github ttlequals0/MinusPod v2.100.22
2.100.22
on GitHub

pre-release2 hours ago

2.100.22

Fixed

  • The Podping listener keeps one HTTP session for its block polling, so the Hive node is no longer resolved and reconnected on every call (two RPC calls every 3 s, each with its own DNS lookup and TLS handshake before).
  • Detected ads cards on phones: the four controls (play, Not ad, Split, Edit) share one line. Only the Ad review card, which adds Confirm, moves Split and Edit to a second line below 430 px.
  • The reviewer no longer pulls a cut back to the last spoken word across wordless audio at the file head or tail, or at a differential-measured edge, so an ad's closing music is cut with it.

2.100.21

Added

  • The auto-process window is a setting with a per-feed override: default 48 hours, up to 720, for instances that run only occasionally (#895).

Changed

  • UI consistency pass. Headings, buttons, tabs and labels are sentence case (product nouns keep their capitals). Errors read "Could not ." and empty lists read "No yet." or "No match these filters." Held markers are "Needs review" everywhere and the duplicate "Pending review" filter is gone. Marker decisions are "Confirm" and "Not ad" in rows, cards, cue matches and the editor footer. Re-running detection is "Re-detect ads" and the full pipeline is "Full reprocess" in every menu. The nav says "Feeds" and "Add feed" to match the page titles. Destructive confirms repeat verb and noun ("Delete pattern", "Delete 3 episodes") under question titles. "and" replaces "&" outside the editor footer, "Close" replaces "Done", and cue candidates are hidden with "Hide".
  • Numbers: confidence and cue scores are whole percents with one label ("99% confidence", "87% match"). Elapsed and summed durations read "2m 5s" or "1h 2m", positions stay m:ss, timestamps share one formatter with "-" for missing values, and run logs are labelled UTC.
  • Components: every button composes one base recipe (size, radius, focus ring, disabled state) with a colour variant. Text fields, selects and textareas share one chrome and height, and labels one recipe. Badges share one shape with a tint, episode status renders through StatusBadge, stat grids use StatTile (label above value, equal heights), and every close X is CloseButton. Lucide icons replace hand-drawn SVGs and every control has a focus ring. The design guide lists the recipes and the copy rules.
  • Layout: dialogs pin their header and footer and scroll only the body, so footer actions stay on screen on phones. Stats and Search use the underline tabs, form pages share one width, and nested cards became dividers or a muted fill. On phones the Add feed modes stack below 560 px, the status bar shows the stage only, the Sponsors search field spans the row, and episode lists hide artwork below 640 px.

Fixed

  • Docs: whisper.cpp needs --no-flash-attn alongside --dtw on Apple Silicon (Metal) too, not only on CUDA; the native Apple Silicon example now passes it. (#894, from velaia)
  • Episode page: the duration beside an ad block's time range uses the same monospace face, so the numerals share a baseline.
  • Review action rows on phones: every button is 44 px tall, the play button's target height. Labels no longer wrap and a row that cannot fit five buttons wraps to a second line instead of stretching.
  • Icon buttons in action rows fill their 44 px target on phones, so the play button matches the text buttons beside it.

2.100.20

Fixed

  • CPU image workflow: the image secret scan allows the Anthropic SDK's managed-agents docstrings, which mention a service-account type and tripped Trivy's GCP key rule after the SDK bump, failing the arm64 leg.
  • Icons are centered in the "Edit feed title", "Open in editor", ad editor "Add new ad" and ad editor "Close" buttons on phones, where the 44 px touch floor had left them packed to the left.
  • The ad editor's hand-drawn plus and close icons are now the shared lucide icons.
  • Held-marker and Ad Review action rows give every button the same height and center their labels.
  • A browser contract measures icon centering on the routes the suite reaches without a backend (home, settings, search, patterns), so regressions in shared chrome and recipes fail CI. Data-driven pages are checked with the mocked harness.

Changed

  • The reject button on a sponsor marker reads "Not ad" instead of "Not an ad", so it stays on one line beside "Confirm ad" on phones.
  • Settings > System Status no longer carries the sentence pointing at the Stats page; the Stats page is in the main navigation.

2.100.19

Added

  • Docs: nine short videos under docs/videos, each in a portrait and a landscape cut. A tour of the web UI, an animated walk through the processing pipeline, and one video each for the review queue, audio cues, processing modes and segment actions, local feeds, providers and cost, the Podcasting 2.0 output, and learned patterns with Pattern Cleanup. The two animations ship with their source and a rebuild recipe. The recordings ship with the Playwright harness that drives them, so a take can be redone against any instance without putting a feed URL or an API key on screen. The project README embeds the tour and the pipeline animation, and the doc for each feature embeds its landscape cut, uploaded as a GitHub attachment so it plays inline.
  • Cue templates can cut every match on their own, with no ad next to it (#874). Three layers, explicit beating inherited: a global "Cut matched cue audio" setting (off), a per-feed override, and a per-template value next to "Remove cue with ad". Each match of an enabled template becomes a marker with the cue type's category (intro, outro, or sponsor for ad-break cues), so the feed's segment category actions decide whether it is removed, beeped, kept or marked. Short stingers are accepted as cuts and the LLM reviewer leaves these markers alone. Cue-only safety still holds an unproven template. A feed whose templates all cut their own matches can run cue-only without start and end templates. Content transition cues are not eligible. Cue telemetry reports the new "cut" outcome.

