v0.6.4

• Stable GitHub Releases and the rolling development prerelease publish the
  unsigned Windows .exe again next to the Microsoft Store listing. The
  checksum files list it.
• On Windows, replaying a Player.log that holds several drafts of the same
  event no longer stops with "Could not append draft audit record". Draft ids
  that Draft Omen makes up for these drafts no longer contain colons, and a
  new draft whose id has a character Windows forbids in file names now fails
  with a clear error. Drafts saved earlier keep loading. (#915)
• On Windows, a draft that an earlier version saved with colons in its id
  now records audit entries and shows pick recommendations again. Its audit
  log file name uses underscores instead of colons. On macOS, an existing
  audit log with colons in its name keeps receiving new records.
• Dismissing an error that is already gone, such as with a second click on
  the dismiss button, no longer logs "Unknown session error" or shows a
  background work failure.
• The website publishes the latest released version at
  https://www.draftomen.com/version.json, so the desktop app can check for
  updates. The release output check fails when the file is missing or out of
  date. (#919)
• Draft Omen can check version.json on the website for a newer release on a
  background thread. Network errors, HTTP errors, a bad response and an
  unknown installed version log a warning and report no update. (#920)
• The desktop app checks the website for a newer release when it starts. If
  one exists, an "Update available" dialog names both versions and links to
  the website, the Microsoft Store listing and the GitHub releases page. The
  dialog appears at most once per launch, and the smoke and screenshot modes
  skip the check. (#921)
• Draft Omen now loads card data for a Quick Draft pack when the log has no
  event join or course line for that draft. Before, those packs got no
  recommendations, and draftomen-tui watch --plain stopped with "Shared live
  session did not score the offered pack". (#918)

v0.6.3

• Draft Omen no longer fails to start the live session when you have
  completed the same event more than once. It now matches a replayed pack to
  the saved draft that recorded it, or else to the most recently updated
  one. If re-reading the Arena log at startup still fails, the app shows the
  error and keeps following Arena and Moxgate drafts. (#909)

v0.6.2

• The Intel macOS app no longer hangs at startup with the Dock icon bouncing
  on macOS 26. On Intel Macs Draft Omen now verifies HTTPS with its bundled
  certificate file instead of the system trust store. (#905)
• A release no longer updates the website. The release workflow stops checking
  website/package.json and the news page. The website version and news page
  move to a separate pull request merged after the GitHub Release is
  published, so the site's macOS download links never point at a release that
  does not exist yet.

v0.6.1

• draftomen-tui replay now scores a draft with the profile for its own
  format, such as Pick-Two with a Premier fallback, instead of always the
  cached Quick Draft profile. It fetches the hosted profile when needed. The
  new --offline-profiles option keeps it on the local cache. (#900)
• Card data export now adds the basic lands that 17Lands cards.csv lists
  and Scryfall has no Arena id for. The new export-set-data --cards-file
  option reads a local copy instead of downloading it. The hosted Reality
  Fracture card data now has its 25 basic lands, so they read Basic and
  score 0 instead of showing as "Unknown card" with a neutral score. (#895)
• Pick-Two and Traditional profiles now use PremierDraft ratings and pair
  rates for cards and color pairs that have no usable data in their own
  format, as Quick Draft profiles already did. Such cards no longer get the
  neutral prior and now read Premier*. Older Draft Omen versions reject
  these profiles and keep their cached Pick-Two or Traditional profile until
  updated. (#896)
• A draft played before Arena logs an account id, as happens after a saved-token
  login, is now saved under the first account id that the same login reports.
  Recovery, the audit log, backtest and draftomen-tui replay now see that
  draft. A later login with another account does not take it over. (#894)
• When Draft Omen is not running, the Moxgate popup now shows how to start it
  from the desktop app, by choosing Moxgate in the draft source selector, as
  well as the draftomen-tui watch --source moxgate command. (#887)
• The Moxgate extension has the Draft Omen icon in the Chrome toolbar and on
  the extensions page. The popup shows the icon and name in a header, a green
  or red dot for whether Draft Omen is running, and follows the system light or
  dark appearance. (#885)
• The Moxgate extension now sends the latest snapshot again every 2 seconds
  while Draft Omen does not answer, so Draft Omen shows the current pack when
  it starts after the draft page is already open. It stops once Draft Omen accepts or rejects
  the snapshot, a newer snapshot replaces it, or the Moxgate page closes. The
  popup shows that the pick is waiting for Draft Omen. (#886)
• The app log now records the Arena log file being followed and, for each new
  draft, the set, event and format. It also records where card data came from,
  the cache or a download with the artifact checksum, and which ratings
  version is in use. A card image host that keeps failing logs one warning and
  a count of later failures. Preferences that cannot be loaded or saved log a
  warning. (#880)
• Fixed the macOS app failing to verify HTTPS certificates on Macs without
  Homebrew OpenSSL. Card data, ratings, 17Lands data and card images could not
  download there, because the bundled OpenSSL looked for its certificate file
  in a Homebrew folder. The app now uses the system certificate store and falls
  back to the certifi file it bundles. (#879)
• The app log now says why a hosted download failed. Card data, the sets
  manifest, ratings and the augmented model each log the URL and the cause,
  such as an HTTP status or a certificate error. A ratings refresh that ends
  with problems logs them in one line. The startup lines also say whether the
  app runs as a bundle and which certificate file OpenSSL uses. An unexpected
  error in a desktop app worker now writes its traceback to the log. (#878)
• Draft Omen tells you when Arena's Detailed Logs are turned off. Arena still
  writes Player.log without them, so before this change the app waited for a
  draft that never showed up. It now reads the header Arena writes at startup,
  shows a dialog with the steps to enable Detailed Logs (Plugin Support) and
  restart Arena, and keeps the same guidance in the main window. The dialog
  and the guidance clear once Arena restarts with the setting on.
• The live draft view keeps the card list beside the card details and pool
  panels when the window is short, as on a 14-inch laptop. The card list
  scrolls, the pool panel shrinks and scrolls, and the view only switches to
  the stacked layout when the window is too narrow.
• Quick Draft uses the PremierDraft ratings when a set has no QuickDraft
  profile yet, as Traditional and Pick-Two already do. Arena opens Quick Draft
  weeks after a set releases, so Quick Drafts and Moxgate drafts of a new set
  such as FRA showed neutral scores for every card. The ratings read
  Premier* until a QuickDraft profile is published. When no format has a
  profile, the ratings status says none are published yet instead of saying
  the refresh failed. (#871)
• Moxgate drafts no longer need a Scryfall bulk file. Draft Omen matches each
  card by the Scryfall id in its hosted card data, or by name for other
  printings, so sets such as FRA that Scryfall has no Arena ids for work on
  release day. (#868)
• When Draft Omen rejects a Moxgate snapshot, for example because a card is
  unknown, the desktop app shows the reason under the Moxgate status and the
  TUI shows it in the status line. The status stays on waiting until a
  snapshot is accepted, and the next accepted snapshot clears the error.
  (#869)
• The app loads hosted card data files that contain fields it does not
  know, so future card data can add fields without breaking this version.
  Required fields are still checked, and export-set-data still rewrites
  any file that is not exactly in the current format. Versions 0.6.0 and
  earlier still reject unknown fields. (#867)

