ttlequals0/MinusPod v2.100.26

2.100.26

on GitHub

Added

• The local upload size limit is now a setting (Settings > Audio > Max upload size, or localUploadMaxMb in the API; default 1 GB, range 1 to 4096 MB). The server's own 413 response names the limit, and the docs explain the Cloudflare and nginx body-size caps that can sit in front of it (#920).

Fixed

• A known ad always cuts. A marker backed by a fingerprint match, a sponsor text pattern, a defined pattern, a sponsor with an active pattern, a confirmation or a reviewer verdict now owns any audio it shares with a keep-category marker. Kept audio never carves it, and a verification catch naming a known sponsor is no longer dropped for lying inside a kept span. Before, an outro pattern whose end was only an average-length estimate could swallow the start of a fingerprinted sponsor read, and every later stage deferred to that estimate. The operator's per-category keep still applies to markers of that category.
• Ollama: the JSON schema response format toggle now applies to Ollama, which sends the detection schema as the native /api/chat format (no probe). In plain JSON mode a model cannot return the ad list the prompt asks for: gemma4 answered with one bare ad object per window or picked its own key (items, results, ad_spans), so a window's second and later ads, or all of them, were lost. Detection windows now send a per-addressing-mode schema that requires the fields the prompt requires and caps free text, since an enforcing decoder otherwise skips category and end_text and can loop inside a string until max_tokens.
• JSON schema responses require what the prompt requires. The per-mode detection schema now also requires start_text, the reviewer schema requires is_ad and reason on every item, and the trim-recovery schema requires both nullable edges, so an enforcing decoder cannot drop the verdict or an edge quote.

Changed

• Docs: Parakeet is now the recommended local transcription backend. The WHISPER_BACKEND default is unchanged.

• CI runs pytest on all runner cores with pytest-xdist. Two run-log tests that log a lone surrogate now route it to the recorder without the root logger, since xdist cannot send that text back from a worker, and a prefetch test waits for the lookahead before changing the chunk size. Tests share one database, so a fixture now reverts any setting a test leaves behind and the missing-stack transcriber tests reset the model singleton; the suite passes under every xdist distribution mode.