private Rome" too. - Folder and file names are read more carefully: "13 Hours" and "3 Body Problem" keep their number, a loose library file loses its ".m4b" in the title, "(Read by Jeff Hays)" is taken as the narrator rather than the author, and "Dungeon Crawler Carl - 04 - The Gate of the Feral Gods" is read as series, number and title, not as an author called "Dungeon Crawler Carl". - When both sides of "Past Tense - Lee Child" look like a name, the catalog decides which one is the author: both readings are searched and the one that finds the better book wins (a tie keeps the folder's order). The swapped reading is only searched when the folder's own reading finds nothing convincing (best score under 0.90), so a confident match never pays for a second set of searches. - A catalog title with a narrator or edition tag no longer hides the book: "Lock In (Narrated by Wil Wheaton)" matches a folder named "Lock In" (the tag is ignored on the catalog side, as it already was on the folder side). - An ASIN in brackets in a downloaded torrent's name ("In Too Deep [B0CLGF5CFH]") is no longer thrown away with the other tags. It is looked up and, if its title matches, added as a candidate. It is graded like every other candidate and never pinned. ### Added - **The watcher budgets the account instead of only obeying the site's moment-to-moment answers.** A freeleech wedge is never refunded and a tracker account has a limited number of unsatisfied snatches, so four new live settings (Settings > Watcher, no restart; `0` = no limit, and for the reserve, none): `WEDGE_MAX_PER_CYCLE` (default 1), `WEDGE_MAX_PER_DAY` (default 3), `WEDGE_RESERVE` (default 1) and `WATCH_MAX_UNSATISFIED` (default 0). The operator sets them; nothing is inferred. - **A wedge ledger.** Two additive columns on the wishlist table record each credit the watcher spends and when. The per-day cap and the reserve are counted from that ledger only - the site's own wedge count is deliberately not read, since no documented endpoint exists to read it. - **Unsatisfied snatches are counted from qBittorrent.** One read of the category per cycle: a torrent the watcher added that is below 100% or below a 1.0 ratio is unsatisfied; one qBittorrent no longer has is not. At `WATCH_MAX_UNSATISFIED` the cycle holds like the incomplete-download guard does, independently of it. - **The watcher panel says where the wedges stand**, e.g. "Wedges: 1 spent in the last 24 hours, 2 left of today's allowance, 1 kept in reserve." With no freeleech parameter it says no credits are ever spent. ### UI quick wins Calm the page, one primary per view; no new features and no layout changes. - **Easier to read.** The dimmest label colour is lighter, the Review dots no longer pulse, and the Held dot is amber like the rest of the app. - **Problem rows are one line.** The reason under a book is cut to a single line with the full sentence in its tooltip, and the shelf starts collapsed unless you opened it this session. - **The rail only shouts when there is something to see.** Review, Held and Failed turn amber only when their count is above 0; the Series count is no longer amber. - **Fewer buttons that do nothing.** Pause after this book and Stop now appear only while a book is encoding (Resume still shows when paused). Dry run is neutral until it is on. - **One primary per view.** Scan, Retry now and Get this are secondary; Convert anyway is the one amber action for a held book. "Scan now" is now **Scan for changes** and "Re-check all" is **Re-match all...**. - **Library page is titled "Library"**; the rail already carries the name. - **No more empty boxes.** The Held glyph and the lock chips no longer use characters some fonts lack ("Locked by you" / "Unlocked" in plain text). - An opened detail panel scrolls clear of the sticky header, and the encode strip's speed, ETA and book title now update live. - Removed `docs/deploy-0.20.0.sh`, which tripped the repo's own no-addresses guard (still in history). ### Needs your review One list for everything that is waiting on you, instead of the same books twice. - **The Needs-Attention shelf is gone.** The library now opens on **Needs your review** whenever something is waiting: the problem books, grouped by reason (Failed, Held, Stopped, Review), with the reason explained once per group instead of once per row. When nothing needs you, it opens on All with one all-clear line. - **A short reason per row.** A new **Why** column says what is wrong in a few words ("Damaged files", "Runtime 73% long", "Not on Audible", "Low match") plus the numbers behind it, instead of a paragraph under the title. The full sentence is still in the row's tooltip and in the detail panel. - **Match is easier to read.** For books that need you it shows the confidence word next to its meter; for books already converted it shows the series position. - **The age is a relative age**, never a date ("7d", "4h"), and it lives in the row's tooltip and the detail panel now that the column is gone. - **One way in.** The rail has one triage entry, Needs your review, instead of a separate Review / Held / Failed block, and the chips read Needs your review / Review / Held / Converted / All. - **The view names itself.** The page heading and the tab title read **Needs your review · 27**, so it is obvious which view you are in - **Library** for All, and the status name for a single filter. - **Length and Waiting are gone from the table.** Length repeated the runtimes the Why column already shows for a held or review book, and Waiting's coloured dot repeated the group heading. The age is still there when you hover a row ("Held · 7d") and in the detail panel. - **The table is Book · Why · Match**, and the Book column now takes the space the dropped columns freed, so long titles are no longer cut off while a wide column sits empty. - Fixed: a held book with damaged source files was told "that gap usually means a different edition". It now leads with the real cause, e.g. "One of your 18 source files can't be read, which probably explains why your files run 1.38 hours when Audible lists 7.72 for this edition (82% shorter)." - Fixed: a long status label could squeeze the age out of the row. The label truncates now, never the age. - Fixed: paths in the rail footer wrap at a slash instead of mid-word. ### The watcher settings live in the app - **No container edit needed.** The tracker URL, session cookie, cookie name, search path, body, method, encoding and category, download path and parameter, freeleech parameter, user agent, VIP, the qBittorrent URL, user, password and category, and the downloads-in-progress limit are all editable on the Watcher tab, saved in the app's own database. - **You can see where each value comes from.** A value saved in the app wins over the container's environment, each field says which of the two it is using ("set here" / "from the container"), and **Reset to container value** puts it back. - **Test qBittorrent** and **Test tracker** check the values you just typed and say what went wrong in plain words, instead of leaving you to find out from an empty cycle. - **Secrets stay out of the page.** The session cookie and the qBittorrent password are never shown back: the box only says whether one is set, and saving with it empty keeps what is already there. - Fixed: the tracker client used to be built once and kept forever, so a new session cookie or URL never took effect. It is rebuilt when those settings change. - Settings that only make sense for an unusual setup (`TRACKER_PROXY`, `TRACKER_HOME_IP`) stay in the container and are listed on the Advanced tab, so it is clear where they live. ### Settings has tabs - The page was one long column; it is now **Library**, **Watcher**, **Account** and **Advanced**, with one save button, and the tab you are on survives a refresh (`/settings?tab=watcher`). ### Changed - **The budget fails closed.** If qBittorrent cannot be read, no credit is spent that cycle; books that are already free still go through. A torrent that is already free never consults the budget. A row a budget refuses is left wanted with the reason in its note, and the cycle log gains "N skipped for the wedge budget" when there are any. A dry run reports what it would have spent and never writes the ledger. - A cycle now makes one read-only qBittorrent request even when there is nothing to send. ## [0.20.0] - 2026-11-15 First-run polish from a real fresh install. ### Changed - **The deprecated access token no longer greets a new user.** The template's `AUTH_TOKEN` field moves behind the Advanced toggle and its description now leads with the upgrade case: an install that still signs in with a token keeps working, and a new install leaves it empty and creates its account on the setup page. The variable, its masking and the sign-in behaviour are unchanged. - **Audio channels default to 2 (stereo), not 1.** Defaulting to mono silently discarded the source channel layout, which contradicted "a richer source keeps its own bitrate". `AUDIO_CHANNELS` (the config default, the template, `docker-compose.yml` and `docs/configuration.md`) now defaults to `2`; `AUDIO_CHANNELS=1` still wins when set, and still halves the file size. - **The Unraid template is pinned to the release tag** (`ghcr.io/stinger303/audiobind:0.19.0`, not `:latest`). Unraid reuses a matching local tag instead of pulling, so an install made once kept an old build on every reinstall. `docs/PUBLISHING.md` now has a release step that bumps the template's tag and ``. - **The Series shelf shows a series you own one book of.** The shelf used to need two of your books; it now gates on the Audible catalog: a series whose listing has two or more numbered entries gets a card whatever you hold, so the series you own one book of and did not know was a series appears. A series the catalog lists with exactly one entry is a one-off and gets no "1 of 1 owned" card. Until a listing is fetched the card says "fetching series size" and never guesses a total. When the catalog has no listing, or cannot be reached, only a series you hold two or more books of stays. Series with two or more of your books sort before single-book ones. Every series you hold now queues a catalog lookup (at most one a day each). - **A book row looks and acts expandable.** The row has a pointer cursor and a chevron that turns to point down when its details are open. It is reachable by keyboard (`tabindex`, `role="button"`, `aria-expanded`); Enter and Space toggle it. The kebab menu, buttons and links in the row still do only their own job. ## [0.19.0] - 2026-11-15 The first public release since 0.18.0: a first-run checklist, Audiobookshelf folder names with a previewed migration, honest delete semantics for torrent-sourced books, and the fixes from the 0.19 release audit. ### Changed - **Sign-in lockout also has a global budget, and counts IPv6 per /64.** On top of the unchanged 5 failures per address in 10 minutes, at most 30 failed sign-ins from all addresses together are allowed in that window, and an IPv6 client is counted by its /64 (an IPv4-mapped address as its IPv4 address). A successful sign-in clears that client's failures from both budgets. - **Upgrade notes for the folder layout.** `docs/install.md` and `docs/configuration.md` now say that converted books stay in their old folders after upgrading, that **Match Audiobookshelf layout** should run before any Re-check all / Retry / Convert anyway on converted books, and that Audiobookshelf may treat a renamed folder as new and lose listening progress. - **New conversions follow the Audiobookshelf folder convention.** `Author/Series/Book/` or `Author/Book/`. Every author is now in the author folder, joined with ` & ` (only the first was used before). A series book is always filed under its series folder: ` - ` with a sequence, plain `<title>` without one (a series book with no sequence used to fall back to `Author/Title/`, which Audiobookshelf reads as a standalone). The sequence is no longer zero padded (`3 - Title`, not `03 - Title`). File names, path sanitising and the `OUTPUT_ROOT` containment check are unchanged; no narrator braces or ASIN brackets are added (the tags carry them). Existing books are not moved by a scan. - **Delete works on ingest books; the torrent keeps seeding.** Delete on a torrent-sourced book used to be refused outright, even though the converted `.m4b`/`.mp3` in the output folder is Audiobind's own file. It now opens the normal confirm modal ("Remove from Audiobind?", button **Remove from library**), removes only the converted copy (same `OUTPUT_ROOT` guard and missing-file handling as a library delete) and forgets the row. `/ingest` is never touched: `POST /api/unit` answers `source_kept: true` with the reason "the files are a torrent that is still seeding in qBittorrent", plus `dismissed: true`. Confirmation is still required, and a library delete is unchanged. ### Added - **Settings > Library: Match Audiobookshelf layout.** Moves books converted under the old naming into the new layout, on request only. **Preview** (`POST /api/library/abs-layout` `{"dry_run": true}`) returns every planned move and every skipped book with its reason, and writes nothing. **Apply** (`{"dry_run": false, "confirm": true}`, behind the confirm modal) moves each book folder (a same-filesystem rename, else `shutil.move`), renames just the file when only the file name differs, updates the row's output, removes folders left empty, and records every from/to in `STATE_DIR/abs-layout-<UTC timestamp>.json`. Never overwrites: skipped when the target exists, the source is missing or outside `OUTPUT_ROOT`, another row shares the folder, or the folder holds another book's audio. One failure is reported and the run carries on; refused while a book is encoding. The card warns that Audiobookshelf may treat a renamed folder as a new book and lose its listening progress. - **Removed ingest books stay removed.** A new `dismissed` table remembers the state key of each ingest book you removed, and the scan skips it before it is checked, prefetched or converted, so the still-seeding torrent is not converted all over again. A dismissal is forgotten once its folder (or single file) is gone from the ingest folder, so the same torrent re-added later converts as new. Pruning only happens while the ingest folder is enabled and reachable: an unmounted or disabled ingest folder never forgets a dismissal. ### Fixed - **Delete only removes an output folder Audiobind wrote.** A book whose output path already existed is stored `ready` (already published) with that file as its output, and Delete used to remove the whole folder anyway. The output is now removed only for a `converted` row, and only when the path recorded with the conversion matches the row's output. Any other row is forgotten and its folder left alone: `output: "kept"`, `output_kept_reason: "that folder was not created by Audiobind"`. The `OUTPUT_ROOT` and shared-folder guards are unchanged. - **A failed output removal is a failed delete.** When the converted copy could not be removed, Delete still forgot the book and answered `ok: true`. It now keeps the row (an ingest book is not dismissed), answers 500 with `ok: false` and the reason, and the toast shows it. - **Re-check no longer re-encodes a converted book into the new layout.** A re-check (Re-check all, Retry, a per-book Re-check) only tested whether the book's NEW path existed, so after the 0.19.0-dev14 layout change every series or multi-author book was encoded again: hours of CPU, a duplicate copy and an orphaned old folder. A `converted` row whose file still exists is now left untouched, with the log line `already converted at <old path> - run Match Audiobookshelf layout to move it`. **Convert anyway** still re-encodes; it is the only action that does (`reencode`, a flag separate from the scan's force). - **The title score is whole-string again; a window of the folder name is no longer an exact match.