v0.6.0

• scripts/patch_scryfall_arena_ids.py adds Arena ids from 17Lands card
  ratings to Scryfall default-cards data. Use its output with
  export-set-data --bulk-file for a set Scryfall has no Arena ids for.
  docs/set-profiles.md describes the steps. (#861)
• The daily profile refresh uses the committed card data for a set when
  Scryfall has no Arena ids for it, so sets such as FRA no longer fail with
  card-metadata-unavailable. The refresh summary lists the sets that used it.
• Draft Omen keeps following a human draft after Arena restarts in the
  middle of it. The parser picks the draft up from the in-progress course in
  the EventGetCoursesV2 course list, so the packs after the restart show up.
  Arena does not log the picks made before the restart. When Draft Omen saved
  them before the restart, recommendations and the pool still use them, and
  the full pool from DraftCompleteDraft fills any gap at the end. (#856)
• watch and replay no longer fail in the middle of a human draft when
  Arena answers a failed pick with the table state. The parser drops the
  failed pick and records the card from the resubmitted request. Other
  unknown pick response bodies are ignored. (#855)
• Before the first draft arrives, draftomen-tui watch --source moxgate says
  it is waiting for the Moxgate extension to send a draft instead of waiting
  for Arena. (#851)
• The desktop app no longer has a Mocked Draft button in the header. Mocked
  Draft is now an option in the draft source selector next to Arena and
  Moxgate, and choosing it opens the Mocked Draft dialog. With Moxgate
  selected, the account shows as moxgate and cannot be changed. (#850)
• The TUI Moxgate snapshot test waits for the receiving phase as well as the
  rendered pack. The pack can render before the phase changes, which failed
  at random on a busy CI runner.
• draftomen-tui watch --source moxgate shows Moxgate drafts in the TUI.
  --plain is no longer required. Quitting the TUI stops the receiver and
  frees the port. (#833)
• The TUI account-cycling tests wait for the ChooseAccount worker to finish
  after pressing "a" instead of pausing once, which failed at random on a
  busy CI runner.
• The desktop app has a draft source selector next to Settings. Choosing
  Moxgate starts the receiver on 127.0.0.1:47326 and shows whether it is
  loading card data, waiting for the extension or receiving a draft. Choosing
  Arena stops the receiver and frees the port. A port already in use shows an
  error naming the port. Card data comes from the Mocked Draft Scryfall bulk
  file set in Settings or passed with --scryfall-bulk-file. (#829)
• draftomen-tui watch --source moxgate --plain saves Moxgate drafts under the
  fixed account id moxgate and prints a deck build sheet after "Draft
  complete". Moxgate snapshots carry no account id, so these drafts were not
  saved before. (#839)
• The Qt adapter startup coalescing test waits for the post-start state to
  arrive instead of reading the observed states as soon as the session is
  ready, which failed at random in the parallel CI gui job. The gui nox
  session now runs the Qt test files one at a time, because several of them
  failed at random with parallel workers on the 4-core CI runner. Other tests
  still run in parallel.
• A Chrome extension in extensions/moxgate reads the Moxgate draft page and
  posts each pack and pool to draftomen-tui watch --source moxgate --plain on
  127.0.0.1. Load it unpacked from chrome://extensions. Its popup shows whether
  Draft Omen is running and what happened to the last snapshot. A new extension
  CI job runs its tests. (#828)
• draftomen watch --source moxgate --plain starts the Moxgate receiver
  instead of following Player.log and shows recommendations for each pack the
  browser extension posts. --moxgate-port overrides the default port 47326.
  Card data for all sets comes from one local Scryfall bulk file, --bulk-file
  or the Mocked Draft download in the app dir, and watch fails at startup when
  that file is missing. The Moxgate source needs --plain for now. (#832)
• Tests run in parallel with pytest-xdist by default. The full suite takes
  about a quarter of the time it took before on a 10-core machine. Pass -n 0
  to run tests one at a time. The Moxgate receiver also checks for shutdown
  every 0.05 seconds instead of 0.5, so stopping it is faster. The CI tests
  job no longer runs the Qt test files, which the gui job already runs, and
  a Qt adapter test now waits for each published state before reading it.
• Add a loopback receiver for Moxgate snapshots. It listens only on a loopback
  address, accepts POST /moxgate/snapshot from chrome-extension:// origins
  that send the X-Draftomen-Moxgate header, caps the body at 256 KiB and
  answers a malformed snapshot with 400. Accepted snapshots go on a queue that
  the session thread turns into draft events. Nothing starts the receiver yet.
  (#831)
• Add a Moxgate snapshot adapter that turns successive draft page snapshots
  into the existing draft events. It derives each pick and its pack and pick
  coordinates from the change between snapshots, and rejects out-of-order,
  skipped or unresolvable snapshots with an error. (#826)
• The website has a News page that shows the release notes of each version,
  taken from this changelog. Releases now write the page with
  scripts/write_release_news.py, and the release check fails if it is missing.

