ttlequals0/Audicle v0.55.4 on GitHub

0.55.4

The reset-to-defaults control is red instead of grey. It is the one control on
the page that discards values, and the grey was hard to find against the card.
The README is a short front page again. The details moved to a docs tree with
a per-page split, a glossary that links every term to the page covering it, a
release flow, a deployment runbook, and fresh screenshots of the current UI.
The app image builds with uv 0.12.3 (#119).

The settings save bar appears only when there are unsaved edits, and shows how
many. It sits bottom right at its own size rather than spanning the page: it is
one action among the per-section saves, not the page's headline.

Each settings card can put its fields back to the values the app ships with.
The control is clickable only where something actually differs, and it edits
the form rather than saving, so a reset is undoable until you press save.

Saving now re-baselines the form. It never did, so every page counted as
edited from the first save onward. Nothing showed it while the save bar was
always on screen; making the bar conditional is what surfaced it.

Settings are grouped by subject: Content, Voice, Publishing, Services, System.
The hand-written sections now sit in those categories rather than trailing the
page, which is what had stranded voices away from the voice settings and left
the last category holding whatever was unclaimed, artwork included.
Three groups renamed for what an operator would look for: the category holding
Connections is Services rather than Integrations, since two near-synonyms at
adjacent levels of one hierarchy only invite the question of whether they mean
different things; Diagnostics is Logging, which is the one setting it holds;
and Pipeline is Job timeouts, which is what its four settings actually decide.
The TTS delivery, Job timeouts, and Timeouts groups have help text. Those are
the groups you reach for while something is failing.

0.55.0 was withdrawn: its app and render images were sound, but the wrapper
image could not start.

The TTS wrapper no longer grows until the kernel kills it. It served every
chunk from one process with no cleanup between them, and on long jobs its
resident memory climbed to about 20 GB until the host OOM killer took it out
mid-chunk. That happened three times in one day, each time destroying a job
that had already run 84 to 91 minutes. Above a soft limit it now collects
garbage, empties the CUDA cache, and returns freed allocator arenas to the
kernel; still above a hard limit it restarts itself between chunks, which
costs a pause instead of an episode. Per-chunk resident size is logged, so the
curve is visible rather than something you reconstruct from a kernel log.
A wrapper that is not listening gets its own retry budget. Restarting it takes
60 to 99 seconds while the models reload, and the old budget of seven attempts
covered 61 seconds of backoff, so jobs died a few seconds before the wrapper
could have answered. Connection failures are now bounded by elapsed time
instead of attempts. A comment claiming that budget was 90 seconds was wrong;
the arithmetic gives 61.
Wrapper logs are no longer all errors. uvicorn's lifecycle logger is named
uvicorn.error and carries ordinary INFO lines like "Application startup
complete", and a shipper that reads a level by scanning the text matched the
word "error" in that name. The whole stream arrived in Loki as errors, which
made an error-rate panel useless.

A search box at the top of Settings filters as you type and paints the matching
words. Matches open themselves, so nothing is buried behind a collapsed header,
and clearing the box restores whatever was open before.
Settings are grouped into Content, Voice, Integrations, and Operations. TTS
settings had been split across three groups with Webhooks in between, and the
sections below the save button belonged to no group at all.
Chatterbox can run on another host (#117). TTS_BACKEND=openai-api points at
any OpenAI-compatible speech endpoint; because a remote host shares no volume
with this app, the client writes the audio it returns to the path the pipeline
expects, and the rest of the pipeline neither knows nor cares which backend
ran. A remote server returns no transcript, so WHISPER_BACKEND selects where
verification happens: the bundled wrapper, a remote OpenAI-compatible
transcriber, or nowhere. The two switches are independent, so local synthesis
with remote transcription works. Asking the local wrapper to verify audio made
elsewhere is refused at startup rather than silently returning no transcripts.
42 settings that previously needed a redeploy are now editable in the UI and
the API, including the TTS retry and timeout knobs that were unreachable
during the incident above. Settings that guard the login path stay env-only on
purpose: reaching the Settings UI must not confer the ability to switch off the
defences protecting it.

The Settings save button sits at the end of the page. It saves the grouped
fields, and the eight sections that used to sit below it each save themselves,
which is exactly why its old position read as the middle of the page.

The feed shows 25 episodes at a time. Past 50 the rest were unreachable from
the UI: it asked the API for the first 50 and rendered whatever came back.
Prev and next appear only when there is more than one page, and deleting the
last episode on the last page drops you back a page instead of showing an
empty list.
A search box on the feed page matches episode titles, source URLs, and
uploaded filenames. It runs on the server, so it searches the whole feed
rather than the page on screen. A typed % or _ is literal text, not a
wildcard.
The upload form takes several files at once, up to 20, and each becomes its
own queued job. Files upload one at a time because the server holds each one
in memory while it reads it. One bad file no longer sinks the batch: a
duplicate or an oversized document is reported against its own filename and
stays in the picker for a retry while the rest go through.

GET /api/v1/episodes returns 25 per page by default instead of 50 and
accepts q. X-Total-Count reports the filtered total.