Changed

  • LLM prompts use minimal timestamps, [987.6-1023.4] instead of [987.6s - 1023.4s], with the unit stated once in the transcript header. That is about 3 fewer tokens per transcript line. Stored default prompts are refreshed to the new form on upgrade when they were never customized. (#886, from mrfootoyou)
  • Segment-ID prompts mark pauses of 1.5 s or longer between transcript lines with a gap cue, so the model can see jingles and silence it otherwise cannot. The System One adapter skips these rows when it reads a prompt back into segments. (#883, #884, from mrfootoyou)
  • Settings > System Status no longer repeats Time Saved, LLM Tokens and Cost. The Stats page already shows them with filters, and the two sets of numbers could disagree. The panel links to the Stats page instead.
  • The reject button on a held or detected marker whose category is not sponsor reads "Keep" instead of "Not an ad" (follow-up to #877).
  • Per-feed detection notes reach the detector, the verification pass and the reviewer as their own prompt block, marked as operator instructions that override the general guidance, instead of riding inside the podcast description. The notes no longer count against the 800-character description budget, so a long show description cannot truncate them.

Fixed

  • Image build: the 30 upstream security patches for ffmpeg, x264, libass, libsndfile, glibc, cairo, cjson and glib ship in the repo instead of being downloaded from their git forges at build time. One forge now answers automated fetches with a bot challenge page and another changed its patch rendering, so the pinned checksums no longer matched.
  • Detection: when a same-sponsor merge folded a span that sat entirely inside the first one, the merged marker took the inner span's end and shrank. A cross-fetch differential span measured to the end of an inserted spot lost its tail to the LLM's last transcribed word. A few seconds of ad music stayed in the output at an episode start and before the credits. A fold now only extends an end, never pulls it inward.
  • Review: the reviewer could move an ad start inward past an ad-break stinger that sat inside the LLM span, because only edges the cue snap had already moved acted as limits. Removable template cues inside the span now hold the edge too, so the stinger is cut with the ad wherever it falls.
  • MP3 stream copy: a final cut that ends inside the two-frame warm-up before the stream end is extended to the end instead of refusing and falling back to a re-encode. A leading frame the mp3 demuxer skips as junk no longer fails validation when source and copy are short by the same frames. Validation reads from decoder pipes instead of writing two 32-bit PCM files, which removes about 3 GB of temporary disk writes per 70-minute episode. Ported from batram's plusminuspod fork.
  • Foldable phones: the page shell, fixed bars and floating buttons pad for the left and right safe-area insets, so nothing sits under a side status bar on a folded outer display. Stat and tile grids use two or four columns between 640 and 1023 px so rows divide evenly at a center fold. The Feeds grid goes to two columns from 600 px, so the unfolded portrait width gets two cards per row. When the browser reports a two-segment viewport, dialogs stay inside the right-hand half instead of spanning the hinge.
  • Chapters: ffmpeg writes ID3v2.4 frame sizes as plain big-endian integers. A CHAP frame of 128 bytes or more (a long chapter title) then failed the chapter restore with "Invalid ID3 synchsafe integer", and the episode was retried four times and marked permanently failed. The parser now accepts either encoding by checking which reading lands on the next frame, and the tag it writes back is synchsafe. A chapter tag that still cannot be read is logged and skipped instead of failing the cut. Fix from ChaosHat. (#888)
  • Recut: a cue-only feed with transcription skipped has no saved transcript segments. "Apply confirmed and recut" and "Recut Audio" both refused with "No saved transcript segments". Recut now runs without segments when the episode never had a transcript, and the pending-recut list marks such episodes ready. (#891)
  • Ad review editor: dragging the current-position pin always started playback. It now follows the "Play audio while dragging pin" setting, like the boundary pins. (#875)
  • Feed settings: the cue pair override is labelled "Create ads from cue pairs", the same name the global setting uses, so search finds both. (#879)
  • Pattern details: the Podcast link opened the API episodes path and returned Not Found. It now goes to the feed page. (#876)
  • Transcript viewer: segments are shown as paragraphs, grouped at render time only. A paragraph breaks at a real pause, at the first sentence end past 20 seconds, or at 45 seconds. A Paragraphs / Lines toggle keeps the old one-row-per-segment view and remembers the choice. Stored segments and the published VTT are unchanged.
  • Episode page: cut markers that the audio processor merges into one span (overlapping, or within 5 s of each other) show as one block in Detected Ads, with the whole span's play button and a segments toggle that opens the member cards. A lone marker renders as before.
  • Web UI: Calls by phase on the Stats page is a table. The pattern details dialog uses the shared button recipes, and on phones it lays its footer out as equal-width text buttons. The Feeds header shows the compact action set below 768 px, so a wide phone no longer draws the labelled buttons over the title. Copy buttons center their icon when the label is hidden. Play buttons are one square size everywhere, 32 px inside the 44 px target on phones and 28 px on desktop, with the Play glyph optically centered. The Cue Matches section uses the same button. The patterns table scrolls inside a focusable region instead of squeezing its columns until the Confirmed and False Pos. headers collide. Dialog footers, close buttons, secondary destructive actions and icon buttons across the app share one set of button recipes instead of per-page class strings.

Security

Check out latest releases or
releases around ttlequals0/MinusPod v2.100.22

Don't miss a new MinusPod release

NewReleases is sending notifications on new releases.

Get notifications