v0.5.0

• Player-facing text now calls set profiles "ratings". The desktop status
  banner and Settings, the TUI, plain watch and the README use the new word.
  Code, URLs, manifest keys, cache paths and CLI commands keep "profile".
  (#820)
• Add Reality Fracture (FRA) card data and early Premier, Traditional and
  Pick-Two profiles. Scryfall has no Arena ids for FRA yet, so this one-time
  export takes the 280 FRA Arena ids from 17Lands ratings and matches them to
  Scryfall prints by name. The 10 Special Guests cards are left out, as in
  other sets, and so are basic lands, because 17Lands does not rate them. The
  daily refresh keeps this file until it is invalid, so a later re-export
  with Scryfall's own ids is a manual step.
• The README, PRD, website and Store listing now describe Quick, Premier,
  Traditional and Pick-Two drafts instead of Quick Draft only. The README has a
  format table and names the limits of human drafts, including that Traditional
  drafting has no real Arena log sample yet. The log-format document covers
  format detection from event metadata, the 1-based human-draft coordinates,
  picks with one or two cards and the UTC_Log thread prefix. Pick scoring and
  audit documents explain the ratings fallback table and how a Pick-Two pick
  counts against the recommendation. The package, Windows app and
  draftomen-tui --help descriptions now read "draft assistant". A new rule
  in AGENTS.md and the README forbids committing or publishing raw Arena
  logs, anonymised third-party samples or player-identifying data. (#264)
• A human draft whose log skips a pick submission now saves Arena's full
  CardPool from DraftCompleteDraft as its pool, so a Pick-Two draft with
  one missing request keeps all 42 cards instead of 40. The missing pick stays
  empty. The pool updates when the CardPool arrives after the draft already
  completed, and replay and plain watch print a second "Draft complete" line
  that counts the added cards. A CardPool that lacks a recorded card leaves
  the saved pool unchanged and logs a warning. (#800)
• Mocked Draft can run a Pick-Two draft with a 4-seat pod and two cards per
  pick, chosen separately from Manual and Auto. The Mocked Draft dialog has a
  Format selector, and a manual Pick-Two pick stages the first card before
  the second press submits both. Quick Draft keeps its 8-seat
  pod and one card per pick. The test-draft command takes --format quick
  or --format pick_two, and a Pick-Two trace line lists both cards. Premier
  and Traditional are not offered, because bots would only relabel a Quick
  Draft. (#658)
• Premier, Traditional and Pick-Two drafts are supported end to end, checked
  on real Arena logs in plain watch, the TUI and the desktop app. The parser
  now reads the "Entering table draft queue" scene line, for example
  MSH_Premier_Draft or HOB_PickTwo_Draft, as the draft event when a log
  has no EventJoin line. An EventJoin for the same set and format keeps
  its dated event name. scripts/verify_real_logs.py prints the format,
  packs, logical picks, cards, gaps and completion of each local private log
  without card IDs or UUIDs. (#779)
• Add hand-written synthetic fixtures for a complete Premier, Traditional and
  Pick-Two draft, each running from ingestion to plain watch and replay
  output. Pick-Two ends with 21 logical picks and 42 cards, the one-card
  formats with 42 picks and 42 cards. Tests derive the variants from these
  fixtures: thread prefix, missing first notify, missing submission, failed
  pick response, repeated notify, truncated last line, and completion with
  and without DraftCompleteDraft. (#778)
• The desktop live view names the detected draft format and shows pack and
  pick progress from the draft rules, for example "Pack 1 of 3 · Pick 4 of 7 ·
  take 2 cards" in Pick-Two. The recent picks gallery lists every card of a
  logical pick. The pre-draft view and the waiting status no longer mention
  Quick Draft: they read "Ready for your next draft" and "Waiting for a
  draft.", and name the format once it is known. (#777)
• The desktop GUI adapter state gains a draft_progress block with the draft
  format name, pack and pick coordinates, and the pack count, picks per pack
  and cards per pick from the active rules. An unknown format publishes a
  neutral block with every value empty. The pool gains
  recent_logical_picks, which groups recent picks by logical pick so a
  Pick-Two pick lists both cards. Existing state keys are unchanged. (#776)
• The TUI names the detected draft format in the waiting title, the detected
  title, and the status bar, and opens with a neutral "Waiting for a draft"
  state. The pack title shows progress from the draft rules, for example "Pack
  1 of 3, Pick 4 of 7, take 2 cards" in Pick-Two. The pool summary shows the
  last pick with every card, and Pick-Two pool counts read "8 cards from 4
  picks". Quick Draft counts stay unchanged. (#775)
• watch --plain names the draft format in each pick status line and shows
  pack and pick progress from the draft rules, for example "Pack 1 of 3, Pick
  4 of 7, take 2 cards" in Pick-Two. A Pick-Two pick prints both cards on one
  "Chosen cards" line. Watch opens with a neutral "Waiting for a draft." line,
  and a Quick Draft start adds a format summary line. (#774)
• Backtest uses the draft rules for logical pick counts and scores Pick-Two
  selections against the top recommendation. Reports show both selected cards,
  and scoring keeps the saved pool before each pick. Quick Draft output stays
  unchanged. (#773)
• Replay names the draft format and uses its rules for pick counts. Pick-Two
  selections keep both cards at one coordinate. Audit records include the
  format and count how many of the top two cards in the active ranking were
  selected. Quick replay intentionally gains a format header. (#772)
• Live draft progress and pool ledger totals follow the active draft rules.
  Pick-Two records 21 logical picks and 42 cards, and pending recovery keeps
  its format, coordinates and selected cards. Mocked Draft keeps its 13 or 15
  picks per pack. (#771)
• Live Premier, Traditional and Pick-Two drafts load card data on the first
  offered pack, so recommendations and watch --plain scored rows appear
  without a restart. The load permits networking as Quick Draft detection
  does, then later packs and retries use local data only. (#801)
• The live session now loads set profiles and ratings for the detected draft
  format instead of always asking for QuickDraft. Quick uses QuickDraft,
  Premier uses PremierDraft, Traditional tries TradDraft then
  PremierDraft, and Pick-Two tries PickTwoDraft then PremierDraft. The
  local cache and the hosted refresh both follow that order, and an exact
  profile replaces a fallback one when it appears. Snapshots expose the
  detected format, and the contextual evidence status reports a fallback.
  Rating labels name the format that supplied the rating, such as Trad, and
  add * to a fallback, such as Premier*, so Quick Draft cards now read
  Quick instead of Profile. An augmented model is used only when it was
  trained on the draft's own format or on PremierDraft; otherwise Augmented
  Intelligence reports it as unavailable for that format. (#770)
• The log parser now reads human-draft picks and completion for Premier,
  Traditional and Pick-Two drafts. Each successful EventPlayerDraftMakePick
  request becomes one pick with 0-based pack and pick numbers, holding one card
  or two for Pick-Two. A rejected pick is dropped, and a repeated request or
  response counts once. A pick with no logged response is still recorded. The
  draft completes once, from the last pick response or from DraftCompleteDraft,
  with the picks the parser recorded. A missing pick stays a gap. The card pool
  Arena reports is only compared with those picks, and a mismatch writes a
  warning to the log. Human-draft packs now carry the pool picked so far, so
  the pool store accepts them. Malformed or unknown human-draft lines are
  ignored. (#259)
• Live Draft keeps the two-column layout on laptop screens such as 1352x799.
  The card details column shrinks from 550 px to 350 px before the view
  switches to stacked tabs, and the card preview gives up height before the
  pool summary, so the whole mana curve stays visible. Windows narrower than
  about 1240 px or shorter than about 770 px still use stacked tabs. (#242)
• Errors the live session shows, such as a failed ratings or card data
  download, a failed deck build or a failed backtest, are now written to
  ~/.draftomen/logs/draftomen.log with their code, id and message. Each error
  is logged once when it first appears. Recoverable errors log at WARNING and
  others at ERROR. (#796)
• Settings has a Troubleshooting card with an "Open logs folder" button. It
  creates ~/.draftomen/logs if it is missing and opens it in the file
  manager. If the folder cannot be opened, Settings shows the folder path in
  an error message. The README now has a Troubleshooting section that lists
  the files to attach to a bug report and where they live on macOS and
  Windows. (#792)
• Draft Omen now writes a rotating application log to
  ~/.draftomen/logs/draftomen.log. The desktop app and draftomen-tui share
  the file. Each run starts with a line holding the app version, the OS and
  the Python version. The log records unhandled exceptions from the main
  thread and worker threads with their tracebacks, Qt warnings, and the errors
  the desktop app also prints to stderr. The file rotates at 1 MB and keeps
  three older files. If the logs folder cannot be created, the app still
  starts and prints the reason to stderr. The log never copies Player.log
  lines. (#791)
• The draft audit log now records multi-card picks. choice_made records
  store a selected_grp_ids list in place of chosen_grp_id, and the audit
  schema moves to version ...