** This replaces the best-window `title_sim` above, which gave 1.0 to any catalog title equal to one run of the folder name's words: `Dune Messiah` picked `Dune` at HIGH when only Dune was returned, and tied the right book at 1.0 when both were. `title_sim` is now the best whole-string score over a few readings of the folder title: as is; without `[...]`, `{...}` and `(...)` groups; without a leading `NN - ` / `NN. `; and `<X> NN - <Y>` reduced to `<Y>` when X and Y are alike. `Dungeon Crawler Carl 01 - Dungeon Crawler Carl` still scores 1.0 and grades HIGH; `Dune Messiah` against only `Dune` is MEDIUM, and beats `Dune` at HIGH when both are returned. Equal scores are broken by the raw title, then the runtime. The best-window measure is kept only for the title floor, and a pick that only reached the floor through a window covering under 60% of the title is never HIGH. The `Private` no-match cases are still refused. - **A removed book stays removed during a scan or a queued action.** Deleting an ingest book while a scan was running, or while a Retry was queued for it, re-created the row and could convert the torrent again: the dismissal was only checked when the scan listed its books. The pipeline now skips a dismissed book on entry, again before converting (also after waiting for the encoder) and before writing its row. Delete is refused (409) while the book is waiting for a manual action. A deliberate Retry, Use this ASIN, Re-check or Convert anyway on a removed ingest book brings it back. - **The healthcheck stays green while paused or during one long encode.** The heartbeat was only written between books, so a paused daemon or a single multi-hour encode read as stalled. It is now also written on every paused tick and, at most every 30 seconds, from the encode's progress. - **Match Audiobookshelf layout hardening.** The target folder is checked against `OUTPUT_ROOT` again after it is resolved (a symlinked author folder could point outside it); a file-only rename is skipped when another book is filed in the same folder, like a folder move; the undo journal records each renamed file as it is renamed, so a later failure no longer loses the earlier ones; and the target is worked out from the book's own `metadata.json` (title, authors, series, sequence) where present, falling back to the row, so a migration files the book where a fresh conversion would. - **Cover cache pruning is cheaper.** A prune asked the database once per kept cover whether a library book used it, all under the global State lock, and ran after every downloaded cover. The library's ASINs are now read once per prune (two queries), a download prunes at most every 25 new covers or 10 minutes, and `units(asin)` / `units(asin_override)` are indexed. - **Docs: Delete asks for confirmation in a dialog, not a typed word.** The README promised a typed confirmation; the modal has one **Delete permanently** button. The docs now say so. - **The title score uses the same best-window measure as the title floor.** `title_sim` is now `floor_sim` (best run of the folder title's words against the catalog title or title + subtitle) instead of a whole-string comparison, so folder baggage around a correct title no longer drags it down: `Dungeon Crawler Carl 01 - Dungeon Crawler Carl` scored 0.85 against "Dungeon Crawler Carl", graded MEDIUM and went to Review; it now scores 1.0 and grades HIGH. The floor, the HIGH/MEDIUM/LOW thresholds and the knockoff, Audnex and runtime rules are unchanged, and the folders that must stay no match (`Private`, `Private London`, `Private: #1 Suspect`) are still refused. - **The scan line says what is encoding.** A scan encodes eligible books inline and one encode can take a quarter of an hour, so `Scanning 37/41...` sat frozen and read as a hang. While a book is encoding the line now names it and shows its progress (`Encoding The Gate of the Feral Gods - 43%...`); `Scanning n/m...` is only shown when nothing is encoding. The label is built server-side (`scan_label` in `/api/state`). ## [0.19.0-dev11] - 2026-11-15 The last three findings of the dev4 security audit. ### Security - **Catalog and image reads are capped at 8 MB** (`catalog.MAX_BODY`, like the tracker client). An oversized answer is refused as a failure, not read into memory. - **The cover cache is pruned** whenever a cover is written: covers whose ASIN is no longer in the library, a catalog-seen result, a series listing or the wishlist are removed, and the cache is held to 2000 covers (oldest unprotected first). A library book's cover is never removed. - **`_attr` escapes `'`** as `'`. The CSP now also sets `object-src 'none'`, `base-uri 'none'` and `form-action 'self'` next to `frame-ancestors`. There is deliberately **no `script-src`**: the page is built on an inline script and inline `onclick=` handlers, so a `script-src` without `'unsafe-inline'` breaks every button and one with it protects nothing. ### Fixed - `_start_manual` returned `None` on one refusal path; it now returns `""` like the others. ## [0.19.0-dev10] - 2026-11-15 A default User-Agent, and a 400 that says what to do. ### Changed - **`TRACKER_USER_AGENT` now defaults to `audiobind/<version>`.** dev5 sent no User-Agent unless the variable was set, but a tracker can answer a request with none by HTTP 400 (proven live: no UA -> 400, any UA -> results), so a fresh install failed every search. An install that never set the variable now sends `audiobind/<version>`. A set value is still sent verbatim. The literal `none` (any case, spaces ignored) sends no User-Agent at all, so the anonymity option remains. urllib's own `Python-urllib` agent is still never sent, and the egress probe is unchanged. - **An HTTP 400 with no User-Agent says so**, for search and download: "the tracker answered HTTP 400. Some trackers refuse requests with no User-Agent - set TRACKER_USER_AGENT". With a User-Agent the message is unchanged. ## [0.19.0-dev9] - 2026-11-14 The search document can be sent form-encoded, not only as JSON. ### Added - **`TRACKER_SEARCH_ENCODING`** (`json` default, or `form`). `json` is unchanged. `form` sends the template as `application/x-www-form-urlencoded` `key=value` pairs: a list repeats its key, a nested object becomes `key[sub]`, booleans are `true` / `false`, numbers their decimal form, `null` is dropped, and the empty-value rule still omits a key that comes out empty. An unknown value stops startup with a message naming the variable. ## [0.19.0-dev8] - 2026-11-14 Finishes the rename: the search request and the qBittorrent category are settings now, nothing in the code speaks one site's dialect. Plus two ideas adopted from a comparable tool. ### Changed - **`TRACKER_SEARCH_BODY` replaces the built-in search request.** A JSON object with `{{title}}`, `{{author}}` and `{{category}}` placeholders (the category from `TRACKER_SEARCH_CATEGORY`), no default: the watcher stays off until it is set, and Settings > Watcher names it. Values are substituted into the parsed JSON, so quotes in a title cannot break the document. An object member or list item that is only placeholders and comes out empty (no author, no category) is left out. `POST` sends the document as a JSON body, `GET` as query parameters. The fixed English-language filter, sort order and page size are gone from the client; a site that needs them puts them in its template. `TRACKER_SEARCH_CATEGORY_PARAM` is removed (the template carries the key). - **`QBIT_CATEGORY`** (default `audiobooks`) is the qBittorrent category torrents are added to; it replaces the hardcoded category and is shown on Settings > Watcher. Set it to your existing category. - Example torrent names in the backlog, tests and verify page are neutral. ### Added - **`WATCH_MAX_INCOMPLETE`** (default 10, `0` = off): a cycle that finds that many torrents of the qBittorrent category still below 99.9 % searches and adds nothing and logs one line. The `downloaded` flip and the library match are never held; a qBittorrent error while counting is logged once and the cycle carries on as before. - **Format preference:** between candidates that already qualify, an m4b/m4a source is picked over an mp3. Series match, title score and the freeleech preference all rank above it. ## [0.19.0-dev7] - 2026-11-14 The acquisition client no longer knows which site it talks to. The site's URL, endpoints, session cookie and category / freeleech parameters are settings you fill in; nothing in the code, the UI, the logs or the new docs names a site. ### Changed - **The client module is now `audiobind/tracker.py`** (no compatibility import). The client, error and backoff classes are `TrackerClient`, `TrackerError` and `TrackerBackoff`, and the freeleech decision is `freeleech_decision()`. The URL guard (`authorized_url()`) still refuses anything but the two configured endpoints, now compared against the configured values. - **New settings, all optional and empty by default:** `TRACKER_URL`, `TRACKER_SEARCH_PATH`, `TRACKER_SEARCH_METHOD` (`POST` or `GET`), `TRACKER_SEARCH_CATEGORY`, `TRACKER_SEARCH_CATEGORY_PARAM`, `TRACKER_DOWNLOAD_PATH`, `TRACKER_DOWNLOAD_PARAM` (default `tid`), `TRACKER_FREELEECH_PARAM` (empty: never sent), `TRACKER_COOKIE_NAME`, `TRACKER_SESSION`, `TRACKER_VIP`, `TRACKER_PROXY`, `TRACKER_USER_AGENT`, `TRACKER_HOME_IP` and `TRACKER_EXPECTED_EGRESS_IP`. The watcher is on only when the URL, both paths, the cookie name and the session are set; otherwise it is off and Settings > Watcher names what is missing. A malformed URL, path, method or name stops startup with a message naming the variable. - **Search results are read through one alias table** (title / name, author_info / authors, numfiles / files, filetype / format, free / freeleech, personal freeleech as a boolean or a datetime, and so on). Anything unrecognised is simply absent, never an error. - **Wording:** the watcher, the Wishlist notes, the egress notices ("the tracker would see ..."), settings hints and log lines say "the tracker" and "a freeleech credit"; the membership setting is labelled "Premium member". - The wishlist table's torrent-id column is now `tracker_tid`; existing rows are copied over on the first start (the old column is left in place). ### Upgrading - **Set the endpoints.** The search path, download path, cookie name and category have no old variable to inherit from, so an existing install's watcher stays off ("not set: ...") until they are configured. The log and Settings > Watcher say which. - **Renamed variables keep working** while the new name is empty, with one startup log line each (never the value): the six variables that existed before are accepted under their old names and report `<old> is deprecated, use <new>`. When both are set the new one wins. ## [0.19.0-dev6] - 2026-11-14 Documentation and repository polish for the public launch. No behaviour change. ### Documentation - **README:** new "What you get" list and "Requirements" block after the intro; the wishlist lifecycle is described as it works (wanted, downloading, downloaded, then the row is removed once the converted book is in the library; not found and error rows stay until Retry or Remove); `/healthz` documented as the public liveness endpoint; a "Behind a reverse proxy" section for `TRUSTED_PROXY` + `TRUSTED_PROXY_IPS` with the dev5 migration; the Docker examples use `:latest` instead of the long-outdated `0.15.1` tag; "never deletes your sources" now says what is true (conversion never touches them, only the explicit Delete does). - **Screenshots** retaken from the redesigned UI: library (`docs/screenshot.png`, same name for the Community Applications template), wishlist and Discover books. The two November 2 screenshots that nothing referenced are gone. - **Template, CA profile and forum post:** removed the stale "AAC 64 kbps" and "input is mounted read-only" claims. ### Repository - New `SECURITY.md` and a bug report issue form. - Removed two unreferenced icons from other projects and the host-specific Unraid deploy scripts and template snapshot under `docs/`. ## [0.19.0-dev5] - 2026-11-14 Security-fix build from the 0.19.0-dev4 audit. Docker users: if you set `TRUSTED_PROXY=true`, **also set `TRUSTED_PROXY_IPS`** (see Security) or your proxy's forwarded headers stop being honoured. ### Security - **Tracker requests no longer send a Python User-Agent.** urllib's default `User-agent: Python-urllib/x` was going out on every search and download, contradicting the documented behaviour. No User-Agent is sent unless `TRACKER_USER_AGENT` is set, then exactly that one. The same fix applies to the egress probe. - **The environment's `HTTP(S)_PROXY` is no longer consulted** by the tracker client or the egress probe. Without `TRACKER_PROXY` the request goes direct; with it, only through that proxy. - **A residential IP address and a LAN layout were removed from the repository**: tests now use documentation addresses, and the host-specific deploy/rollback scripts and the stored container-args file are gone (a test keeps the address out). `docs/PUBLISHING.md` describes the release steps generically. - **Forwarded headers are only trusted from listed proxies.** New `TRUSTED_PROXY_IPS` (comma-separated IPs/CIDRs). `X-Forwarded-For`, `-Host` and `-Proto` are honoured only when `TRUSTED_PROXY=true` **and** the connecting peer is in that list; `TRUSTED_PROXY=true` alone no longer trusts anyone. **Migration:** add your reverse proxy's address, for example `TRUSTED_PROXY_IPS=172.18.0.0/16`. - **The failed-sign-in limit holds under parallel requests.** Each attempt is counted before the password is checked and cleared on success, and the password hashing is bounded to a couple of concurrent runs. This covers the sign-in form, Bearer requests and Settings > Account. - **Concurrent connections are capped** (`MAX_CONNECTIONS`, default 64; extra connections get a 503), and request headers must arrive within 10 seconds in total. - **The manual-action cap works again**: a manual action released its slot twice. - **Sign out is POST-only** and goes through the same origin check as other POSTs. - **The session cookie gets `Secure`** when the request arrived over https through a trusted proxy (`X-Forwarded-Proto: https`). - **Legacy `AUTH_TOKEN` sessions end when the token changes**, restarts included. - **The container refuses `PUID=0` / `PGID=0`** unless `ALLOW_ROOT=1` is set. - **Control characters in a title or override** can no longer break publishing or put a line break in a folder name. - **The template and compose file no longer say an empty `AUTH_TOKEN` means no sign-in.** With it empty the first visit opens the account setup page, which anyone who can reach the port can complete until someone does: finish setup right after the first start. ### Fixed - **A blocked egress guard no longer skips local bookkeeping.** The qBittorrent progress step (`downloading` to `downloaded`) now runs before the guard and regardless of its verdict; it makes no tracker request. - **`/healthz` is reachable without a session** once an account exists, for uptime monitors. It only ever answers `ok` or `stale`. Nothing else became public. - **An audiobook with more than 255 chapters** no longer gets a malformed MP3 chapter list: the table of contents lists the first 255. ### Tests - The User-Agent and proxy tests assert what goes over a real local socket. Added: slot release through the manual path, concurrent failed sign-ins, the connection cap and header deadline, trusted-proxy peers, sign out, the `Secure` flag, token-bound sessions, `/healthz`, the entrypoint's root refusal, the 300-chapter table of contents, control characters in names, the guard-blocked bookkeeping, and a check that the residential address never returns. ## [0.19.0-dev4] - 2026-11-14 ### Fixed - **The title floor compares against the best contiguous run of words, not the whole name.