podkit — turn a playlist or video URL into a private podcast feed https://git.terst.org/podkit.git/atom?h=main 2026-09-21T07:05:24+00:00 2026-09-21T07:05:24+00:00 Claude noreply@anthropic.com 2026-09-21T07:05:24+00:00 urn:sha1:e3ecae12d7a309eb5a6c8e62910257096cb99375 The machine that hosts a feed and the machine that downloads the audio do not have to be the same, and usually shouldn't be: a datacentre IP is treated very differently from a home connection by the larger video sites. Downloading where the person actually is sidesteps that entirely, and means the server needs no credentials for or route to the source. - push/ingest: `podkit push` uploads a local library to a remote `serve --ingest`, asking for a manifest first so only the difference is sent. Uploads land via a temp file and rename, so an interrupted transfer cannot be picked up as a short episode. Ingest is off by default and refuses to start without a token. - get: downloads a direct audio URL or every enclosure in an RSS feed, with no extractor involved. Numbers RSS items in reverse so a mirrored series plays in order rather than backwards. Two bugs found by running it rather than by testing it: - A tokened server produced enclosure URLs with no token. The feed loaded, listed every episode, and all of them 404'd in a real client — the whole point of the tool, broken in the default configuration. Feeds now carry the token; regression test asserts enclosures from a tokened feed actually resolve against a tokened server. - Ingest ran filepath.Base before validating, so "../evil.mp3" was stored as "evil.mp3" — contained, but under a name the client never asked for, leaving push and the server disagreeing about the library. Names are now validated raw and rejected. Also: identify with a real user-agent (libsyn 403s Go's default), write feed.xml via rename so a polling client never reads a truncated feed, and match ingested file permissions to downloaded ones. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018AwnALZL6bD4VBtATDVv3B 2026-09-21T06:50:42+00:00 Peter Stone thepeterstone@gmail.com 2026-09-21T06:50:42+00:00 urn:sha1:293a627358553c9f61802caf9d4bc76af9c2e239 One Go binary with fetch / tag / feed / serve / all. The stages run separately on purpose: the download is the slow part and should not be repeated while adjusting titles or the base URL. Built from a shell + Python + Go sketch. The pieces that needed fixing on the way, all of which produce a feed that loads wrong or not at all: - Enclosure URLs were unescaped. Track titles routinely contain spaces and ampersands, which make both the URL and the XML invalid — the likeliest way for a feed to fail to load entirely. Only the final path segment is escaped; escaping the base would mangle scheme and host. - Ordering came from filename sort, which is correct only while zero-padding is applied consistently. One unpadded file puts lecture 10 after lecture 1. Order now comes from the parsed track number. - No pubDate and no itunes:type. Podcast clients default to newest-first, which starts a lecture course at the final lecture. Items now carry distinct ascending dates and the feed is marked serial. - Missing channel description and link, which several clients reject outright. - A struct field with no xml tag emitted a stray <EnclosureURL> element. - guid had no isPermaLink="false", so a reader may treat it as a fetchable URL. - Errors from ReadDir, Info and Marshal were discarded, so a missing directory produced an empty feed that looked like a working setup with nothing in it. - os.Args indexed without a length check. Beyond the sketch: a download archive so an interrupted series resumes, an up-front check for yt-dlp and ffmpeg before the long download rather than a post-processing failure after the bytes are on disk, non-audio files excluded by extension so cover art and .part files cannot become episodes, and a serve command with token gating that 404s rather than 401s. Found by the first live run against a LibriVox collection: sources whose own titles begin with a track number came out as "01 - 01 - April.mp3", because the output template prepends the playlist index regardless. Collapsed, but only when the two numbers agree — "01 - 1964 - Lecture" is a year and "05 - 12 Angry Men" is a title, and neither is a repeat. 47 tests. Verified end to end against public-domain audio: downloaded, converted, tagged, served, and the delivered file is a playable 128kbps MP3 with correct ID3 title/artist/album/track frames.