v0.4.2

• Add docs/microsoft-store.md with what the first Microsoft Store
  submission needs: the Store listing text, runFullTrust justification,
  certification notes, age rating guidance and a publication record to fill
  in. Add Store screenshots rendered from a recorded MSH Quick Draft, two of
  them without card art. (#328)
• The Windows CI job now builds the MSIX with MakeAppx and zips it into a
  .msixupload. It installs a test-signed copy, runs the mock smoke test from
  the installed package, and removes the package. It then uploads the
  .msixupload and the SHA-256 of the .msix inside it as a CI artifact.
  (#756)
• Add the Microsoft Store MSIX package definition for the Windows app: a
  manifest with the Partner Center identity, the Store logos, and
  scripts/msix_package.py, which stages the package folder, maps X.Y.Z to
  X.Y.Z.0, and rejects a release version that does not increase. (#755)
• WOE now has a published augmented model with a Card Pairs table. On held-out
  PremierDraft picks, Pool Shape raises top-1 from 0.4912 to 0.5398 and MRR
  from 0.6692 to 0.7110 over Basic DO. The Card Pairs test passed with a
  held-out gap of +3.88 pp [+3.56, +4.26]. At the selected weight of 0.25,
  pairs raise top-1 to 0.5424 and MRR to 0.7127.
• When the Card Pairs correction moves a card up the ranking, the GUI card
  preview and the TUI focused card details show "Pairs well with ". The
  named card is the pool card with the largest pair score. Cards that pairs did
  not move up show no line. (#745)
• The TUI can now use Augmented Intelligence. The configuration opened with
  c has an Augmented Intelligence switch, off by default and saved in
  tui-preferences.json. When it is on, the TUI loads the set's model, from the
  cache or the published manifest, and DO scores include its adjustment. (#745)
• Pick scores now use the Card Pairs table to break close calls. Each offered
  card gets the sum of its pair scores against the current pool, centered on
  the pack average, times a weight, and clipped to ±8 points. The Card Pairs
  and Pool Shape corrections together stay within ±8 points. The build picks
  the weight with the best validation MRR, then top-1, where weight zero is
  Pool Shape alone. It stores the weight in the table and prints it with
  held-out top-1 and MRR with and without pairs. An artifact without a table, or with a table that
  has no weight, scores as before. (#744)
• build-augmented-set now also downloads the set's 17Lands game dump
  through the checksummed draft dump cache. It builds a Card Pairs table from
  games drafted before Model C's validation period and tests it on the later
  games. The table is published under the artifact's training field only
  when decks with better pairs win more games than decks of equal card
  strength in the same colours and player rank, with a 95% lower bound above
  zero. A failed test publishes the model without a table. The build prints
  the gap, its range and the result. Released clients still load the
  artifact, because they accept any object under training. (#743)
• build-augmented-set now matches pool columns to pack columns by card name,
  so dumps such as WOE that list them in a different order load. It still
  fails when the two column sets name different cards. (#741)
• Add scripts/compare_augmented_candidates.py, an offline spike that ranks
  held-out picks with three candidate models next to Basic DO and Pool Shape,
  the name now used for Model C. Card Pairs scores cards by how often they win
  together in 17Lands game data. Player Picks learns the card players take
  from the exact pack and pool, and Winner Picks does the same from 7-win
  drafts only. The script uses Model C's chronological split and reports
  top-1 and mean reciprocal rank with bootstrap ranges, a view limited to
  7-win drafts, and the rank-adjusted win-rate gap of drafts that followed
  each candidate. Its --games-only mode tests the pair signal on held-out
  17Lands games: whether decks with better pairs win more than decks of equal
  card strength in the same colours and player rank. It reorders pool columns
  for dumps such as WOE that list them in a different order. No app code
  changes. (#717)
• Show 17Lands card-rating counts in the Profile Refresh summary. Each
  successful set and format pair lists the fetched rows, the rows accepted
  into the profile, and the rejected rows grouped by reason. It also says
  whether those counts add up and how many profile ratings came from other
  formats. The static card-data counts are now labelled as set artifacts.
  (#402)
• Update the release skill for signed macOS releases: the new asset names, the
  run time, and a check that a downloaded DMG is notarized and stapled.
• Fix a segfault when the live adapter shuts down while its worker thread is
  finishing. shutdown() now queues the stop through an adapter signal
  instead of calling the worker, which the thread may be deleting. (#727)

v0.4.1

• Publish the signed macOS DMGs as draftomen-vX.Y.Z-macos-arm64.dmg and
  draftomen-vX.Y.Z-macos-x86_64.dmg, and the checksum file as
  draftomen-vX.Y.Z-sha256sums.txt. The Windows executable keeps its
  unsigned name. The website download buttons, the README and the release
  docs use the new names. The README covers installing, Gatekeeper and
  notarization troubleshooting, and uninstalling, and docs/releasing.md
  adds a post-release check that opens a browser-downloaded DMG on a Mac that
  has never run Draft Omen. (#730)
• Sign, notarize and staple the macOS DMGs in tag releases. release.yml
  calls native-bundles.yml with sign_macos: true, and only the macOS jobs
  of that call use the macos-release environment. Each job builds with
  Nuitka's Developer ID signing, copies the verified app to Draft Omen.app
  with ditto so the data-file signatures survive, signs and notarizes the
  DMG, and runs the smoke test on the app inside the stapled DMG. An
  always-run step deletes the temporary keychain and credential files.
  The mounted smoke test now retries a busy hdiutil detach before forcing
  it. Development builds stay unsigned. The release asset names still say
  unsigned until #730 renames them. (#729)
• Document the macOS signing credentials in docs/releasing.md: the
  Developer ID Application certificate, the App Store Connect Team API key,
  the protected macos-release environment that only v* tags can use, the
  secret and variable names, and how to rotate or revoke each credential.
  (#728)
• Stop publishing Draft Omen to PyPI and Homebrew. A tagged release now only
  builds the native macOS and Windows bundles and attaches them to the GitHub
  Release. The release workflow drops the wheel build, the wheel smoke tests,
  the PyPI publish job, and the Homebrew job. homebrew.yml, the formula
  template, scripts/render_homebrew_formula.py, and tests/package_smoke.py
  are removed. Version 0.4.0 stays on PyPI and in the Homebrew tap but gets no
  further updates. The README now points to the GitHub Release downloads.
• Fix the Homebrew formula build. bidict, a dependency of
  python-socketio, builds with uv_build, which Homebrew compiles from
  source with Rust. The formula now declares rust as a build dependency.