** A long, typo'd folder such as `Steeleart Reckoners Brandon Sanderson` scored 0.38 against the right book (*Steelheart*) and would have been refused; its best window `Steeleart` scores 0.95. The floor (still 0.60) now takes the maximum `sim` over every run of consecutive whitespace-separated words of the wanted title (the whole string is one of them). The held note quotes that window score. The wrong matches stay refused; HIGH/MEDIUM/LOW thresholds, knockoff rules, the Audnex tie-break and the runtime check are untouched; existing rows are not rewritten. ### Tests - Steelheart regression, the three Patterson cases and `Partials` still refused, single-word titles unchanged, the note quotes the window score. ## [0.19.0-dev3] - 2026-11-14 ### Fixed - **A catalog hit with an unrelated title is no match, not a LOW match.** `Private`, `Private London` and `Private: #1 Suspect` had been adopted as *Step on a Crack*, *The Country Road Murders* and *Judge Stone*. If the best candidate's title similarity (the matcher's own `sim`) is below 0.60, the unit is now no match and goes to `held` with a note such as `no catalog match: closest was 'Step on a Crack' (title 0.18, needs 0.60)`. Candidates at or above the floor grade exactly as before (thresholds, knockoffs, Audnex tie-break and runtime check untouched). Library and ingest units share the rule. Existing rows are not rewritten; only new scans change. - **A finished torrent no longer reads DOWNLOADING.** Each watcher cycle asks qBittorrent (one read-only `GET /api/v2/torrents/info`) for the progress of every `downloading` row's torrent; at 100% the row becomes `downloaded` ("downloaded - waiting for the converted book"). `downloaded` is never searched again, still counts on the Wishlist badge and in "N books wanted", shows a DOWNLOADED badge, keeps Remove and has no Retry. The row is still deleted when the converted book lands. A qBittorrent failure is logged once and leaves the rows alone. ### Tests - No-match cases for the three real books plus the two must-still-match cases; progress 1.0 / 0.4 / no hash / qBittorrent error; `wishlist_due`, badge count and page render for `downloaded`. Not verified here: a live qBittorrent. ## [0.19.0-dev2] - 2026-11-14 ### Fixed - **MP3 argv put `-map 0:a:0` before the cover `-i`.** ffmpeg binds an output option that precedes a later input to that input and refuses ("cannot be applied to input url"). The mp3 command now lists every input first, then the maps. The m4b argv is unchanged. ### Changed - The CTOC flags byte (0x03, top-level + ordered) is now documented in `id3.py`. ### Tests - New structural test: no `-map` may appear before the last `-i`, for mp3 with and without a cover and for m4b. Not yet verified: a real ffmpeg run of the mp3 path. ## [0.19.0-dev1] - 2026-11-14 ### Added - **Output format: M4B, MP3 or both.** A new Settings choice (`OUTPUT_FORMAT`, default `M4B`, so existing installs change nothing). `MP3` publishes one `.mp3` per book (not one per chapter) encoded with libmp3lame at the same bitrate rules and channel settings, with the usual tags and the cover as an APIC frame. `M4B + MP3` publishes both in the same book folder, the m4b first; the unit's output is the m4b. - **MP3 chapters.** ffmpeg silently drops chapters when it writes an mp3, so Audiobind now writes them itself (`audiobind/id3.py`): CHAP frames, one per chapter, plus a CTOC frame are appended to the ID3v2.3 tag ffmpeg wrote, every existing frame kept byte for byte. A file with no tag gets a fresh v2.3 tag; a v2.4, unsynchronised or extended-header tag is replaced by a fresh one rather than corrupted. ffprobe-based players (Audiobookshelf, Plex) read these chapters; Apple's Music/iTunes does not. - Verification runs for mp3 as for m4b: an mp3 audio stream, duration within the usual tolerance, the chapter count, and the cover. A failing file is never published. ### Behaviour - With `M4B + MP3`, both files are encoded and verified in the work directory first and published only when both pass. If the second encode fails, the unit fails and nothing is published: the library never holds an m4b without the mp3 that was asked for, and the retry starts clean. - The series shelf recognises an `.mp3` output path as well as `.m4b`. - `docs/verify-page.py` and the unit tests cover the new setting and the tag writer without ffmpeg. ## [0.18.0] - 2026-11-14 The first public release since 0.15.1. It folds every banked build from `0.16.0-dev1` through `0.18.0-dev12`; the per-build detail is kept as subsections below. **Acquiring books.** The wishlist watcher talks to a private tracker: it searches the one authorised search endpoint, downloads a book with a freeleech wedge only when the torrent is not already free, and hands it to qBittorrent. Off unless `TRACKER_SESSION` is set. It refuses to run unless the tracker would see your VPN address (a home-IP denylist and an optional expected-IP allowlist), sends no User-Agent, and can be pointed through an HTTP proxy. How many books a cycle takes, and how often it runs, are both editable in Settings > Watcher. **The wishlist is a want-list.** Its own rail tab with a count, and a book that lands in your library leaves the list instead of lingering. A row can carry its own search title, so a book the tracker spells differently can be fetched without lowering the match bar for everything else. **Discover books.** Search and recommendations on one screen, no chips: suggestions come from the authors, narrators and genres you already own. Series cards say how much of a series you have and can add every missing book in one click. **Ingest.** A read-only folder of finished torrent downloads is converted, matched and filed exactly like your library, with a fallback to the files' own tags when the folder name matches nothing, torrent-name cleaning, and single-file torrents found. Audiobind never writes to it, so the torrent keeps seeding. **Matching.** Summaries, study guides and review knockoffs are recognised and never promoted to HIGH, and Audnex settles close calls between near-identical candidates. **Your own account.** First-run setup creates a username and password; sessions survive a restart; "Keep me signed in for 30 days"; `audiobind reset-password` for a forgotten one. `AUTH_TOKEN` still works on an install that has no account yet, and is deprecated. **Logging.** A rotating log file, on by default at `/config/audiobind.log`. ### 0.18.0-dev12 (2026-11-14) ### Added - **A wishlist row can carry its own search title.** The new **Edit title** action (next to Retry / Remove, on every row) opens an inline input pre-filled with the effective search title. The watcher searches `search_title or title`, and the 0.90 title score compares the override against the tracker result, so a book whose catalog title near-missed (`Missing` vs `The Missing`, 0.85) can be re-titled to the tracker title and land. The row keeps showing the catalog title and notes `searching for "..."` while an override is set. - `POST /api/wishlist/title` with `{"asin", "title"}` (same auth and cross-site guard as the other wishlist endpoints) returns the row's public shape, now including `search_title`. A blank or whitespace-only title clears the override. ### Changed - Saving or clearing a search title puts the row back to `wanted` with the note and miss count cleared, so the next cycle searches it immediately. The display title never changes. - The state DB gains a nullable `wishlist.search_title` column through the usual add-column migration; an existing database upgrades in place and loses nothing. ### 0.18.0-dev11 (2026-11-14) ### Added - **The watcher schedule is editable in Settings > Watcher.** "Check the wishlist every" (10-1440 minutes) and "Books per cycle" (1-20) sit above Run a cycle now, with the usual "custom (was N from env)" mark. `WATCH_INTERVAL_MIN` and `WATCH_MAX_PER_CYCLE` remain the defaults and are shown when nothing is saved. A saved value applies from the next cycle, no restart: every cycle reads the effective value. With the tracker unconfigured the Watcher block, and so both fields, stays absent. ### Changed - The politeness clamp stays the enforcement: a saved value past a bound (0 or 99 books, 1 minute) is pulled back to the bound (1 / 20 / 10) and that is what is stored, never the raw number. The field bounds equal the clamp bounds and are not widened by an unusual env value. - Saving either field reloads the page so the Watcher status line ("every N min, at most M books per cycle") shows what the next cycle will do. ### 0.18.0-dev10 (2026-11-14) ### Added - **Wishlist is its own rail tab**, directly below Discover books, at `/wishlist`, with the wishlist count badge. With the tracker unconfigured there is no Wishlist item and `/wishlist` shows the library page. `/discover/wishlist` still works as an alias. ### Changed - **Discover is renamed "Discover books"** (rail, page title and heading; the route stays `/discover`) and is one screen: the search form on top, the recommended books below it. A search shows its results on the same page in place of the recommendations; clearing the query brings them back. The "Search", "For you" and "Wishlist" chips are gone (`/discover/for-you` shows the same page). - **A wishlist book that reaches the library leaves the list.** The library match now deletes the row (one log line per removal) instead of marking it `in_library`; rows that do not match (`wanted`, `not_found`, `error`) are never touched, and the removal is idempotent. Leftover `in_library` rows from earlier versions are cleaned up the same way. ### 0.18.0-dev9 (2026-11-14) ### Added - **"Keep me signed in for 30 days"** on the sign-in form (unticked by default, for that browser only). Ticked, the session and its cookie `Max-Age` are 30 days (2592000); unticked is the usual 7 days. The lifetime is per session (`Sessions.create(version, ttl=...)`). - **Remembered username:** a successful sign-in sets an `ab_user` cookie (`Path=/`, 30 days, `HttpOnly`, `SameSite=Strict`) that pre-fills the username field. It never authenticates anything, and signing out leaves it. ### Changed - **Sessions survive a restart.** They are now kept in a `sessions` table in the state DB (a SHA-256 of the session id, its expiry and the credential version - never the id itself), loaded at startup, with expired rows pruned and the 256-session cap kept (oldest dropped). A password or username change still ends every other session, and a session from before a change stays invalid. ### 0.18.0-dev8 (2026-11-14) ### Changed - **The main-page Dry run no longer silences the tracker watcher.** One switch used to do two unrelated jobs: stop the library pipeline writing files, and make the watcher search without downloading. They are now two switches. The main-page **Dry run** button only previews the library scan. The watcher has its own **Dry run** in Settings > Watcher (searches and reports, never downloads), kept across restarts, and the two can be on or off in any combination. - New `WATCH_DRY_RUN` environment variable (default `false`). Like `DRY_RUN`, `true` always starts that switch on whatever was saved; the UI can turn it off afterwards. Startup logs one line per switch. `POST /api/control` accepts `{"action": "watcher_dry_run", "value": true|false}`; `dry_run` keeps meaning the library only. - `GET /api/watcher` and the watcher's cycle summary report `watcher_dry_run` instead of `dry_run`. - **Upgrading:** if you relied on the main-page Dry run to keep the watcher from downloading, turn on the watcher's Dry run (or set `WATCH_DRY_RUN=true`). ### 0.18.0-dev7 (2026-11-14) ### Added - **A log file that survives a container recreate.** `LOG_FILE` (default `<STATE_DIR>/audiobind.log`; empty string = off) gets a rotating file log, 5 MB x 3 backups, UTF-8, same format and level as the console. stdout is unchanged. A file that cannot be opened costs one console warning and nothing else. Every CLI subcommand goes through the same setup. - README: "Bring your own downloader". ### Changed - **A non-tracker install no longer shows tracker surfaces.** Without `TRACKER_SESSION`: no Wishlist tab (and `/discover/wishlist` lands on Search), no wishlist count in the rail, no Get this / Want / Want all missing buttons (Discover and Series), and Settings > Watcher is one short block explaining that the tracker is optional and which variables turn it on. The rail says that files in the ingest folder are converted whether or not the tracker is configured. ### 0.18.0-dev6 (2026-11-14) ### Fixed - **A successful qBittorrent 5.x login is no longer reported as a failure.** 5.x answers 204 and names its session cookie after the WebUI port (`QBT_SID_8083`), not `SID`, so the check failed and the watcher refused to download. Either name is now accepted (a non-empty value is still required; no cookie or `Fails.` is still a failure). - **`DRY_RUN=true` now always starts dry.** The saved switch used to override it, so a container started with `DRY_RUN=true` could still run a live cycle. The environment is now a floor: with `DRY_RUN=true` the watcher starts dry whatever was saved (the UI switch can turn it off afterwards, and that choice persists). One startup line says which source won. ### 0.18.0-dev5 (2026-11-14) ### Added - **`TRACKER_VIP`** and **Settings > Watcher > "VIP member"** (No / Yes, default No, saved in the database, effective on the next cycle). The tracker marks some torrents VIP-only (`vip=1` in the search results); only a VIP member can download them. VIP status cannot be read from the API (`vip_expire` is `0` even for a VIP account), so it is a setting. ### Fixed - **A non-VIP install no longer tries to download VIP-only torrents.** They are skipped, and when nothing else qualifies the book's note reads "only match(es) need VIP membership (N) - skipped". With VIP on, they are normal candidates; they are VIP freeleech, so the wedge decision is unchanged (no wedge spent). `tracker.select_candidate` gained `vip: bool = False`. Still one request per book. ### 0.18.0-dev4 (2026-11-14) ### Fixed - **The wishlist watcher searched the wrong tracker category and found nothing.** It sent `tor[cat][]=14`, but `tor[cat][]` takes a leaf category id and the tracker answers "Nothing returned" for 14 on every query, so books that are plainly on the tracker were reported "not found". It now searches the audiobook main category (`tor[main_cat][]=13`). Still one request per book. ### 0.18.0-dev3 (2026-11-14) ### Added - **`TRACKER_EXPECTED_EGRESS_IP`**: an optional allowlist for the egress guard. When set, the tracker must see exactly this address (your VPN's exit IP) or the cycle is refused with **zero tracker requests**, because a tracker session is tied to one IP. Empty = off (the default); a value that is not an IPv4/IPv6 address stops startup with a message naming the variable. - **Settings > Watcher, "Only reach the tracker from one IP"**: the same value as a text setting, saved in the database and effective on the next cycle with no restart. The environment value is the default until something is saved; blank turns the check off; an invalid address is refused with a clear message and not stored. "Reset to defaults" hands back the environment value. - Guard order: (1) the home IP, when it matches: blocked with the home-connection reason; (2) expected set and different: blocked, "the tracker would see X, but this installation only allows Y"; (3) expected set and equal: ok; (4) nothing armed: off; unverifiable: unknown (fail closed). The probe now runs when either check is armed. - `GET /api/watcher` gains `egress_expected`, and `egress_reason` says which rule produced the verdict. Settings > Watcher shows the address the tracker sees, the expected one, the verdict, and what to change when blocked. ### Changed - The blocked notice on Settings and the Wishlist bar now quotes the guard's reason rather than always saying "home connection", since a block can come from either check. ### 0.18.0-dev2 (2026-11-14) ### Added - **`TRACKER_PROXY`**: `http://host:port` or `https://host:port` (an HTTP CONNECT tunnel for the tracker's https). Search and download share one proxied opener. **Fail-closed:** a failing proxied request is an error and is never retried or sent directly. A malformed value stops startup with a message naming `TRACKER_PROXY`. SOCKS is not supported (standard library only). - **Egress guard, `TRACKER_HOME_IP`**: at the start of every watcher cycle (dry run included), before any tracker request, the IP the tracker would see is looked up (`ipinfo.io`, then `api.ipify.org`, through `TRACKER_PROXY` if set). Equal to the home IP: the cycle is refused. Lookup failed: refused too (fail closed). Either way **zero tracker requests**. Empty = off, no lookup at all. `GET /api/watcher` gains `egress_ip`, `egress_checked_at`, `egress_guard` and `egress_reason`; Settings > Watcher shows what the tracker sees, when it was checked and the guard, with a notice when blocked, and the Wishlist bar says the watcher is stopped because the connection is not the VPN. - **README "Running the tracker through a VPN"**: share a VPN container's network namespace (any VPN container, not just Gluetun), an HTTP proxy, and the tripwire, with a way to verify it. `docker-compose.yml` gains a commented-out VPN example and the Community Applications template gains the watcher, proxy, guard, qBittorrent and ingest settings. - `docs/live-args.json` and `docs/rollback.sh` now describe a container in a VPN network namespace (`network_mode`, no published ports); the rollback refuses to run while that container is not running. ### Changed - **No fingerprint.