Fixed
- release: the release workflow no longer builds the container images a second time, for amd64 only, over the same tags.
docker.ymlis the only image build now.
Changelog
Every change that shipped, newest first, and the project in numbers. Built in the open on GitHub.
What comes next? See the roadmap.
docker.yml is the only image build now.npm install vozonda is no longer listed as a way to install, because the npm package installs no code. It stays as a pointer to the app.imghdr no longer exists), and an upload for an unknown feed leaves no file behind./img and the peaks upload refuse paths outside the media directory before touching the disk, and /music/reset accepts only all, intro or outro.strict_map_key rejects a key; it comes in through librosa, and Vozonda decodes no msgpack from outside)podcast:guid (UUIDv5 of its URL, stored once per feed so it survives a new address or feed key); a show can carry its Podcast Index verification token (podcast:txt) to be claimed there, and be marked podcast:locked; podcast:updateFrequency comes from a watchlist's schedule or the gaps between the last episodes; an episode with its own cover gets it as item artwork (with the feed key on private episodes)GET/PUT /pronunciations, GET/PUT /shows/{slug}/pronunciationsPOST /llm/probe); it was a URL parameter, which access logs and browser history keep. GET /llm/probe now refuses a real keyValueError: ... was created in a different Context ("Task exception was never retrieved"): the job runner stopped at done/failed/review and left the job generator suspended, and the event loop closed it later in another context, where resetting the local-only flag failed. It is now closed in the runner's own taskValueError (sf.read result unpacked wrongly), so a resumed render always failed; the turn cache key now covers text, voice, instruction and language, and old position-only cache files are no longer reused (they could replay a line that had since been edited)vozonda --help started the server; it now prints usage, and --host, --port, --version work (defaults still from VOZONDA_HOST / VOZONDA_PORT)publish-pypi skips files PyPI already has (0.7.4 was published before its tag) and runs in the pypi environmenttest_generator_deterministic skips with the reason when headless Chrome cannot start (e.g. in a sandbox without unix sockets) instead of failing every gate therevozonda on PyPI)ghcr.io/vozonda/vozonda)vozonda on npm)description field for custom shows in the UI and API/feeds (previously only available for the default show)main show (default) and pinned to the top of the shows listinternal only, private link, and public directoryqrcode-generator (MIT) as a dependency and lib/qr.ts for QR codes rendered in the browser (the private feed link to a phone; nothing leaves the device)address.public, wins over VOZONDA_PUBLIC_URL; only a plain http(s) address); GET /distribution reports it with its scope (this computer, private network or VPN, internet) and whether it answers, plus public per showPUT /shows/{slug}/public and POST /feed/key/rotate (a new feed key; old links stop working)show.<n>.public, show.default.public), no longer one switch for every feed; a new show starts private; shows from before follow the old feed.public. The master feed /feed.xml without the key lists only public shows (404 when there is none); with the key it lists every show that has a feed. A private episode's media links keep the key in every feed opened with itnpm run preview -- --host): feed links pointed at 127.0.0.1:8787, because the preview rewrites the host; feeds now use VOZONDA_PUBLIC_URL when it is set, like share pages and webhooks, so a podcast app on another device can play the episodes/{show}/{name}/feed.xml) to the API; it now accepts the host of VOZONDA_PUBLIC_URL (plus VOZONDA_ALLOWED_HOSTS) and proxies per-show feedsgit pull && docker compose up -d --build; it asks GitHub at most every six hours, only while the UI is open, and "stop checking" turns it off (setting update.check)/shows, /styles, /plugins, /music, /storage, /clips, transcripts and per-show feeds were not passed to the API by the bundled nginx, so those screens and links failed in the Docker setup/auth/session and other API answers from its asset cache; every API path is now network-firstVOZONDA_TOKEN; before, a reverse proxy in front of the web UI made the API act as if every request were local, so the episode list, sources, transcripts, audio and settings were open, and episodes could be started. The web UI asks for the token once and keeps an HttpOnly session for 30 days. Local use on 127.0.0.1 is unchanged and needs no tokennpm run preview -- --host) passes the client address on; before, the preview made every request look local. VOZONDA_LOCAL_HOSTNAMES adds names that count as localVOZONDA_TOKEN before updatingmetadata_errors to av.open); av is held below 19 in the stt extra and the Docker image<podcast:transcript> tags are extracted without re-transcribing.VOZONDA_AUDIO_MAX_BYTES); exceeding it raises an error instead of truncating; punctuation restoration processes text in ~300-word chunks with per-chunk validation and enable_thinking=false; Whisper model is configurable via VOZONDA_WHISPER_MODEL (default base).VOZONDA_WITH_QWEN_TTS=1 in .env and run docker compose up -d --build (#26). Upgrade note: if you use the qwen_tts engine in Docker, set that variable before you rebuild.source-map-js 1.2.1 -> 1.2.2 (event-loop denial of service advisory); transformers/accelerate advisories stay blocked by the qwen-tts pin, tracked in #26+https://vozonda.com), as Wikimedia's robot policy asks; without it Wikipedia answered 403 to the Docker image, so the quickstart example failed on a clean machinesrc is required in the prompt and repaired once when missing), stored scripts keep src, and the script review keeps it and drops values outside 1..n (#5)POST /sources, POST /sources/upload, GET|PATCH|DELETE /sources/{id}, POST /sources/{id}/retry; POST /jobs takes sources (ids with role main/context), and every episode keeps a copy of its sources: GET /jobs/{id}/sources (public: uploads by title only) and POST /jobs/{id}/sources/clone for create variant (same text, no re-fetch)GET /distribution, PUT /shows/{slug}/rss, user guide in docs/distribution.mdstyle_registry.py): one place per style for order, docs, hook brief, script params, template, TTS speaker instructs, UI group and icon; GET /meta serves style_meta; guide in docs/adding-a-style.mddaily@HH:MM / weekly@<day>@HH:MM per watchlist with timezone, one digest per slot, a missed slot caught up once, retry on the next poll after a feed failure, manual check forces a digestp tag, show metadata is not republished after a restart, default Blossom servers accept long episodespdftotext (PDF sources)show.<n>.nostr keys are accepteddisclosure.ai_label wrote no AI_GENERATED tag (EU AI Act Art. 50); '0' still turns it offlessac (trained on a non-commercial corpus)<br> from h1, changed "Two voices" to "Voices"/source with making-of link