v0.4.0

• Publish an Intel macOS DMG next to the Apple Silicon one. A second macOS
  build job runs on the macos-15-intel runner, and each macOS job fails when
  lipo -archs reports the wrong architecture for the app executable. The
  macOS release assets are now named unsigned-macos-arm64.dmg and
  unsigned-macos-x86_64.dmg, the checksum file lists both, and the website
  offers both downloads. (#347)
• Fix the Windows native build, which had two separate faults:
  • Since #332 it failed its bundled-profile check. A Windows checkout
    converted line endings in the HOB profile snapshot and the bundled
    baseline profile, so their bytes no longer matched the recorded sizes and
    SHA-256. A new .gitattributes turns off line-ending conversion for
    profile-snapshots/ and draftomen/baseline_profiles/.
  • Since #555 the built executable failed to start because the
    pyside6-deploy specs excluded QtTest, which removed Qt6Test.dll
    while qt_gui.py imports PySide6.QtTest. The specs no longer exclude it.
• Add a Sets page to the website at /sets/, linked from the navigation. It
  lists the supported sets, those with card data and 17Lands ratings, and
  marks the ones with an augmented model. A second list shows the sets that
  have no 17Lands ratings yet. (#707)
• Load per-set card data in the CLI replay, build, backtest,
  benchmark-picks and refresh-structure-targets commands, the same data
  the desktop app uses. They no longer read ~/.draftomen/carddb.json, and
  ratings metadata recovery no longer writes it. --bulk-file still
  replaces the card data for fixtures and CI. refresh-data now only prints
  that card data downloads per set, and an existing carddb.json can be
  deleted. (#706)
• Keep pipeline caches out of the user data directory. build-augmented-set
  now caches draft dumps in .draftomen/corpus-cache/profile-input-cache, and
  refresh-profile-data caches 17Lands ratings in
  .draftomen/corpus-cache/17lands, both relative to the working directory.
  Both commands take --cache-dir to use another root, and
  refresh-profile-data no longer accepts the hidden --app-dir option.
  Nothing writes ~/.draftomen/profile-input-cache or
  ~/.draftomen/set-enrichment any more, so both can be deleted.
  ~/.draftomen/17lands stays for the CLI build, replay and backtest
  ratings. (#705)
• Check downloaded card data against the sets manifest. With network access,
  the card-data client reads sets/manifest.json once per client, rejects a
  download whose SHA-256 differs from the manifest, and downloads again when a
  cached file no longer matches it. A failed update keeps the valid cached
  file, and an unreachable manifest leaves cached sets in use. (#704)
• Publish sets/manifest.json on the website. It lists every set with card
  data, and gives each set's name, card-data URL, size and SHA-256, profile
  formats with their maturity, and augmented model metrics. export-set-data,
  build-augmented-set and the Profile Refresh publication regenerate it.
  (#703)
• Make build-augmented-set report its progress. It prints one line as each
  stage starts, the dump download at most every 10% and 5 seconds, and how
  many picks training and evaluation use. Basic DO scoring now reports in 25%
  steps instead of once per 5,000 picks. (#710)
• Publish the DFT PremierDraft augmented model. On 889,392 held-out picks it
  raises Basic DO top-1 from 0.520 to 0.562 and mean reciprocal rank from
  0.690 to 0.726.
• Make build-augmented-set train against the set profile the website
  publishes in website/public/profiles, with its checksums verified. It used
  to read only the local app cache and fall back to a generic profile without
  ratings, which calibrated DFT against the wrong baseline and reported it
  only after training. The build now stops before any download when the set
  has no published profile with ratings, and stops before training when the
  dump's format has none.
• Let build-augmented-set train on draft dumps with bonus-sheet cards that
  the set card data does not list, such as the 10 DFT Special Guests. Their
  metadata comes from the local Scryfall bulk file, preferring an Arena
  printing, and the set's own printing wins when both exist. (#715)
• Raise the profile input cache limits to 300 MiB per entry and 1 GiB in
  total. The DFT PremierDraft public draft dump is 205 MB, so the old 128 MiB
  limit rejected it and build-augmented-set DFT trained on the much smaller
  TradDraft dump instead.
• Let build-augmented-set read 17Lands draft dumps that have no pick_2
  column. 17Lands added that column for Pick Two Draft, so older dumps such
  as DFT failed with "Draft dump is missing required columns: ['pick_2']".
  Without the column, no row counts as a second pick.
• Make the weekly historical profile refresh process at most 6 set and format
  pairs per run, keeping each set whole, and skip pairs that already have a