** Tracker requests no longer carry `User-Agent: audiobind/<version>`; no User-Agent override is sent unless `TRACKER_USER_AGENT` is set, which is used verbatim. - `TRACKER_HOME_IP` that is not an IP address stops startup, naming the variable. ### 0.18.0-dev1 (2026-11-14) ### Added - **The wishlist watcher (a private tracker).** Off unless `TRACKER_SESSION` is set; with it unset no thread starts and nothing touches the network. With `TRACKER_SESSION` and the qBittorrent WebUI login (`QBIT_USER` / `QBIT_PASS`, at `QBIT_URL`, default `http://localhost:8083`) a background thread works through the wishlist every `WATCH_INTERVAL_MIN` minutes (default 60, 10-1440): up to `WATCH_MAX_PER_CYCLE` wanted books (default 5, 1-20), never-checked first and oldest first, then the least recently checked. Each is searched on the tracker's Torrent Search JSON API (audiobooks, English). A match needs a title similarity of 0.90 or more, a matching author when the book has one, audio files and not already snatched; files present, the exact series and sequence, and an already-free torrent are preferred. The torrent is downloaded through the tracker's Download Torrent API and handed to qBittorrent (the configured category), and the book becomes **downloading**. - **Freeleech wedges are spent only when needed.** A torrent that is already free, VIP freeleech or a personal freeleech is downloaded without `&fl`; any other is downloaded with it. Exactly one decision per torrent, written into the book's note. - **Wishlist states:** wanted, downloading, in library, not found and error, moved only along a fixed set of transitions. A book with no good match stays wanted with a "not found on the tracker (checked ...)" note and is searched again later; after 3 misses in a row it becomes not found. Nothing below the threshold is ever forced. **Retry** puts a not-found or failed book back to wanted. - **Books resolve on their own.** After every scan (watcher on or off), a wishlist book whose ASIN, or title and first author, matches a converted book becomes **in library**. - **Wishlist tab:** a state chip, when the book was last checked, the watcher's note, and Retry / Remove on each row, plus a line saying whether the watcher is off, disabled, paused or on (with its next cycle and a dry-run warning). - **Settings > Watcher:** status (enabled, off, disabled, paused, dry run), schedule, qBittorrent host, last and next cycle, last error and any tracker backoff. **Run a cycle now** runs one cycle in the background; **Pause watcher** is kept across restarts, like the conversion pause. Missing qBittorrent credentials show a warning here and in the log; the watcher stays disabled instead of failing. - `GET /api/watcher`, `POST /api/watcher` (`run`, `pause`, `resume`) and `POST /api/wishlist/retry`. `/api/wishlist` rows now include `note`, `checked_at`, `updated_at`, `tracker_tid` and `torrent_hash`. ### Safety - The watcher calls only the two tracker endpoints the tracker authorizes for automation, checks every URL before sending it, never parses HTML and never follows a redirect. - One tracker request at a time, at least 3 s apart. Any non-200, an unreachable host or an answer that is not what the endpoint documents starts a backoff (60 s, doubling to 30 minutes; a longer `Retry-After` is honoured) and ends the cycle. The backoff survives a restart, and a restart never fires a cycle at once. - qBittorrent is logged into before anything is downloaded, so a qBittorrent that is down never costs a wedge. A failed download or a qBittorrent refusal marks the book error and is never retried on its own (wedges are not refunded); a Retry searches again first. - Dry run applies: the watcher searches and notes what it would download, and downloads nothing. - `TRACKER_SESSION` and `QBIT_PASS` are never shown in a page, the API, the log or a `repr()`. ### Changed - The Discover tab's Wishlist list shows the state chips above instead of the bare status. ### 0.17.0-dev1 (2026-11-13) ### Added - **A wishlist.** A new `wishlist` table in the state DB stores ASIN, title, authors, when it was added and status `wanted`. Adding a book that is already there changes nothing. `GET /api/wishlist` lists it, `POST /api/wishlist` adds a book and `POST /api/wishlist/remove` takes one off. Books go on it only when you click. Nothing is downloaded, and no tracker is contacted. A later version picks books up from here. - **Discover tab, Search.** A new rail item between Series and Settings searches the Audible catalog by title, with an optional author. Each result shows its cover, authors, narrators, runtime, series and release year. A book already in your library is marked **In your library**: same ASIN, same title and first author, or marked "I have this". Knockoffs carry their badge. **Get this** adds a book to the wishlist. Searches go one request at a time through the catalog's usual pacing, and only the US store (`api.audible.com`) is searched. Searching several regions is not in this build. `GET /api/discover/search` returns the same results as JSON. - **Discover tab, For you.** Suggestions based on what you own: more by your top 5 authors, books read by your top 3 narrators, and your top 2 genres (from the `metadata.json` beside each published book). Books you own or already want are left out, and so are knockoffs, other languages and second editions. A result only counts when the seed author or narrator is really credited on it. Suggestions are stored with the time they were fetched. Opening the page reads the stored list and never queries the catalog. A refresh runs in the background when nothing is stored yet, when the list is more than 7 days old, or when you press **Refresh**. One refresh makes at most 10 catalog requests. If none of them reach the catalog, the last good list is kept. `GET` / `POST /api/discover/recommendations` give the stored list and start a refresh. - **Series completion.** Each series card shows "owned X of Y" and says the total comes from the Audible catalog, with when it was checked. The missing list (title, sequence, runtime) has a **Want** button on each book. **Want all missing** adds every missing book in the series in one click and one transaction, skipping any already wanted. It uses the server's current missing list, so a book you got after the page loaded isn't added. `POST /api/wishlist` with `series` and `"missing": "all"` does the same. ### Changed - On the Series page, **Want** is now the main button on missing books. **Copy Audible link** is still beside it. - `/api/series` now says where each series total comes from (`source`: `catalog`, `pending`, `not_found` or `unreachable`). It also includes `last_error`. ### Fixed - **A series with no total now says why.** Before, a series card said "fetching series size…" even when the catalog had answered that no such series exists, or couldn't be reached. Now the card shows **total unknown** and the reason: still fetching, Audible has no listing, or the catalog couldn't be reached and will be tried again. The detail panel's series block says the same. ### 0.16.0-dev4 (2026-11-13) ### Added - **Knockoffs are recognised.** A catalog entry counts as a knockoff, not an edition, when its title or subtitle says summarized, summary, study guide or workbook, or starts with `[Review]` or `Review:`. An entry with no runtime also counts when another candidate in the same search has a real one. The case of the words does not matter. - **Audnex settles close calls.** When another candidate scores within 0.10 of the best, Audiobind asks Audnex about the best and up to three of those close rivals. A candidate Audnex agrees with (same title and author, runtime within 5%) beats one it does not agree with. A rival Audnex does not back can no longer narrow the margin of a pick it does back. When nothing is that close, nothing is looked up, so a book with no Audnex record grades exactly as before. Some ASINs have no Audnex record in a region. - **Knockoff badge in the detail panel.** Knockoffs still show among the other candidates, with a Knockoff badge and the reason: a summary, review or catalog stub never counts against the pick. "Use this edition" still works on them. When Audnex backed the pick, its card says "Audnex agrees". ### Changed - **A best match that is itself a knockoff grades LOW.** It is still matched and shown, but a person decides. LOW goes to review under every auto-convert setting, "medium" included. To keep a summary edition you really want, pin its ASIN or use it from the other candidates. ### Fixed - **A knockoff no longer costs the real edition its HIGH grade.** The margin a pick needs (0.05 for HIGH) is now measured against the best real candidate of a different work. Knockoffs are left out. With no real runner-up the pick is decisive. *The Butcher's Masquerade* (Matt Dinniman) was an exact match: score 1.0, runtime 23.57 hrs against 23.55. It was parked at MEDIUM behind "[Review] The Butcher's Masquerade (Matt Dinniman) Summarized." Now it grades HIGH and converts on its own. ### 0.16.0-dev3 (2026-11-12) ### Added - **Ingest falls back to the files' own tags.** If an ingest book's name matches nothing in the catalog (confidence NONE), Audiobind reads the embedded tags of its first audio file and searches again. The title comes from `title`, or from `album` first when there are several files. The author comes from `album_artist`, `artist` or `composer`. This is one read-only ffprobe. If the second search finds something, its match is used, and the usual grading and auto-convert tiers still apply. The detail records which names won (`name_source`: `folder` or `embedded`). The "why this match" panel and the row note say when the tags were used. Library books, ingest books whose name already matched, and books with a pinned ASIN never read tags. ### Changed - **Torrent names are cleaned before an ingest book is matched.** Tags in square brackets or braces are dropped with their contents (`[CA]`, `[tracker v5]`, `{ASIN:...}`). So are Unabridged and Abridged, a leading sequence number (`01 `, `03. `, `Book 2 - `, `Book 1: `), and extra spaces. Numbers that are the title stay: `1984` and `2001 A Space Odyssey` survive. The author is still read the way it was. Only ingest books are cleaned. Library folders are parsed exactly as before. The state key and the path on disk keep the raw torrent name. ### Fixed - **Single-file torrents are found.** A torrent of one audio file lands as a bare file in `INGEST_DIR`, for example `The Guilty Will Robie, Book 4 (Unabridged).m4b`. These were never discovered, because only folders counted. Now every audio file directly in the ingest root is a book of its own (m4b, m4a, mp3, flac, ogg, opus, wav and the other audio types the library already knows; the extension's case does not matter). It is keyed on the file name and its title is the file name without the extension. It goes through the same match, convert and publish steps. Hidden and IGNORE-listed files are skipped. Two files side by side are two books, never grouped. The file itself is only read. - **`01 Convergence [CA]` was searched as "Convergence CA"** and matched nothing. It is now searched as "Convergence" (see Changed). ### 0.16.0-dev2 (2026-11-11) ### Added - **Your own account: username and password.** The first time you open Audiobind it asks you to create one: a username (1-64 characters) and a password (at least 8, typed twice). Creating it signs you in. There is only one account. Once it exists the setup page redirects to the library and a second setup is refused. Until it exists, every page goes to the setup page. Only the setup page, the icon and `/healthz` answer. - **Sign-in asks for both fields.** A wrong username and a wrong password get the same message ("That username or password is not right."), and the server does the same work for either, so a guess never learns which half was wrong. Five failed sign-ins from one IP still lock that IP out for 10 minutes. Sessions still last 7 days. - **Settings → Account:** change the password (current password, new, confirm) or the username (current password required). The browser you do it from stays signed in. Every other browser is signed out. Wrong current passwords count towards the same lockout. - **`audiobind reset-password`** for a forgotten password: `docker exec -it audiobind audiobind reset-password`. It asks for the username (Enter keeps the current one) and a new password twice, hidden. It replaces the account, or creates it on a fresh database, and signs every browser out, even ones signed in to a daemon that is already running. `--username NAME` skips the username question. Without a terminal it refuses unless `--username` is given, and then it reads the password from stdin. There is deliberately no reset link on the web page. - Scripts can send the account password as `Authorization: Bearer <password>`. A password that was already accepted is remembered (in memory, until it changes), so a polling script does not cost a full password check on every request. ### Changed - **`AUTH_TOKEN` is deprecated, but existing installs are not locked out.** While it is set and no account exists, the setup page is skipped. You sign in with any username (for example `admin`) and the token as the password, and `Bearer <token>` keeps working. Settings → Account offers to create the account, confirmed with the token. As soon as an account exists it takes precedence: the token no longer signs in, and sessions that signed in with it end. - The password is stored as a salted PBKDF2-SHA256 hash (600,000 rounds, a fresh 16-byte salt every time it is set) in a new `credentials` table in the state database. The table is created automatically. Passwords, hashes and Bearer values are never logged or returned by the API. - The lockout message now reads "Too many failed sign-ins" instead of "Too many wrong tokens". A wrong Bearer is logged as "wrong API credentials". ### Fixed - An install without `AUTH_TOKEN` no longer serves the library, the controls and Delete to anyone who can reach the port. It asks for an account first. ### 0.16.0-dev1 (2026-11-10) ### Added - **Ingest - convert finished torrent downloads.