  published profile. The old run refreshed all 128 historical pairs at once,
  17Lands stopped answering after about 105 requests, and the run published
  nothing. A
  historical pair with no 17Lands games, such as IKO TradDraft, is now listed
  as skipped and no longer fails the run. (#709)
• Close the Mocked Draft dialog once the draft started from it is under way,
  so the live drafting view is visible. A failed start keeps the dialog open
  with its error, and reopening the dialog during a draft to leave it keeps it
  open.
• Let Mocked Draft offer booster cards that the set card data does not list,
  such as SPG, MUL, BIG, and OTP bonus-sheet cards and 32 KTK cards. A card
  maps to another Arena printing of the same card when the set has none, and
  its metadata comes from the local Scryfall bulk file, or from the card
  fields Draftmancer sends when the printing never reached Arena. These cards
  carry no ratings and score the neutral prior. (#696, #697)
• Make the GUI replay smoke check wait for a pending hosted profile refresh
  before it captures. A fast replay finished before the MSH profile download,
  so the check showed fallback ratings even though the session adopts the
  profile once it arrives. (#683)
• Let Mocked Draft play Draftmancer packs that are not 14 cards. Pack events
  carry the pack size, and the pick stage, remaining picks, and pool ledger use
  it, so 15-card sets such as LCI and 13-card sets such as BLB draft all 45 or
  39 picks and reach a deck build. Arena drafts keep 14-card packs. (#695,
  #682)
• Fix Mocked Draft booster cards that failed to map to an Arena card. The
  Scryfall mapping now reads every set in the bulk file, uses the first face's
  oracle id when the card has none, and picks the lowest Arena id when a card
  has several Arena printings, such as the LCI Travel Poster basics. This fixes
  booster decoding for LCI, M21, and SNC and startup for TDM and ECL. Booster
  errors now name the card by name, set, and collector number. (#682)
• Add a QML regression test showing that the Mocked Draft set selector keeps
  the chosen set, such as LCI, through state updates and after Start, and that
  Start sends that set. The fix shipped with the set picker in #688. (#681)
• Make "Refresh hosted ratings" work during a Mocked Draft. The refresh now
  uses the simulated set's profile client and updates only the Mocked Draft
  session, so the Arena profile is never touched. If no hosted profile exists,
  the ratings banner says so, and fallback scores stay active. (#687)
• Remove the "DO recommendation" comparison sentence and the repeated
  "Pack 1, pick 6." status line from the desktop Live Draft header in wide
  and narrow layouts. The pack heading, ranked cards, and controls stay. (#655)
• Remove the narrative score paragraph from the desktop Live Draft card
  details in wide and compact layouts. The card image and the structured
  score, win rate, grade, ALSA, and fit fields stay. (#654)
• List every Draftmancer-supported set in the Mocked Draft picker by full name
  and code, such as "The Hobbit (HOB)", in alphabetical name order. A set
  without cached card data shows a "Download set card data" button, and Start
  stays disabled until a validated download completes. A failed download shows
  the set and the reason, and Start stays disabled. (#657)
• Label the desktop card-detail score "Augmented DO Score" when a validated
  per-set model adjusted it, in both wide and narrow layouts and in the
  accessible description. Scores keep the plain "DO Score" label when the
  switch is off or the set has no model. (#656)
• Remove the unused role-profile compile, load, dump, and rebuild helpers and
  the CompiledRoleProfile type. resolve_card_roles and
  RoleClassifier.resolve no longer take a profile argument and always use
  the local classifier with reviewed overrides. (#680)
• Republish the HOB and DFT profiles as clean schema-4 profiles and remove the
  unused enrichment-publications.json record. Retrain the HOB augmented model
  against the clean Basic DO baseline. It passes both held-out gates on 274,932
  picks: top-1 rises from 0.5489 to 0.5922 and mean reciprocal rank from 0.7115
  to 0.7474. (#664)
• Remove the enrich-set and list-enrichment commands, the offline
  semantic-enrichment extraction, review, and projection code, and its
  OpenRouter and guide clients. Profile publication and refresh no longer
  retain enriched manifest entries or report enrichment conflicts. Card-data
  export and ratings-based profile generation are unchanged. (#662)
• Remove typed relationship scoring, pool-ledger support, and advice. Keep
  empirical SetProfile scoring, local card-role classification, generic synergy,
  and existing audit and draft-history readers. (#673)
• Generate clean schema-4 SetProfiles. Ignore retired semantic fields when
  reading schemas 1–3 and reject them in schema 4. Keep confirmed enrichment
  artifacts standalone, treat old semantic-only...

Draft Omen v0.3.1 development build 2026-09-25 #10

Draft Omen v0.3.1 development build 2026-09-25 #10

• Build identifier: v0.3.1-dev.20260925.10
• Source commit: 98713bd7a80b276620eb35c839259ec1c51d891c
• Package version: 0.3.1
• UTC date: 2026-09-25
• Progressive run number: 10
• Warning: This is an unsigned native development build.
• Workflow run: https://github.com/andreagrandi/draftomen/actions/runs/36182054052

Changes since previous release

• Publish an Intel macOS DMG next to the Apple Silicon one. A second macOS
  build job runs on the macos-15-intel runner, and each macOS job fails when
  lipo -archs reports the wrong architecture for the app executable. The
  macOS release assets are now named unsigned-macos-arm64.dmg and
  unsigned-macos-x86_64.dmg, the checksum file lists both, and the website
  offers both downloads. (#347)
• Fix the Windows native build, which had two separate faults:
  • Since #332 it failed its bundled-profile check. A Windows checkout
    converted line endings in the HOB profile snapshot and the bundled
    baseline profile, so their bytes no longer matched the recorded sizes and
    SHA-256. A new .gitattributes turns off line-ending conversion for
    profile-snapshots/ and draftomen/baseline_profiles/.
  • Since #555 the built executable failed to start because the
    pyside6-deploy specs excluded QtTest, which removed Qt6Test.dll
    while qt_gui.py imports PySide6.QtTest. The specs no longer exclude it.
• Add a Sets page to the website at /sets/, linked from the navigation. It
  lists the supported sets, those with card data and 17Lands ratings, and
  marks the ones with an augmented model. A second list shows the sets that
  have no 17Lands ratings yet. (#707)
• Load per-set card data in the CLI replay, build, backtest,
  benchmark-picks and refresh-structure-targets commands, the same data
  the desktop app uses. They no longer read ~/.draftomen/carddb.json, and
  ratings metadata recovery no longer writes it. --bulk-file still
  replaces the card data for fixtures and CI. refresh-data now only prints
  that card data downloads per set, and an existing carddb.json can be
  deleted. (#706)
• Keep pipeline caches out of the user data directory. build-augmented-set
  now caches draft dumps in .draftomen/corpus-cache/profile-input-cache, and
  refresh-profile-data caches 17Lands ratings in
  .draftomen/corpus-cache/17lands, both relative to the working directory.
  Both commands take --cache-dir to use another root, and
  refresh-profile-data no longer accepts the hidden --app-dir option.
  Nothing writes ~/.draftomen/profile-input-cache or
  ~/.draftomen/set-enrichment any more, so both can be deleted.
  ~/.draftomen/17lands stays for the CLI build, replay and backtest
  ratings. (#705)
• Check downloaded card data against the sets manifest. With network access,
  the card-data client reads sets/manifest.json once per client, rejects a
  download whose SHA-256 differs from the manifest, and downloads again when a
  cached file no longer matches it. A failed update keeps the valid cached
  file, and an unreachable manifest leaves cached sets in use. (#704)
• Publish sets/manifest.json on the website. It lists every set with card
  data, and gives each set's name, card-data URL, size and SHA-256, profile
  formats with their maturity, and augmented model metrics. export-set-data,
  build-augmented-set and the Profile Refresh publication regenerate it.
  (#703)
• Make build-augmented-set report its progress. It prints one line as each
  stage starts, the dump download at most every 10% and 5 seconds, and how
  many picks training and evaluation use. Basic DO scoring now reports in 25%
  steps instead of once per 5,000 picks. (#710)
• Publish the DFT PremierDraft augmented model. On 889,392 held-out picks it
  raises Basic DO top-1 from 0.520 to 0.562 and mean reciprocal rank from
  0.690 to 0.726.
• Make build-augmented-set train against the set profile the website
  publishes in website/public/profiles, with its checksums verified. It used
  to read only the local app cache and fall back to a generic profile without
  ratings, which calibrated DFT against the wrong baseline and reported it
  only after training. The build now stops before any download when the set
  has no published profile with ratings, and stops before training when the
  dump's format has none.
• Let build-augmented-set train on draft dumps with bonus-sheet cards that
  the set card data does not list, such as the 10 DFT Special Guests. Their
  metadata comes from the local Scryfall bulk file, preferring an Arena
  printing, and the set's own printing wins when both exist. (#715)
• Raise the profile input cache limits to 300 MiB per entry and 1 GiB in
  total. The DFT PremierDraft public draft dump is 205 MB, so the old 128 MiB
  limit rejected it and build-augmented-set DFT trained on the much smaller
  TradDraft dump instead.