** A new setting, `INGEST_DIR`, points Audiobind at a folder of finished downloads (mount it read-only, for example as `/ingest`). Each folder directly inside it is treated as one book. It is matched, converted and filed into your output library exactly like a book from your library: the same confidence rules, the same runtime check, the same Review and Held decisions, the same dry run. Disc and part folders inside a download are combined into one book, the same as in your library. Loose files sitting in the folder itself and hidden folders are ignored. Audiobind only ever reads this folder. It never writes, moves or deletes anything there, so the torrent keeps seeding. Leave `INGEST_DIR` empty (the default) and nothing changes. - Books from the ingest folder show a **⇣ ingest** badge, in the same table with the same columns as your library books. Type "ingest" in the search box to see only them. One that needs a decision lands on the Needs attention shelf with the usual reason and buttons. Its details show the folder it came from, for example `/ingest/Andy Weir - Project Hail Mary [tracker]`. - `/api/units` now says where each book came from (`"source": "library"` or `"ingest"`). `/api/state` adds `ingest_dir` and `ingest_enabled`. ### Changed - **Delete won't touch a download.** Delete on an ingest book is refused with "delete is for library sources; remove the torrent in qBittorrent". The files belong to a torrent that is still seeding. Nothing is deleted, not even the published copy. Library books delete exactly as before. - If `INGEST_DIR` is set but the folder doesn't exist when Audiobind starts, a warning is logged and ingest stays off until the next restart. Your library keeps converting. The same happens if the folder is, or sits inside, your output or library folder. - If a download is removed between scans, its row stays as it was. If it is removed while Audiobind is reading it, that one book is skipped and the rest of the scan carries on. - The database gains a `source` column. It is added automatically the first time this version starts, and every existing book is marked as a library book. ### Fixed - Nothing in this build. With `INGEST_DIR` unset, Audiobind behaves exactly as 0.15.0-dev4 did. ## [0.15.1] - 2026-11-11 Every recommendation from the v0.15.0 security review is in: 1 high, 5 medium, 10 low and 3 info findings. ### Before you upgrade - action needed - **Set `ALLOWED_HOSTS`, or the web UI only answers on the server itself.** Add the address you type into the browser, for example `ALLOWED_HOSTS=localhost,127.0.0.1,192.168.1.10,tower,tower.local` (ports don't matter). Until it's set, the log says so on start-up and other machines get a short "Host is not allowed" page. Unraid: the template has a new **Allowed hosts** field. ### Security - **Websites you visit can no longer reach Audiobind through DNS rebinding.** Every request, page or API, GET or POST, must be addressed to a name in `ALLOWED_HOSTS`; anything else gets a 403. Before, a hostile page could point its own domain at your server's LAN address, read your library and press Delete on every book. `X-Forwarded-Host` is only believed with the new `TRUSTED_PROXY=true`, for a reverse proxy you run yourself. - **Optional sign-in.** Set `AUTH_TOKEN` and the web UI asks for it once per browser (the session lasts 7 days; there's a "sign out" link in the side rail). Scripts send `Authorization: Bearer <token>` to `/api/...`. Five wrong tokens from one address lock it out for 10 minutes. Leave it unset to keep today's open behaviour. - **Delete never takes another book's converted copy.** When a second rip of a book ended up "already published", both rows point at the same output folder, and deleting the duplicate used to remove the good book. Now the source goes as normal, the output stays, and you're told "output kept: also used by 1 other book(s)". - **Delete re-checks every folder right before it removes it.** If a folder (or one of its parents) was swapped for a link after Delete checked it, it stops, removes nothing more, and keeps the book in the list. - **Manual actions can't be piled up.** Retry now, Convert anyway and Use this ASIN now refuse a book that is already encoding or waiting, like Re-check already did. At most two run at once (`MANUAL_WORKERS`); a third click gets "another action is already running". - **ffmpeg only opens local files.** Every ffmpeg and ffprobe run is limited to files and pipes, so a booby-trapped download (a playlist renamed `.mp3`, say) can't make it fetch anything from the network. - **Odd file names are held, not merged.** A book whose file or folder name contains a line break or NUL byte is now **Held** with "unsafe file name: ..." and the advice to rename it. Such a name could have slipped extra instructions into the merge list. Convert anyway doesn't override this one. - **Tags can't break out of their line.** Titles, authors and other tag values have line breaks and other control characters turned into spaces, and `#` escaped, so a value can no longer add tags or chapters of its own to the m4b. - **The container runs locked down.** The compose file, the Docker run example and the Unraid template now use a read-only root filesystem (with `/tmp` in memory), drop every Linux capability except the three needed to switch to `PUID`/`PGID`, and set `no-new-privileges`. Your library mount is unchanged; a commented `:ro` line shows the safer option and notes that it stops Delete from removing sources. - **The image build is pinned and patched weekly.** Every GitHub Action is pinned to an exact commit, CI runs with read-only permissions except where a job publishes, the base image is pinned by digest, and `:latest` is rebuilt every week from scratch so ffmpeg and Debian security fixes arrive without waiting for a release. - **Hardening headers on every response**: the UI can't be framed by another site (no clickjacking), and browsers won't guess content types. ### Fixed - **A malformed or oversized request gets a clean error.** Bodies over 64 KB (or with a negative length) get 413, and a JSON body that isn't an object gets 400, instead of a traceback in the log and a dropped connection. A client that stalls is cut off after 15 seconds. - **Opening a page never calls Audible.** The Series page, the library page and `/api/series` read what's stored. Stale series lists are fetched by the background daemon after each scan, or when the Series page asks with a POST as it opens. - **Very long titles no longer fail after the encode.** Each folder and file name is capped at 200 bytes, cut on a whole character, so long non-English titles fit the file system's limit. - **Error messages don't show internal paths.** A failed delete tells you the kind of error (`EACCES`, `EIO`, ...) and points to the log, which keeps the full detail. - A library folder literally named `ingest:...` is listed as `./ingest:...`, so it can't be mistaken for an internal key. - `docs/retire-sources.sh` keeps its scratch files in a private temporary folder, stops on errors, always resumes the daemon when it exits, and accepts `AUDIOBIND_TOKEN` when sign-in is on. ## [0.15.0] - 2026-11-10 ### Fixed - The library table no longer runs past the right edge of the page: the Match column now takes whatever room is left (and shortens with "..." earlier, full text on hover), cell padding is tighter, and the actions column has a small fixed width, so the kebab menus are never cut off, down to a 1280px-wide window. - On the Series page, clicking a series name now jumps to its missing list, the same as "Show the N missing". The name is a real link: keyboard focusable, underlined on hover. ### Added - **Series - see what you're missing.** A new **Series** tab sits under Library in the side rail. Its amber number is how many books are missing from your series. The page shows a card for every series where you have two or more books, for example "Jack Reacher - Lee Child - 18 of 31 owned - 13 MISSING". Each card has a strip of your covers in series order, and dashed amber tiles for the numbers you don't have. Below the cards, each series lists what's missing: "Missing from Jack Reacher - 13 books", with the cover, title, length and ASIN of each one. Every missing book has two buttons. **I have this** is for a book you own without a rip: it stops counting as missing, and you can undo it. **Copy Audible link** copies the book's Audible page so you can go and get it. - Only numbered books count. Unnumbered novellas, "between the numbers" entries like #23.6, and box sets are left out of the missing count. They're tucked behind a "Show the not-numbered entries" toggle. - The full series list comes from Audible, at most once a day per series. It's fetched in the background, so opening a page never waits on Audible. If Audible can't be reached, or doesn't know the series, the card shows just how many books you have and "fetching series size…". It never guesses a total. - **In every book's details**, a book in one of those series now gets a "See what you're missing" block. It says where you are ("Jack Reacher - you're 18 of 31") and what's next ("13 books missing from this series, starting with #19 Personal"), with a mini cover strip. The button takes you to that series' card on the Series page and briefly highlights it. - The series a book belongs to now comes from your metadata overrides first. After that it comes from the catalog match, then from where the book was filed in your output library (`Author/Series/14 - Title`), and finally from the source folder name. Books recorded before this version start carrying the catalog series on their next re-check. Until then, the output path and folder name fill in. - `/api/series` returns the same information as JSON. - Covers for books you don't have yet load through the app's own cover proxy, like every other cover. `/cover?asin=` now also fetches an ASIN that appears in a series list Audible returned. Nothing else changes: https only, Amazon/Audible image hosts only, and a failed cover still isn't retried for six hours. Any other ASIN still gets nothing. - **"Why this match?" - see the evidence, not just the score.** A book's detail panel now shows why the app picked its Audible match: the match card with its score, then four bars - how closely the title and author agree with your folder, how far your files' runtime is from the catalog's (green-blue inside your runtime tolerance, amber outside it, with both lengths and the tolerance spelled out), and the catalog language (amber with the -15% penalty when it isn't English). Below that, every runner-up the search scored is listed with its score and a plain reason it lost - "a different book by the same author", "Right title, wrong language", "your files are 65% shorter, over the 25% tolerance" - and its own Use this edition button. Nothing about how books are matched or graded has changed; the panel only shows numbers the matcher already worked out. Books recorded before this version keep the old panel until their next re-check (Retry now, or the next scan that looks at them). - **Metadata you edit now locks (0.15.0-dev3).** Saving a title, author, narrator, series or series index in a book's Metadata box locks that field. So does setting an ASIN with Use this ASIN, Edit ASIN or Use this edition. A locked field shows a green "🔒 Locked by you" chip. Rescans, Re-check all, the new per-book re-check and re-rips never overwrite it. One click on the chip unlocks it, and the field follows Audible's catalog again. Unlocking the ASIN clears the ASIN you set, so the next re-check matches the book on its own. Emptying a box and saving also unlocks that field. Locks are kept in the database, so they survive restarts, rescans and upgrades. Any correction you typed before this version starts out locked. - **Every row has a visible ⋮ menu.** It has Open details, Edit ASIN, **Re-check against Audible** and Delete…. Re-check against Audible is new: it is Re-check all for a single book. It asks Audible about that book again and refreshes the match, the other editions and the confidence. It keeps an ASIN you set, and auto-convert rules still apply. The menu opens with a click or the keyboard (Enter, Space or ↓, then the arrow keys). It closes on Escape or a click anywhere else. Delete… still asks for confirmation first. Edit ASIN works as before. It's also available to scripts as `POST /api/unit` with `{"action": "recheck"}`. ### Changed - **The library table says what it means (queue hygiene, 0.15.0-dev3).** Each row's status is now one labeled dot, never color alone: Converted (green), Held (blue), Review (amber), Failed (red). The word sits beside the dot, followed by how long ago: "Converted 2d ago", "Held 5h ago", "Review waiting 3h". Only Review pulses; finished states hold still. Ages read like "5h ago" for the first 7 days, then show the date (for example "Oct 28"). - The Tier column is now **Match**. It reads "Jack Reacher #14 · High Confidence", or "Current ASIN · Low Confidence" for a book whose ASIN you set. Confidence is always spelled out (High / Medium / Low Confidence). Runtime is now **Length** and reads "13.18 h". Numbers line up in tabular figures, and no length, match or status ever wraps onto a second line. The Status and Match columns are a little wider so each one fits on a single line. The table has the same columns as before. - An unknown value (no author, no length, no match, never scanned) now shows "—" instead of a blank or "-". - Column headers are real buttons, so you can sort with the keyboard. Only the active column shows an arrow. - The hover-only Retry now / Convert anyway / Edit ASIN strip on table rows is gone; the ⋮ menu replaces it. Retry now and Convert anyway stay on the Needs attention cards and in each book's details. Hovering a row still highlights it. - Plainer wording, and "locked" now means only a field you locked. "Pinned by you" in the match evidence is now "you picked this edition". The toast now says "setting ASIN … - re-checking…". A dead ASIN now reads "The ASIN you set (…) is not in Audible's catalog". Collection advice now says "set it to one exact Audible entry". ### Fixed The book currently encoding no longer lingers in the review/held/failed lists - only the ones waiting in the queue do. ### Fixed - The library table no longer runs past the right edge of the page: the Match column now takes whatever room is left (and shortens with "..." earlier, full text on hover), cell padding is tighter, and the actions column has a small fixed width, so the kebab menus are never cut off, down to a 1280px-wide window. - On the Series page, clicking a series name now jumps to its missing list, the same as "Show the N missing". The name is a real link: keyboard focusable, underlined on hover. ### Added - **Series - see what you're missing.** A new **Series** tab sits under Library in the side rail. Its amber number is how many books are missing from your series. The page shows a card for every series where you have two or more books, for example "Jack Reacher - Lee Child - 18 of 31 owned - 13 MISSING". Each card has a strip of your covers in series order, and dashed amber tiles for the numbers you don't have. Below the cards, each series lists what's missing: "Missing from Jack Reacher - 13 books", with the cover, title, length and ASIN of each one. Every missing book has two buttons. **I have this** is for a book you own without a rip: it stops counting as missing, and you can undo it. **Copy Audible link** copies the book's Audible page so you can go and get it. - Only numbered books count. Unnumbered novellas, "between the numbers" entries like #23.6, and box sets are left out of the missing count. They're tucked behind a "Show the not-numbered entries" toggle. - The full series list comes from Audible, at most once a day per series. It's fetched in the background, so opening a page never waits on Audible. If Audible can't be reached, or doesn't know the series, the card shows just how many books you have and "fetching series size…". It never guesses a total. - **In every book's details**, a book in one of those series now gets a "See what you're missing" block. It says where you are ("Jack Reacher - you're 18 of 31") and what's next ("13 books missing from this series, starting with #19 Personal"), with a mini cover strip. The button takes you to that series' card on the Series page and briefly highlights it. - The series a book belongs to now comes from your metadata overrides first. After that it comes from the catalog match, then from where the book was filed in your output library (`Author/Series/14 - Title`), and finally from the source folder name. Books recorded before this version start carrying the catalog series on their next re-check. Until then, the output path and folder name fill in. - `/api/series` returns the same information as JSON. - Covers for books you don't have yet load through the app's own cover proxy, like every other cover. `/cover?asin=` now also fetches an ASIN that appears in a series list Audible returned. Nothing else changes: https only, Amazon/Audible image hosts only, and a failed cover still isn't retried for six hours. Any other ASIN still gets nothing. - **"Why this match?" - see the evidence, not just the score.** A book's detail panel now shows why the app picked its Audible match: the match card with its score, then four bars - how closely the title and author agree with your folder, how far your files' runtime is from the catalog's (green-blue inside your runtime tolerance, amber outside it, with both lengths and the tolerance spelled out), and the catalog language (amber with the -15% penalty when it isn't English). Below that, every runner-up the search scored is listed with its score and a plain reason it lost - "a different book by the same author", "Right title, wrong language", "your files are 65% shorter, over the 25% tolerance" - and its own Use this edition button. Nothing about how books are matched or graded has changed; the panel only shows numbers the matcher already worked out. Books recorded before this version keep the old panel until their next re-check (Retry now, or the next scan that looks at them). - **Metadata you edit now locks (0.15.0-dev3).** Saving a title, author, narrator, series or series index in a book's Metadata box locks that field. So does setting an ASIN with Use this ASIN, Edit ASIN or Use this edition. A locked field shows a green "🔒 Locked by you" chip. Rescans, Re-check all, the new per-book re-check and re-rips never overwrite it. One click on the chip unlocks it, and the field follows Audible's catalog again. Unlocking the ASIN clears the ASIN you set, so the next re-check matches the book on its own. Emptying a box and saving also unlocks that field. Locks are kept in the database, so they survive restarts, rescans and upgrades. Any correction you typed before this version starts out locked. - **Every row has a visible ⋮ menu.** It has Open details, Edit ASIN, **Re-check against Audible** and Delete…. Re-check against Audible is new: it is Re-check all for a single book. It asks Audible about that book again and refreshes the match, the other editions and the confidence. It keeps an ASIN you set, and auto-convert rules still apply. The menu opens with a click or the keyboard (Enter, Space or ↓, then the arrow keys). It closes on Escape or a click anywhere else. Delete… still asks for confirmation first. Edit ASIN works as before. It's also available to scripts as `POST /api/unit` with `{"action": "recheck"}`. ### Changed - **The library table says what it means (queue hygiene, 0.15.0-dev3).** Each row's status is now one labeled dot, never color alone: Converted (green), Held (blue), Review (amber), Failed (red). The word sits beside the dot, followed by how long ago: "Converted 2d ago", "Held 5h ago", "Review waiting 3h". Only Review pulses; finished states hold still. Ages read like "5h ago" for the first 7 days, then show the date (for example "Oct 28"). - The Tier column is now **Match**. It reads "Jack Reacher #14 · High Confidence", or "Current ASIN · Low Confidence" for a book whose ASIN you set. Confidence is always spelled out (High / Medium / Low Confidence). Runtime is now **Length** and reads "13.18 h". Numbers line up in tabular figures, and no length, match or status ever wraps onto a second line. The Status and Match columns are a little wider so each one fits on a single line. The table has the same columns as before. - An unknown value (no author, no length, no match, never scanned) now shows "—" instead of a blank or "-". - Column headers are real buttons, so you can sort with the keyboard. Only the active column shows an arrow. - The hover-only Retry now / Convert anyway / Edit ASIN strip on table rows is gone; the ⋮ menu replaces it. Retry now and Convert anyway stay on the Needs attention cards and in each book's details. Hovering a row still highlights it. - Plainer wording, and "locked" now means only a field you locked. "Pinned by you" in the match evidence is now "you picked this edition". The toast now says "setting ASIN … - re-checking…". A dead ASIN now reads "The ASIN you set (…) is not in Audible's catalog". Collection advice now says "set it to one exact Audible entry". ### Fixed The book currently encoding no longer lingers in the review/held/failed lists - only the ones waiting in the queue do. ## [0.14.4] - 2026-11-09 Every finding from the v0.14.3 code audit is fixed: 4 high, 3 medium, 11 low and the 3 style notes. ### Security - **Cross-site requests are refused.** Every POST must be `Content-Type: application/json` (415 otherwise). When the browser sends an `Origin` or `Referer`, it must name this server's own host (or `X-Forwarded-Host` behind a proxy), or the request gets a 403. Another website can no longer press Delete, Stop or Save settings for you. curl and scripts that send neither header keep working. - **The series index can't steer output outside `OUTPUT_ROOT`.** The sequence goes through `safe_name` like every other path component, and `m4b_path` checks that the resolved path stays inside the output root. - **ASINs are validated and quoted.** Use this ASIN accepts only real ASIN shapes (8-14 letters and digits). Every ASIN placed in an Audnex, Audible or Amazon URL is percent-quoted. - **Cover downloads are allow-listed.** Only https URLs on Amazon/Audible image hosts are fetched, including after redirects, so `file://` and arbitrary hosts are refused. `/cover?asin=` only fetches, and only leaves `.miss` markers, for ASINs that a book in your library is matched or pinned to. ### Fixed - **One brace no longer blanks the library page.** The detail panel ran `str.format()` over text that already held user data, so a folder named `Dune {Unabridged}`, a title override like `Book {1}` or an ffmpeg error with braces crashed the whole page. It's now built with f-strings. - **Books sharing a parent folder no longer collide, and Delete can't take their siblings.** Part sets that share a parent (`Author/Foo Part 1..2`, `Author/Bar Part 1..2`) used to both be keyed `Author`, overwriting each other's row, ASIN pin and overrides. Delete then removed the entire `Author` folder. Each now gets its own key (`Author#foo`) and guesses its title from its own name. Delete removes only what the book owns: its folder, or its part folders. It refuses if another book's audio would go with them, and never touches a folder it can't match to the book (for example a pre-0.14.4 `Author` row). - **Manual actions always leave the queue.** Retry now, Use this ASIN and Convert anyway used to stay "Waiting to convert", hidden from every list until a restart, whenever the result wasn't an encode: still in review, held, dry run, already published or an error. The queue is now cleared however the action ends. - **Chapters follow the play order.** Chapter timelines walked the files sorted by name (`Track 1, 10, 2`) while the audio was joined in natural order (`1, 2, 10`), so titles and boundaries were wrong. Both now use the concat order. - **The settings page shows your real environment.** The defaults are read from the live config, not a hardcoded copy, so `MAX_RUNTIME_DELTA=0.40` shows 40%. Reset returns the tier in the select's casing, so the select is never left blank. `SCAN_INTERVAL_MIN` can be any value from 1 up, and an unusual env bitrate, region or tolerance is offered as a choice. Save had also been posting nothing at all, because it read the value off each setting's wrapper `<div>`. It now reads the real control and posts only the fields you changed. - Audnex chapters start at `startOffsetMs` when Audnex provides it, so an intro or a gap no longer shifts every chapter after it. - A dead ASIN pin says "so nothing was matched" (no search runs any more). - Delete errors are always valid JSON. A permission error is reported separately from a read-only mount, and a source that was only partly deleted is reported as such: the row and the output are kept instead of claiming the source was kept. - The re-check banner's flag is cleared atomically, so a re-check that lands at the same moment is no longer overwritten. - Convert anyway in dry run records a dry run, not "already published". - The output is re-checked after taking the encoder lock, so a manual action and the daemon can't both encode the same book. An already-published "ready" book keeps its output path and isn't re-matched on every scan. - `?q=` searches and `/cover?path=` are decoded once, not twice, so paths and searches containing `%` work. - Edit ASIN on a shelf card works when the book isn't in the table (another status filter, past the row limit, or hidden by the search box): the page reloads on a view that has the book, with its panel open. - "ffmpeg failed:" always carries ffmpeg's error text, because the stderr reader is joined before its text is read. ### Changed - `unitAction` and `rowAction` are merged into one `abUnitAction(btn, action)`. A refused action now shows its error instead of reloading silently. - Each conversion builds its chapter plan once, from the durations the audit already probed, and hands it to `encode()`. Source files are no longer ffprobed again for a second plan. - `UnitState.overrides` is gone. Operator overrides are owned by `set_overrides()`/`get_overrides()`, and `save()` never writes that column. ## [0.14.3] - 2026-11-08 ### Changed - **The Audible match block shows the author.** "Author: Michael J. Sullivan" sits right under the matched title, next to the narrator line that 0.13 added. ## [0.14.2] - 2026-11-08 ### Changed - **Delete means permanently deleted.** The owner's call: the library mounts read-write again (README, docker-compose, CA template), so Delete removes the source folder *and* the published output, exactly as the dialog says. The read-only graceful mode stays: mount the library :ro and Delete adapts - it removes the published output, keeps the source folder, and says so in the dialog and toast. Both behaviors are covered by tests. - The code audit found and fixed the bugs in 0.14.0/0.14.1 first; the only writes to the source library remain the explicit Delete path. ## [0.14.1] - 2026-11-08 ### Fixed - **Delete works on a read-only library — and tells the truth.** The hard delete was built when the library was mounted read-write; on the recommended read-only mount it errored with "source removal failed: Read-only file system". Now a delete always removes the published output and forgets the book; when the library is read-only it says so plainly in the dialog and the toast: the source folder stays on disk and the book reappears on the next scan unless the folder is removed from the share. Other source-removal failures (permissions, disk errors) still fail loudly. Regression test covers the read-only path. ## [0.14.0] - 2026-11-08 ### Added - **The conversion queue is visible.** Every Convert anyway / Retry now / Use this ASIN click joins a real queue. Queued books leave the Review/Held/ Failed lists and the sidebar counts (they were sitting there doing nothing), and a "Waiting to convert" strip shows what's next plus how many are in line; the Now-encoding strip adds "N waiting in queue". If the container restarts, the queue empties and the books reappear in their lists. - **A full code audit landed** (four parallel deepseek-v4-pro reviews over all 5,186 lines plus the test suite). The bugs below came out of it; the audit also confirmed the test suite is genuine and mapped its gaps. ### Fixed - **Books without a cover no longer produce a broken ffmpeg command.** The tags/chapters input index was hardcoded to 2, but without a cover the metadata file is input 1 - every no-cover book lost its tags and failed the pre-publish verification. The index is now computed, and the muxer is told to ignore any chapters riding on the concat input so they cannot double the chapter list. - **The Auto-convert tier setting actually works now.** The settings page saves HIGH/MEDIUM/OFF in uppercase and the tier lookup used lowercase keys, so every override fell back to HIGH - setting Medium or Off in the UI did nothing (and Off silently behaved as High). Regression tests cover it. - **A dead ASIN pin can no longer convert a different edition.** When a pinned ASIN has no catalog data (or the catalog hiccups), the book now goes to review instead of falling through to a fresh search that might auto-convert something else. - **Non-English matches are capped at MEDIUM.** The language penalty applied only to the ranking score; grading read the raw similarity, so a foreign-language edition with an exact title/author match graded HIGH and auto-converted. - **The MEDIUM grading path now needs a margin.** A match tied with a different work (margin 0) no longer converts when the tier is Medium. - **The review panel's "why" text names the real tier** instead of always saying "Auto-convert is set to High", including an explicit message when auto-convert is off. - **A re-check banner shows once.** The re-open flag is consumed the first time the page displays it, so the panel stops popping open on every refresh afterwards. - **A non-numeric MAX_RUNTIME_DELTA env value no longer crashes startup.** ## [0.13.4] - 2026-11-08 ### Fixed - **A re-check of an already-converted book can no longer downgrade it.** `convert_unit` returns "nothing written" both for a dry run and for "the output already exists", and the recorder treated both as `ready` - wiping the row's output path and chapters. Found live: Re-check all hit one converted book whose m4b already existed and flipped it to "ready" with a wrong dry-run note. Now the recorder distinguishes the two cases: a dry run is still `ready`, and an already-published book with an existing converted row is left untouched (its row keeps the output path, chapter count and detail). A book with no previous row gets an explicit "already published - the output file exists" note instead of a phantom dry-run message. ## [0.13.3] - 2026-11-08 ### Added - **"Scan finished" toast.** A pass can end faster than the page's first status poll, so even the spinner can be missed. When a page loads within ~20 seconds of a scan pass ending, it now toasts the outcome: "Scan finished - 98 checked, 0 converted, 0 held, 0 failed". Quiet passes are now provably quiet instead of invisible. ## [0.13.2] - 2026-11-08 ### Added - **Re-check all.** A new button next to Scan now re-audits every book against Audible, including ones that already have a verdict. This is what fills in the "Other editions seen" panel for a whole existing library in one pass (and refreshes chapter data along the way). Books still follow the auto-convert rules - only a HIGH-confidence match converts. The scan progress indicator covers the pass; expect it to take a while (98 books = 98 catalog lookups). ### Fixed - **The scan indicator explains quiet passes.** A plain Scan now skips books that already have a verdict and can finish in under two seconds - faster than the first status poll, so the indicator was invisible. The controls row now explains what Scan now covers in its tooltip, and Re-check all is the button for a full pass. ## [0.13.1] - 2026-11-08 ### Added - **Scan progress is visible.