• Let build-augmented-set read 17Lands draft dumps that have no pick_2
  column. 17Lands added that column for Pick Two Draft, so older dumps such
  as DFT failed with "Draft dump is missing required columns: ['pick_2']".
  Without the column, no row counts as a second pick.
• Make the weekly historical profile refresh process at most 6 set and format
  pairs per run, keeping each set whole, and skip pairs that already have a
  published profile. The old run refreshed all 128 historical pairs at once,
  17Lands stopped answering after about 105 requests, and the run published
  nothing. A
  historical pair with no 17Lands games, such as IKO TradDraft, is now listed
  as skipped and no longer fails the run. (#709)
• Close the Mocked Draft dialog once the draft started from it is under way,
  so the live drafting view is visible. A failed start keeps the dialog open
  with its error, and reopening the dialog during a draft to leave it keeps it
  open.
• Let Mocked Draft offer booster cards that the set card data does not list,
  such as SPG, MUL, BIG, and OTP bonus-sheet cards and 32 KTK cards. A card
  maps to another Arena printing of the same card when the set has none, and
  its metadata comes from the local Scryfall bulk file, or from the card
  fields Draftmancer sends when the printing never reached Arena. These cards
  carry no ratings and score the neutral prior. (#696, #697)
• Make the GUI replay smoke check wait for a pending hosted profile refresh
  before it captures. A fast replay finished before the MSH profile download,
  so the check showed fallback ratings even though the session adopts the
  profile once it arrives. (#683)
• Let Mocked Draft play Draftmancer packs that are not 14 cards. Pack events
  carry the pack size, and the pick stage, remaining picks, and pool ledger use
  it, so 15-card sets such as LCI and 13-card sets such as BLB draft all 45 or
  39 picks and reach a deck build. Arena drafts keep 14-card packs. (#695,
  #682)
• Fix Mocked Draft booster cards that failed to map to an Arena card. The
  Scryfall mapping now reads every set in the bulk file, uses the first face's
  oracle id when the card has none, and picks the lowest Arena id when a card
  has several Arena printings, such as the LCI Travel Poster basics. This fixes
  booster decoding for LCI, M21, and SNC and startup for TDM and ECL. Booster
  errors now name the card by name, set, and collector number. (#682)
• Add a QML regression test showing that the Mocked Draft set selector keeps
  the chosen set, such as LCI, through state updates and after Start, and that
  Start sends that set. The fix shipped with the set picker in #688. (#681)
• Make "Refresh hosted ratings" work during a Mocked Draft. The refresh now
  uses the simulated set's profile client and updates only the Mocked Draft
  session, so the Arena profile is never touched. If no hosted profile exists,
  the ratings banner says so, and fallback scores stay active. (#687)
• Remove the "DO recommendation" comparison sentence and the repeated
  "Pack 1, pick 6." status line from the desktop Live Draft header in wide
  and narrow layouts. The pack heading, ranked cards, and controls stay. (#655)
• Remove the narrative score paragraph from the desktop Live Draft card
  details in wide and compact layouts. The card image and the structured
  score, win rate, grade, ALSA, and fit fields stay. (#654)
• List every Draftmancer-supported set in the Mocked Draft picker by full name
  and code, such as "The Hobbit (HOB)", in alphabetical name order. A set
  without cached card data shows a "Download set card data" button, and Start
  stays disabled until a validated download completes. A failed download shows
  the set and the reason, and Start stays disabled. (#657)
• Label the desktop card-detail score "Augmented DO Score" when a validated
  per-set model adjusted it, in both wide and narrow layouts and in the
  accessible description. Scores keep the plain "DO Score" label when the
  switch is off or the set has no model. (#656)
• Remove the unused role-profile compile, load, dump, and rebuild helpers and
  the CompiledRoleProfile type. resolve_card_roles and
  RoleClassifier.resolve no longer take a profile argument and always use
  the local classifier with reviewed overrides. (#680)
• Republish the HOB and DFT profiles as clean schema-4 profiles and remove the
  unused enrichment-publications.json record. Retrain the HOB augmented model
  against the clean Basic DO baseline. It passes both held-out gates on 274,932
  picks: top-1 rises from 0.5489 to 0.5922 and mean reciprocal rank from 0.7115
  to 0.7474. (#664)
• Remove the enrich-set and list-enrichment commands, the offline
  semantic-enrichment extraction, review, and projection code, and its
  OpenRouter and guide clients. Profile publication and refresh no longer
  retain enriched manifest entries or report enrichment conflicts. Card-data
  export and ratings-based profile generation are unchanged. (#6...