** Pressing Scan now shows a spinning "Scanning n/m…" indicator in the controls row, and the Scan button disables and relabels itself for the duration. The page stops auto-refreshing while a scan pass runs (it used to reload on every change, fighting the scan), then refreshes once when the pass ends so the fresh results appear. A page loaded mid-scan shows the indicator immediately. ### Changed - **Removed the "auto: high" tag** from next to the version badge. The tier belongs in Settings, where it is explained; the header just showed a code fragment nobody asked for. ## [0.13.0] - 2026-11-08 Four ideas from the Chaptarr / Listenarr comparison landed. The fifth (metadata consensus) is deferred - see the backlog for why. ### Added - **Chapters survive the merge when the sources have them.** The m4b builder now reads the chapter tags inside the source files themselves and uses them for the merged book's chapter list, keeping the publisher's own titles ("Chapter One" beats "01 - Part_01.mp3"). Across a multi-disc rip the markers stack onto one continuous timeline so disc two's chapter one starts where disc one ended. The rules, in order: source tags win for single-part m4b/mp3 sources; for multi-part merges or odd containers the Audible chapter data is preferred (per-file tags can repeat or drift at part boundaries); with no tags anywhere and no usable Audible data, it falls back to one chapter per file as before. The log line now says which won - `chapters: 31 from source-embedded (31 chapters)`, audnex, or per-file. - **Other editions are one click away.** While matching, the top runners-up from the Audible search (usually other editions of the same book - unabridged vs abridged, different narrators) are now saved with the book. A new "Other editions seen" section in the book's detail panel lists them with a "Use this edition" button that locks the match to that ASIN and re-checks the book, reusing the existing pin flow. - **Named quality profiles on the settings page.** A select at the top presets the two conversion settings: Balanced (HIGH tier, 128k floor), High (HIGH tier, 256k floor) and Source-match (HIGH tier, 64k floor, which keeps the source bitrate as-is). Editing either setting afterwards flips the profile to Custom so it is visible when the settings stopped being a preset. All six existing settings behave exactly as before. - **Narrator now surfaces in the UI.** The detail panel's "The Audible match" block shows "Narrated by …" when the catalog knows the narrator, and the metadata editor's narrator box pre-fills its suggestion from the match (the narrator tag in the m4b and the editor field itself existed already - this closes the loop from match to editor). ### Fixed - **The settings page's Save button works again.** The page's JavaScript was being served with doubled braces from a templating artifact, which made the whole script a syntax error - Save, Reset and toast feedback silently did nothing (the underlying POST handler was fine; the buttons just couldn't call it). The script is now plain JavaScript and parses. ## [0.12.5] - 2026-11-08 ### Changed - **The ASIN row speaks plain English.** "Pin an exact ASIN:" is now "Lock this book to one exact Audible entry:", the hint explains what locking actually does (every future scan checks this book against this exact entry instead of guessing), and the Edit ASIN / Use this ASIN tooltips say "lock" instead of "pin". - **The README now tells the truth about source safety.** The beta disclaimer explains the real mechanics: the source library is only ever read, converted books are written to the separate output library, and the hard delete is the one destructive path. (A stale note claimed converted sources were moved to a trash folder - that behavior never shipped, and the README no longer claims it.) ## [0.12.4] - 2026-11-08 ### Changed - **The detail panel layout matches the approved mockup.** The cover now sits beside the explanation instead of below the buttons; "Why it's waiting for you" and "Your options" stack in one column next to it; Convert anyway is the filled amber primary and Retry now the ghost button; the ASIN pin row carries a "Pin an exact ASIN:" label, the current ASIN pre-filled, and the hint line underneath. Delete moves to its own row at the very bottom, away from the action buttons. - All detail section headers (including Your source files / What you'll get / The Audible match) render in the amber accent, and the big cover grew to 190px, as in the mockup. ## [0.12.3] - 2026-11-08 ### Changed - **The review panel speaks plain English.** "Why it is review / What to do / Before / After / Decision" are now "Why it's waiting for you / Your options / Your source files / What you'll get / The Audible match", with status-specific headers everywhere ("Why it failed", "Why it's paused", ...). The field copy reads like sentences: "32 audio files in 1 part", "31.79 hours of audio", "Audio format not detected in the files". The match block names the book and links the ASIN to Audible, and a LOW/MEDIUM/HIGH confidence meter sits under the explanation. - **Every advice string was rewritten** in the approved plain-English voice (converted, ready/dry-run, interrupted, failed, held, unreadable sources, no-match, collection, low confidence, medium-confidence fallback, dead pinned ASIN, never-scanned). - **Book action buttons now live inside the detail panel** - Convert anyway / Retry now / the ASIN pin row sit right under "Your options" instead of only on hover at the end of the table. ### Added - **Manual re-checks that still need a decision re-open their panel.** Retry now and Use this ASIN run a manual re-check; if it still ends below your auto-convert level (or held/failed), the detail is flagged, the page reloads and the panel re-opens by itself with an amber banner saying what the re-check found and what to do next. If the re-check passes, the book simply starts converting and the panel stays closed. ### Fixed - **A same-status re-check used to go unnoticed.** The page only reloaded when book counts changed, so a Retry now that ended in the same status refreshed nothing while a panel was open. The page now watches a change signature that moves on every row write. - **Detail panels no longer persist across refreshes.** Every panel starts closed on page load; the only panels that open by themselves are the ones a manual re-check asked to re-open. ## [0.12.2] - 2026-11-08 ### Changed - **Both "Audiobind" names are links home.** The sidebar brand and the page title now take you back to the main library view from anywhere (Settings included) - they used to be dead text. - The lone "auto: high" indicator that floated at the far right of the header now sits with the version badge instead of hanging by itself. ## [0.12.1] - 2026-11-08 ### Fixed - **Column sorting works again - on every column.** A shadowed variable in the row renderer replaced each book's five sort attributes with a literal `{}`, so clicking any column header (not just Status and Tier) silently sorted nothing. The reason-line block was overwriting the sort-attribute string with the parsed detail blob; it now uses its own variable. Regression test asserts every rendered row carries all five sort keys, with and without a reason line. ### Added - **What the tiers mean, in Settings.** The Auto-convert tier dropdown now explains HIGH (strong title+author match with a clear margin - safe to auto-convert), MEDIUM (good match, weaker evidence), LOW (held for review - fix with Edit ASIN), and COLLECTION/NONE (never auto-converted). ## [0.12.0] - 2026-11-08 ### Added - **Settings page** at `/settings` with six controls: auto-convert tier, output quality floor, wrong-edition tolerance, scan interval, CPU priority (nice), and catalog region. Values are persisted in the database and override environment defaults. The daemon wakes early when the scan interval is lowered and every consumer reads the effective value. - **Settings link** in the left rail (gear icon, below Library). - **`/api/settings` endpoint** with server-side validation; `/api/state` now includes an effective `settings` object and uses the effective `auto_convert` for the top-bar indicator. - **README installation section** with Unraid (manual template), `docker run`, and Docker Compose paths, plus a first-run paragraph. ### Changed - **Row counter removed** from the toolbar next to the search box — the static "N books" / "N of M shown" label, the `.ab-shown` CSS, and the JS that updated it are gone. The empty-state message is unchanged. ## [0.11.8] - 2026-11-08 ### Fixed - **The expanded detail panel no longer forces the page sideways.** The detail row inherited the books table's no-wrap cell rule, so the Why/What text, the Before/After blocks and the metadata hint could not wrap and pushed the table wider than the screen. The detail row now wraps text normally; the compact book rows keep their no-wrap cells. ## [0.11.7] - 2026-11-08 ### Fixed - **The hover action buttons no longer clip at the screen edge.** The three row buttons (Retry now / Convert anyway / Edit ASIN) were wider than the actions cell at typical viewports, and the cell's hidden-overflow clipped them mid-word. They now wrap to a second line inside the cell when space is tight, with tighter padding so all three fit on one line at 1080p. ## [0.11.6] - 2026-11-08 ### Fixed - **The table can no longer grow past its columns.** The books table now uses `table-layout: fixed`, so long titles can never push the STATUS/TIER/RUNTIME columns (and the hover actions) off-screen again - the max-width cap alone did not hold under auto layout. The BOOK column is capped at `min(560px, 30vw)` with the title truncating on ellipsis, and the freed width flows into the actions column where the hover buttons sit, right of Last Scan like the mockup. The hover overlay is clipped inside its cell, so no width produces a horizontal scrollbar. ## [0.11.5] - 2026-11-08 ### Fixed - Rail footer now wraps long host paths inside the sidebar (overflow-wrap:anywhere, 10.5px). - BOOK column constrained to 38%/520px; title and author truncate with ellipsis. - Hover action buttons anchored inside a fixed-width last column (170px), no longer cause horizontal scroll. ### Changed - Delete confirmation is now a shared modal overlay instead of an inline strip. - Row-hover "Use this ASIN" renamed to "Edit ASIN"; clicking it opens the detail panel and focuses the ASIN input (shelf cards too). Panel button unchanged. ## [0.11.4] - 2026-11-08 ### Fixed - The rail footer still said the library was "(read-only)" after 0.11.3 made it read-write for the delete feature. The stale note is gone. ## [0.11.3] - 2026-11-08 ### Added - **A plain-English reason line under the title for problem rows.** Held, review and failed units now show the "why" text from their advice explanation directly in the books table (amber for held/review, red for failed), so you can see what is wrong without expanding the detail panel. - **Delete a book — permanently.** A Delete button in the detail panel's action row opens a warning dialog ("This permanently deletes the book's source folder and its published files. This cannot be undone."); deleting requires clicking the red "Delete permanently" button. The deletion removes the book's source folder under the library and its published output folder under the output library, then drops the book from the state database. It refuses while the book is encoding and on any path that would escape the library or output roots. (The library mount must be read-write for the source deletion.) - **Hover tooltips on every control.** Scan now, Pause, Stop, Resume, Dry run and the per-book actions each explain what they do before you click. - **Host paths in the rail footer.** Optional `HOST_LIBRARY` / `HOST_OUTPUT` environment variables are shown when set, so the footer names the real Unraid folders instead of the container's `/library` and `/output`. The stale "(read-only)" note was removed - the library is mounted read-write since the delete feature exists. ### Changed - **The books table is more compact.** The permanent action-buttons column is gone — Retry now / Convert anyway / Use this ASIN now appear on row hover as a compact overlay with a solid background, matching the approved mockup. The column no longer reserves ~350px of width, so the table fits at 1080p without horizontal scrolling. - **The Dry run button sits with the other controls** (Scan now / Pause / Stop) instead of alone at the right edge of the control row. ### Fixed - No fixes in this release. ### Added - **The rail footer shows the real host paths.** The container only ever knows `/library` and `/output` - meaningless names on a NAS. Two optional environment variables, `HOST_LIBRARY` and `HOST_OUTPUT`, are now displayed in the rail footer when set (the Unraid template ships them with the actual share paths); without them the container paths are shown as before. ## [0.11.1] - 2026-11-08 ### Added - **Every manual action now answers.** Retry now / Convert anyway / Use this ASIN show a toast ("pinning B013ILGDHQ - re-checking…") immediately, and the button itself flips to "working…" — previously a click was followed by silence until the next render. - **A pinned ASIN that Audible will not serve is explained, not swallowed.** Some ASINs are ghost catalog entries: the API returns `REGION_UNAVAILABLE` (checked across five regions) and no metadata at all. When a pin falls back to the search result, the book's Why/What panel now says so in plain English and notes that the pin stays saved. ### Fixed - **The sticky table header stopped one column short.** The Book header cell carried the `ab-bookcell` class, which also styles the body's title cell as `display:flex` - the same class-collision family as the old `.bar` bug, and it broke the sticky bar's layout so it did not span the full width. The header cell now has no cell class and the rule is scoped to `td.ab-bookcell`. - The rail's Review glyph was the invalid entity `&flag;` (rendered as literal "&flag;" text) and Library used the nonexistent `└`; both are real unicode glyphs now. - The rail footer's paths are labelled (`library /library (read-only)`, `output → /output`) instead of two bare paths. ## [0.11.0] - 2026-11-08 ### Added - **Left sidebar rail** with live unit counts for Library, Review, Held, and Failed. Clicking an attention item sets the corresponding status filter; "Library" clears all filters. - **Needs-attention shelf**: collapsible card strip showing held/review/failed units (failed first, capped at 12 with "+N more"), each with cover, explanation, and quick-action buttons (Retry now, Convert anyway, Use this ASIN). Collapsed state persists in localStorage. - **Pinned now-encoding strip** showing live title, percent, speed, and ETA with Pause/Stop. Hidden entirely when no encode is running — no empty stub. - **Hover row actions** on problem rows (held/review/failed) in the table: Retry now, Convert anyway, and Use this ASIN appear on row hover without needing to open the detail panel. ### Changed - **Full page redesign** matching the approved dark-theme mockup: left sidebar + main content layout, CSS custom properties, Inter/JetBrains Mono approximated with offline-safe system font stacks (no Google Fonts — the container works without internet). - Stats display replaced with a slim monospaced stats line (hours, GB, progress with mini bar). - Filter chips restyled as rounded pills matching the mockup's toolbar. - Status pills and tier badges redesigned with coloured dots and soft backgrounds. - Detail panel restyled with darker background and tighter spacing; metadata editor unchanged. - Search box and controls row restyled to match the new design vocabulary. ### Fixed - All existing API endpoints, polling behaviour, sessionStorage persistence, and dry-run toggle preserved unchanged. Page polls every 2s with the same busy-guard rules (open panels, focused search, interactive elements prevent reload). ## [0.10.1] - 2026-11-07 ### Fixed - **Most book covers showed a placeholder.** The ASIN-derived Amazon URL pattern resolves for only some ASINs; the rest return a 43-byte 1x1 tracking pixel with HTTP 200, which the download correctly rejected but then remembered as a six-hour miss. `/cover` now asks the Audible catalog for the real `product_images` URL for the ASIN first (the lookup is memoised with the same 15-minute cache as every other catalog call, so repeats are free), and falls back to the pattern only when the catalog has nothing. Books whose match data already carries an image URL still use it without any lookup. Miss markers from the pattern-URL era were cleared so every cover gets a fresh attempt on first view. ## [0.10.0] - 2026-11-07 ### Added - **Search and column sort in the books table.** A live search box filters on title, author and path (case-insensitive; every whitespace-separated term must match), and the column headers sort on title, status, confidence tier, runtime and last scan, with the direction toggling on re-click and an arrow marking the active column. Search composes with the existing status chips. The rules are not JS-only: `web.search_matches()` and `web.sort_value()`/ `web.sort_units()` run the same logic server-side, and the chosen column and direction are stored in the state database, so a restart restores them. - **Cover art in the table (48px) and the detail panel (150px).** Covers come from the `cover.jpg` already published next to a finished m4b, or from the catalog image URL now persisted with the match data (`Product.image`); books matched before 0.10.0 fall back to the Amazon URL derived from their ASIN. Everything is served by `/cover`, which fetches an image at most once into `STATE_DIR/cover-cache` and serves it from disk forever after, so the page still renders with the network down. A failed fetch leaves a six-hour `.miss` marker so a dead cover is not re-requested on every poll, and a miss falls back to a neutral placeholder. Only an ASIN-shaped string can become a cache filename, and a published cover must resolve inside `OUTPUT_ROOT`. - **A metadata editor in the detail panel.** Title, author(s), narrator(s), series and series index can be corrected per book and are stored as overrides in the units table, beside the ASIN override they follow; `save()` leaves the column alone, so a correction survives later rescans. The overrides are written into the catalog payload before anything reads it, so the m4b tags, `metadata.json` and the author/title filing layout all agree. Corrected fields are marked *custom* in the panel and the book carries a badge in the table. An empty box means "no override" rather than "blank this field", so clearing a box hands the field back to the catalog; the catalog value is shown as the box's placeholder. Saving is one request with inline feedback - no page reload, and no library scan. - **A catalog lookahead during the encode.** While a book encodes, one background thread does the next eligible book's catalog search and chapter fetch, so the main loop does not wait on the network when it gets there. Successful catalog responses are memoised by URL (15 minute TTL, bounded at 256 entries; failures are never cached), which is where the pipeline already looks. Exactly one lookahead runs at a time, the thread only reads `State` through the shared `RLock` and writes nothing, and any failure is logged and forgotten - the normal path simply fetches again. - **Stats cards on the dashboard**: total hours of audio converted, total GB written to the output tree, and a library progress bar (converted vs total discovered units) in the existing card style. The numbers are computed from the state database on each render - output sizes are recorded at publish time, so nothing scans the output tree - and the same block is served on `/api/state` as `stats`. ### Changed - The page no longer reloads while a detail panel is open or while the search box is focused or non-empty, so a refresh cannot interrupt reading or typing; open rows persist in `sessionStorage`. Table columns have fixed widths so no header truncates. - The stylesheet and the script moved out of the `PAGE` format template into their own constants. CSS and JS braces no longer have to be doubled, which is what produced the template bugs in the 0.9.x line. Every class added in 0.10.0 is `ab-` prefixed and a test enforces it. - The test suite covers the new work: the search filter, the sort comparator, override application in tag and `metadata.json` writing, cover caching with the network down, prefetch thread safety under concurrent `State` access, and the stats computation. `docs/verify-page.py` additionally boots the real server and checks the served HTML and inline JS. ### Fixed - **The test suite collected nothing outside the container.** `tests/` was not a package, so `python3 -m unittest discover` silently ran zero tests on Python 3.12+, where namespace-package discovery was removed. With `tests/__init__.py` in place the documented command collects the whole suite - 185 tests at this release. - **Total GB written was understated by about 2.4%.** The stats block divided megabytes by 1024, but `out_mb` is recorded as bytes/1e6, so a GB here is 1000 MB; 2560 MB read as 2.50 GB instead of 2.56 GB. - The library progress bar reused the live encode bar's `.ptrack` class. It now has its own `.ab-statbar`/`.ab-statfill`, so restyling the encode bar cannot deform a stat card - the same class collision that once crushed the toolbar. - A second catalog lookahead could start while the first was still being spawned: the guard read `Thread.is_alive()`, which is still false between construction and `start()`. The busy flag is now set inside the lock before the thread starts. - `State.last_run()` read the database without taking the shared lock. Harmless while only the daemon and the web handlers existed; a background prefetch thread makes it a real race. - The two tests that serve real HTTP endpoints bound fixed ports (8096/8098) and could fail with `OSError: [Errno 98] Address already in use` when a listener from a previous run lingered. They now ask the kernel for a free ephemeral port. ## [0.9.3] - 2026-11-01 ### Added - The status page now declares a favicon and theme colour, so the browser tab shows the Audiobind icon instead of a blank page glyph. The icon is served from the container at `/favicon.svg`, and the image carries `icon.svg` so it works without internet access. ## [0.9.2] - 2026-11-01 ### Fixed - **The daemon could wedge and stop converting anything.** The daemon thread, the web handlers and the manual Retry/Convert actions all shared one SQLite connection with no lock and implicit transactions, which surfaced as `sqlite3.OperationalError: cannot start a transaction within a transaction` and left 21 books sitting in `ready` forever. Writes are now serialised behind an `RLock`, the connection runs in autocommit with a 30 s busy timeout, and there is a test that hammers it from four threads. - Only one book is encoded at a time: a manual action now waits for the daemon's current book instead of running a second ffmpeg alongside it. ## [0.9.1] - 2026-11-01 ### Fixed - Books recorded before 0.9.0 keep their runtime comparison in the note rather than the detail blob, so their "why it is held" explanation only showed half the story. Both numbers are now recovered from the note. ## [0.9.0] - 2026-11-01 ### Added - **"Why it is held / in review" and "What to do"** at the top of every book panel, in plain English. A held book now shows both runtimes and the percentage difference and explains that it is usually a different edition or missing parts; a review book names the actual reason (no catalog match, weak match, unreadable source files, looks like several books) and the specific next step - Retry, paste an ASIN, Convert anyway, or split the folder. - The catalog comparison (matched title, ASIN, runtime) is stored with each book, so the explanation survives restarts and does not depend on re-querying Audible. ## [0.8.1] - 2026-11-01 ### Fixed - An expanded book panel closed itself after a few seconds: the page refreshed on a fixed 10 second timer. Open panels are now remembered in `sessionStorage` and restored after any refresh, and the page only reloads when the status counts actually change. ## [0.8.0] - 2026-11-01 ### Added - **Per-book actions**, so `review` is no longer a dead end. Expand a row and you get: - **Retry now** - re-audit and convert under the normal rules (useful after you fix a source file or change `AUTO_CONVERT`). - **Convert anyway** - the operator override: ignore the confidence tier, the runtime hold and the damaged-source hold. The output is still verified before it is published. - **Use this ASIN** - pin the book to a specific Audible entry when the automatic match picked the wrong edition; the override is stored and used on every later scan. - `POST /api/unit {"path": ..., "action": "retry" | "force" | "set_asin", "asin": ...}`. ### Fixed - Rows recorded before 0.7.0 showed no size in the detail panel even though the columns held it. ## [0.7.0] - 2026-11-01 ### Added - **Filter chips by status** (`all`, `converted`, `ready`, `review`, `held`, `failed`, `interrupted`) with live counts, so the books waiting for review are reachable instead of being buried below the fold. - `/api/units?status=ready` returns the filtered list as JSON for scripting. - The detail panel now shows the full source path, not just its size and codec. ### Fixed - The table silently showed only the newest 40 rows even when there were far more books. It now lists every matching unit (capped at 400 with an explicit note), sorted by title when filtered. ## [0.6.1] - 2026-11-01 ### Fixed - The row expander used a Unicode triangle that renders as a tofu box in the default UI font; it is now a plain `+` / `-`. - Rows recorded before 0.6.0 say "codec not recorded" instead of inventing `0 MB / unknown codec`. ## [0.6.0] - 2026-11-01 ### Added - **Before/after detail on every row.** Click a book in the activity table to expand a panel showing what it was (parts, files, size, runtime, codec/channels), what it became (file, size, runtime, chapters and where they came from, full path) and the decision (match confidence, ASIN, notes). Stored as JSON on the unit, so it survives restarts. ### Fixed - **A book that is already an AAC `.m4b`/`.m4a` is no longer re-encoded.** It is remuxed (`-c:a copy`), so chapters, tags and cover are rewritten while the audio samples are untouched. Anything not AAC (e.g. ALAC, or mp3 that only pretends to be m4b) still converts normally. ### Changed - **Default bitrate is now 128 kbps** (was 64), and it acts as a *floor*: a source richer than the floor keeps its own bitrate (a 256 kbps mp3 stays 256 kbps), anything below is lifted to it, and the result is capped at 320 kbps. ## [0.5.3] - 2026-11-01 ### Fixed - A book with unreadable source files is now held for review **before** encoding, instead of spending minutes producing a short file that then fails verification (`1984` hit this: 168 files, one unreadable, output came out at 5.2 h against 9.7 h of sources). - When the merged output is shorter than its sources, the error now says so in hours and quotes the first ffmpeg complaint, instead of a bare pair of second counts. ## [0.5.2] - 2026-11-01 ### Fixed - `last scan` now updates at the start of a pass as well as after each book, so it reads fresh while a long scan is still running instead of showing the previous run's age. ## [0.5.1] - 2026-11-01 ### Fixed - The control bar collapsed into colored slivers: the new progress track used the class `bar`, which is already the class of the toolbar container, so `height:8px; overflow:hidden` clipped every button. The track is now `.ptrack` and the toolbar is `.toolbar`. - `last scan` no longer shows a stale time during a long pass - the heartbeat is updated after every book, not only when the whole scan ends. ## [0.5.0] - 2026-11-01 ### Added - **Progress bar for the book being encoded.** ffmpeg runs with `-progress pipe:1` and the percentage, realtime speed, elapsed time and ETA are published on the status page, refreshed every 2 seconds without reloading the page. The same values are in `/api/state` under `progress` for scripting. ### Fixed - `render()` shadowed its `state` argument while building the progress text, which would have broken the activity table. ## [0.4.0] - 2026-11-01 ### Added - **Stop now** button. Unlike pause (finish this book, start nothing new), stop aborts the running ffmpeg immediately. Stopping is safe by construction: the encode writes into the work directory and only moves into the library after verification, so an aborted conversion leaves the partial file in the work directory, which is deleted. The book is marked `interrupted` and retried on the next scan from scratch (nothing is resumed, so a half-written file can never be published). - The page names the book currently converting, so pause-vs-stop is an informed choice. ### Changed - The activity table's `hours` column is now `runtime` - it is the on-disk running time of the source files, used to prove a multi-part set is complete. ## [0.3.0] - 2026-11-01 ### Added - **Dry-run toggle on the status page.** `DRY_RUN` no longer requires editing the container; the button switches between deciding-only and writing, and the choice survives a restart. The environment variable remains the default for a fresh install. - The button turns amber when dry run is on (clicking it arms real writes) and red when writing. ### Changed - `pipeline` reads the effective dry-run state from the run controls rather than the environment. ## [0.2.0] - 2026-11-01 ### Added - **Controls on the status page**: `Scan now`, `Pause after this book` and `Resume`. Pausing stops the daemon before it starts another book (the one in flight finishes), and the paused state is remembered across container restarts. - `POST /api/control` with `{"action": "scan" | "pause" | "resume"}` for scripting. - The page shows a running/paused pill and refreshes itself every 10 seconds. ## [0.1.4] - 2026-11-01 ### Added - Every release tag now publishes a GitHub Release whose body is this changelog section, generated by `scripts/release_notes.py` in the build workflow. - New application icon: a tetra-teal spined book with a waveform across the cover, verified legible at 32 px. ## [0.1.3] - 2026-11-01 ### Fixed - Dry runs no longer record books as `converted`. They get their own `ready` status, shown as a separate card on the status page, so a dry-run pass cannot be mistaken for real output. ## [0.1.2] - 2026-11-01 ### Fixed - `audiobind health` prints "no scan recorded yet" instead of the nonsensical "infs" before the first scan completes. ## [0.1.1] - 2026-11-01 ### Fixed - `audiobind version` and `audiobind health` no longer fail when no volumes are mounted. The output-path write check now runs only for the commands that actually write (`audit`, `scan`, `convert`, `daemon`), which is what a container healthcheck needs. ## [0.1.0] - 2026-11-01 First public release. ### Added - Book *unit* discovery: sibling `Part N` / `Disc N` folders (including CamelCase names such as `StoneColdUnabridgedPart2`) are grouped into one book, and a bare `Disc1` keys on its parent so a 9-disc set does not become nine one-hour books. - Catalog verification against the public Audible search API plus Audnex: title, author, narrator, series, ASIN, genres, description, cover. - Confidence grading (`HIGH` / `MEDIUM` / `LOW` / `COLLECTION` / `NONE`) with the match margin measured against the best *different* work, so duplicate editions do not count as ambiguity. - Completeness proof: summed runtime is compared with the catalog runtime; a set that is short is reported, never silently merged. - Merge and convert to a single m4b (`ffmpeg` concat, AAC 64 kbps mono by default) with Audnex chapter marks scaled onto the real timeline, embedded cover art and a full tag set. - Publishing to `Author/Title/Title.m4b` or `Author/Series/NN - Title/`, with `cover.jpg` and an Audiobookshelf-compatible `metadata.json`. - SQLite manifest so re-runs are idempotent, plus a status page and JSON API on port 8080. - Unraid template, GitHub Actions image build to GHCR, and unit tests for parsing and grading.